diff --git a/.beads/interactions.jsonl b/.beads/interactions.jsonl
index 868eca3..0c91efc 100644
--- a/.beads/interactions.jsonl
+++ b/.beads/interactions.jsonl
@@ -27,3 +27,4 @@
 {"id":"int-27353fbf","kind":"field_change","created_at":"2026-06-27T13:28:58.859954398Z","actor":"dawn","issue_id":"klbr-fdr","extra":{"field":"status","new_value":"closed","old_value":"open","reason":"Duplicate/stale: lexical ranking policy was already retired in closed klbr-wmz.14; current code uses FTS-native candidate generation and packet fusion. Optional trigram/BGE sparse work can be filed separately if needed."}}
 {"id":"int-e9f123e3","kind":"field_change","created_at":"2026-06-27T13:28:59.01170437Z","actor":"dawn","issue_id":"klbr-g5i","extra":{"field":"status","new_value":"closed","old_value":"open","reason":"Duplicate of existing blocked klbr-1yn, which already tracks running the official LongMemEval QA evaluator once the upstream evaluator command/environment is available."}}
 {"id":"int-c183c25d","kind":"field_change","created_at":"2026-06-27T14:56:52.022821587Z","actor":"dawn","issue_id":"klbr-9yo","extra":{"field":"status","new_value":"closed","old_value":"in_progress","reason":"Implemented tool-required router calibration metrics and threshold selection; filed klbr-b4z for regenerating saved router artifacts."}}
+{"id":"int-12aecbfc","kind":"field_change","created_at":"2026-06-27T15:08:17.198468383Z","actor":"dawn","issue_id":"klbr-b4z","extra":{"field":"status","new_value":"closed","old_value":"in_progress","reason":"Regenerated router linear artifact out-router-linear-iter-15 with tool-required metrics, updated benchmark helper default to the fresh model, and filed klbr-mpy for residual tool→memory quality work."}}
diff --git a/.beads/issues.jsonl b/.beads/issues.jsonl
index 44ff78a..ca7712d 100644
--- a/.beads/issues.jsonl
+++ b/.beads/issues.jsonl
@@ -29,6 +29,7 @@
 {"_type":"issue","id":"klbr-wmz.8","title":"Validate the LoCoMo adapter and comparison protocol","description":"docs/memory-benches.md recommends LoCoMo as the secondary public suite for memweaver-style comparison, and docs/memory-implementation-status.md says a flexible locomo adapter exists. Current source parses locomo-ish shapes inside klbr-bench/src/longmemeval.rs, but there is no verified dataset fixture, loader test, scoring protocol, or documented run result proving the adapter matches the intended benchmark semantics.","design":"Keep this as a comparison harness task, not a claim about beating another system. The output should make dataset version, reader model, token budget, and scoring method explicit.","acceptance_criteria":"A small LoCoMo-shaped fixture or documented local dataset path exercises the adapter; loader tests cover supported input shapes and session ordering; the benchmark manifest/report clearly labels LoCoMo runs and avoids claiming comparability without matched reader/budget/scoring; docs include the exact command and current verified result or blocker.","status":"closed","priority":2,"issue_type":"task","assignee":"dawn","owner":"90008@klbr.net","created_at":"2026-06-26T17:54:09Z","created_by":"dawn","updated_at":"2026-06-26T23:44:57Z","started_at":"2026-06-26T23:42:41Z","closed_at":"2026-06-26T23:44:57Z","close_reason":"Added LoCoMo smoke fixture, loader tests for session/haystack/conversation shapes, protocol notes in report/manifest, docs with exact command/result, and verified locomo retrieval-only smoke.","labels":["architecture","benchmarks","locomo","memory"],"dependencies":[{"issue_id":"klbr-wmz.8","depends_on_id":"klbr-wmz","type":"parent-child","created_at":"2026-06-26T20:54:09Z","created_by":"dawn","metadata":"{}"}],"dependency_count":0,"dependent_count":0,"comment_count":0}
 {"_type":"issue","id":"klbr-wmz.5","title":"Upgrade episodic notes from transcript cards to source-grounded event cards","description":"klbr-core/src/pipeline.rs render_episode_card creates an episode note from timestamp, source refs, and truncated role lines. docs/memory-arch.md asks for source-grounded scene memory that preserves who, when, where, what changed, what was said, available attachments, and supporting raw refs. The current implementation is addressable, but still too transcript-shaped for temporal/update reasoning.","design":"Do not synthesize unsupported vivid details. The event card should summarize only what source turns or attachments support, with raw refs retained as the escape hatch.","acceptance_criteria":"Episode artifacts have a stable structured representation for time anchors, participants/entities, changes/decisions, salient quotes or snippets, attachments when present, and source refs; the markdown rendering remains human-editable; retrieval/context packets can expose the structured gist without losing provenance; tests cover an episode with multiple turns and verify source refs, session id resolution, and useful promptable chunks.","status":"closed","priority":2,"issue_type":"feature","assignee":"dawn","owner":"90008@klbr.net","created_at":"2026-06-26T17:53:44Z","created_by":"dawn","updated_at":"2026-06-26T23:36:56Z","started_at":"2026-06-26T23:33:59Z","closed_at":"2026-06-26T23:36:56Z","close_reason":"Episode notes now render as structured episode_event_card artifacts with event metadata, source-grounded timelines, stated facts/decisions, attachment markers, source refs, and promptable retrieval coverage; added core regression.","labels":["architecture","episodic","memory","provenance"],"dependencies":[{"issue_id":"klbr-wmz.5","depends_on_id":"klbr-wmz","type":"parent-child","created_at":"2026-06-26T20:53:44Z","created_by":"dawn","metadata":"{}"}],"dependency_count":0,"dependent_count":0,"comment_count":0}
 {"_type":"issue","id":"klbr-wmz.4","title":"Materialize profile and procedural lanes as first-class notes","description":"The schema and MemoryGarden know about profile and procedural lanes, but the production pipeline observe_session path currently writes raw turns plus episodic notes/memories. Stable preferences, standing instructions, and workflows still depend mostly on tags or model-authored remember calls instead of a first-class markdown-note flow with source refs and update policy.","design":"Build on MemoryGarden and upsert_markdown_note rather than adding a new store. Treat legacy memory tags as routing hints, not the durable source of truth for profile/procedural knowledge.","acceptance_criteria":"There is an explicit writer/reflection path for profile_note and procedural_note artifacts; new notes include source refs, frontmatter, stable paths under profile/ or procedural/, and canonical refs/chunks; updates use supersession or versioning instead of silent overwrite; user-confirmation or policy gates are documented for stable profile changes; tests cover creating and updating one profile note and one procedural note from source turns.","status":"closed","priority":2,"issue_type":"feature","assignee":"dawn","owner":"90008@klbr.net","created_at":"2026-06-26T17:53:36Z","created_by":"dawn","updated_at":"2026-06-26T23:42:13Z","started_at":"2026-06-26T23:37:20Z","closed_at":"2026-06-26T23:42:13Z","close_reason":"Added write_memory_note reflection tool for source-grounded profile/procedural markdown notes, ref supersession updates, source/policy frontmatter, stable lane paths, and tests for profile/procedural create/update paths.","labels":["architecture","markdown","memory","profile"],"dependencies":[{"issue_id":"klbr-wmz.4","depends_on_id":"klbr-wmz","type":"parent-child","created_at":"2026-06-26T20:53:36Z","created_by":"dawn","metadata":"{}"}],"dependency_count":0,"dependent_count":0,"comment_count":0}
-{"_type":"issue","id":"klbr-b4z","title":"Regenerate router calibration artifacts","description":"After klbr-9yo, rerun the router bench with the calibrated tool-required metrics and commit fresh benchmarks/models/router model/report outputs so future sweeps track tool→memory and tool-required recall from generated artifacts.","status":"open","priority":3,"issue_type":"task","owner":"90008@klbr.net","created_at":"2026-06-27T14:56:42Z","created_by":"dawn","updated_at":"2026-06-27T14:56:42Z","dependencies":[{"issue_id":"klbr-b4z","depends_on_id":"klbr-9yo","type":"blocks","created_at":"2026-06-27T17:56:51Z","created_by":"dawn","metadata":"{}"}],"dependency_count":1,"dependent_count":0,"comment_count":0}
+{"_type":"issue","id":"klbr-mpy","title":"Improve tool-required router separability","description":"Fresh router artifact benchmarks/models/router/linear/out-router-linear-iter-15 reports test tool-required false-memory rate 0.6913 and tool-required recall 0.3087 after adding the metric/calibration surface. Improve training data, features, or model shape so tool-required queries stop looking like memory queries without regressing memory false-abstain.","status":"open","priority":3,"issue_type":"task","owner":"90008@klbr.net","created_at":"2026-06-27T15:08:07Z","created_by":"dawn","updated_at":"2026-06-27T15:08:07Z","dependencies":[{"issue_id":"klbr-mpy","depends_on_id":"klbr-b4z","type":"blocks","created_at":"2026-06-27T18:08:16Z","created_by":"dawn","metadata":"{}"}],"dependency_count":1,"dependent_count":0,"comment_count":0}
+{"_type":"issue","id":"klbr-b4z","title":"Regenerate router calibration artifacts","description":"After klbr-9yo, rerun the router bench with the calibrated tool-required metrics and commit fresh benchmarks/models/router model/report outputs so future sweeps track tool→memory and tool-required recall from generated artifacts.","status":"closed","priority":3,"issue_type":"task","assignee":"dawn","owner":"90008@klbr.net","created_at":"2026-06-27T14:56:42Z","created_by":"dawn","updated_at":"2026-06-27T15:08:17Z","started_at":"2026-06-27T15:01:12Z","closed_at":"2026-06-27T15:08:17Z","close_reason":"Regenerated router linear artifact out-router-linear-iter-15 with tool-required metrics, updated benchmark helper default to the fresh model, and filed klbr-mpy for residual tool→memory quality work.","dependencies":[{"issue_id":"klbr-b4z","depends_on_id":"klbr-9yo","type":"blocks","created_at":"2026-06-27T17:56:51Z","created_by":"dawn","metadata":"{}"}],"dependency_count":1,"dependent_count":1,"comment_count":0}
 {"_type":"issue","id":"klbr-bqj","title":"add continuous-loop memory benchmark coverage","description":"benchmark suite mostly tests static query retrieval. add a loop-style harness that simulates multi-session observe/compact/reflect cycles and then evaluates whether reflection notes, markdown sync, and passive recall stay aligned over a longer horizon.","status":"open","priority":3,"issue_type":"task","owner":"90008@klbr.net","created_at":"2026-06-27T13:15:48Z","created_by":"dawn","updated_at":"2026-06-27T13:15:48Z","dependency_count":0,"dependent_count":1,"comment_count":0}
 {"_type":"issue","id":"klbr-6an","title":"unify memory edge tables behind canonical refs","description":"memory.rs still has memory_edges and ref-level edges as separate provenance graphs. design and migrate toward one transactional canonical-ref links graph so packet planning, provenance, supersession, and markdown-note evidence traverse the same substrate.","status":"open","priority":3,"issue_type":"task","owner":"90008@klbr.net","created_at":"2026-06-27T13:15:44Z","created_by":"dawn","updated_at":"2026-06-27T13:15:44Z","dependency_count":0,"dependent_count":1,"comment_count":0}
diff --git a/benchmarks/models/router/linear/out-router-linear-iter-15/router_model.json b/benchmarks/models/router/linear/out-router-linear-iter-15/router_model.json
new file mode 100644
index 0000000..6fa13c5
--- /dev/null
+++ b/benchmarks/models/router/linear/out-router-linear-iter-15/router_model.json
@@ -0,0 +1,2074 @@
+{
+  "kind": "linear_softmax_v1",
+  "version": 1,
+  "embed_model": "bge-m3-FP16.gguf",
+  "embed_dim": 1024,
+  "trained_on": "train",
+  "dataset_id": "router-multi(internal-eval-starter-v2+router-train-pack-v1+router-dev-pack-v1+router-abstain-v2+router-abstain-v3+router-memory-lane-v3+router-read-file-repo-v3+router-read-file-repo-v4+router-shell-state-v3+router-holdout-tools-recall-v1+router-holdout-abstain-toolish-v1+router-holdout-memory-toolish-v1+router-shell-state-v1+router-shell-state-v1b+router-shell-state-v2+router-read-file-repo-v1+router-read-file-repo-v1b+router-write-file-v1+router-memory-lane-v1+router-memory-lane-v1b+router-abstain-v1+router-memory-adversarial-v1)",
+  "labels": [
+    "memory",
+    "abstain"
+  ],
+  "weights": [
+    [
+      0.0909187,
+      -0.10178791,
+      0.11678524,
+      0.05916823,
+      0.070548564,
+      0.055725943,
+      0.038828287,
+      0.13350812,
+      0.06834853,
+      0.053196143,
+      -0.0385138,
+      0.07048024,
+      0.016072715,
+      -0.023768598,
+      -0.027139591,
+      0.048342563,
+      0.12934941,
+      -0.05374762,
+      0.11772681,
+      0.15524524,
+      -0.092097424,
+      -0.0031241039,
+      0.037270628,
+      0.06646499,
+      -0.13187897,
+      0.011146338,
+      0.0047771796,
+      0.085530005,
+      -0.12908898,
+      -0.0928597,
+      -0.047699794,
+      0.27952304,
+      0.030982267,
+      0.036123082,
+      -0.0049592783,
+      -0.025655432,
+      -0.034489375,
+      0.004012741,
+      0.10140389,
+      -0.25641724,
+      -0.044225235,
+      0.10645814,
+      -0.010120653,
+      0.02796328,
+      0.117799245,
+      -0.07337869,
+      0.10935339,
+      -0.033037018,
+      -0.029130008,
+      0.09551594,
+      -0.04300705,
+      -0.03437312,
+      -0.090900406,
+      0.078975394,
+      0.055480488,
+      -0.13741483,
+      0.09314146,
+      -0.01475522,
+      0.1041528,
+      0.012438029,
+      0.03960217,
+      -0.1029107,
+      0.0713567,
+      -0.051731355,
+      0.09445201,
+      -0.15577719,
+      -0.02391808,
+      0.0067842794,
+      -0.046360753,
+      -0.13958193,
+      -0.026215952,
+      -0.2007777,
+      0.063762814,
+      0.011291224,
+      0.022197852,
+      -0.01546266,
+      0.026640097,
+      -0.010218101,
+      0.13750845,
+      -0.04044518,
+      0.035794824,
+      0.062463112,
+      0.17495327,
+      0.04420096,
+      -0.049676433,
+      -0.13711317,
+      0.056510847,
+      -0.021217274,
+      0.10655209,
+      -0.24909946,
+      0.071961224,
+      0.1314598,
+      -0.106257886,
+      -0.029360218,
+      -0.032705307,
+      0.07170856,
+      0.073151775,
+      -0.007486854,
+      -0.101780094,
+      0.046930544,
+      0.035663072,
+      0.014383753,
+      -0.07339751,
+      -0.08912065,
+      0.100061595,
+      -0.09376313,
+      -0.017745597,
+      0.011771549,
+      -0.07223824,
+      0.015942672,
+      0.013549201,
+      -0.08492904,
+      0.049977064,
+      0.0030273998,
+      -0.14002652,
+      0.024056265,
+      0.04748304,
+      0.11573629,
+      0.19070493,
+      0.10864362,
+      -0.015834443,
+      0.007529914,
+      0.060697675,
+      0.0888037,
+      0.040031034,
+      -0.042574655,
+      -0.06335538,
+      -0.27767628,
+      -0.15719801,
+      -0.10027293,
+      -0.06005399,
+      -0.095936164,
+      0.040199954,
+      0.021806598,
+      0.05969676,
+      -0.010764068,
+      -0.0074787457,
+      0.0017246295,
+      -0.14612263,
+      0.13989076,
+      0.05256759,
+      -0.13347758,
+      -0.011251092,
+      0.038681954,
+      -0.13007946,
+      0.27539828,
+      -0.045249272,
+      0.05831534,
+      -0.046448134,
+      -0.02713023,
+      -0.014157591,
+      -0.13915585,
+      -0.077046916,
+      0.045503356,
+      0.11997959,
+      -0.11390483,
+      0.09459816,
+      0.15865034,
+      -0.029030869,
+      -0.006999421,
+      -0.09143082,
+      0.19256212,
+      -0.11315463,
+      -0.01221871,
+      0.04000163,
+      0.03570344,
+      0.016955221,
+      0.16437307,
+      -0.09222379,
+      0.020818336,
+      -0.040878497,
+      0.022809286,
+      0.08395719,
+      0.015888765,
+      0.0116329165,
+      -0.08944471,
+      0.13184665,
+      -0.062960155,
+      0.0504186,
+      0.007710603,
+      -0.03072083,
+      -0.06868697,
+      0.15584622,
+      -0.01815849,
+      0.064208604,
+      -0.10350769,
+      0.057464138,
+      -0.2722616,
+      -0.019180324,
+      0.016849259,
+      0.044826373,
+      -0.09458674,
+      0.022836355,
+      0.008754259,
+      -0.011425031,
+      0.13440506,
+      -0.074882835,
+      0.07931913,
+      0.14062048,
+      -0.03560135,
+      0.14323571,
+      0.09257049,
+      -0.09006065,
+      0.008161624,
+      -0.17046283,
+      0.1371611,
+      -0.076164976,
+      -0.17762338,
+      0.062453173,
+      0.2814841,
+      -0.019066462,
+      0.094567865,
+      0.10161007,
+      0.00808024,
+      0.08481967,
+      -0.06828979,
+      -0.037755657,
+      0.0687689,
+      0.006235402,
+      0.100143686,
+      0.012343185,
+      -0.12530406,
+      -0.14283738,
+      -0.32921645,
+      0.09596835,
+      0.21708255,
+      -0.0063855127,
+      -0.12804818,
+      -0.10758542,
+      -0.046675142,
+      0.030714657,
+      0.09501781,
+      -0.03233305,
+      -0.034008082,
+      -0.18698777,
+      -0.12576406,
+      0.08127337,
+      -0.12850589,
+      0.0040291753,
+      -0.041365903,
+      0.035329875,
+      -0.09262305,
+      0.018961063,
+      0.09265923,
+      0.019368619,
+      -0.09696165,
+      -0.1703949,
+      -0.13549255,
+      0.06260995,
+      -0.02586587,
+      0.0813209,
+      0.11605256,
+      0.049140953,
+      -0.14482489,
+      0.121894896,
+      -0.029212832,
+      0.04796723,
+      0.04732705,
+      0.05929951,
+      -0.036168512,
+      0.07790013,
+      -0.089916855,
+      -0.05808762,
+      -0.06261247,
+      0.025491863,
+      0.016963989,
+      -0.15251179,
+      -0.09780863,
+      0.044852488,
+      0.15522851,
+      0.004163399,
+      0.060386475,
+      0.052862376,
+      0.075335175,
+      0.096873686,
+      -0.043037854,
+      -0.08131523,
+      -0.050964788,
+      0.018023998,
+      -0.015964346,
+      0.030023769,
+      -0.11333087,
+      0.062238727,
+      0.0087387655,
+      -0.03900802,
+      -0.07718013,
+      0.055245925,
+      -0.09055882,
+      -0.018206658,
+      0.10128244,
+      0.11391827,
+      0.014042869,
+      -0.013369883,
+      0.08306669,
+      -0.10710397,
+      -0.23349969,
+      -0.119926244,
+      0.13514437,
+      -0.08946301,
+      0.0052589774,
+      -0.18481515,
+      -0.09958813,
+      0.035068393,
+      -0.11119286,
+      -0.059356723,
+      0.07889465,
+      0.08490952,
+      -0.1109816,
+      0.030580774,
+      0.056059625,
+      0.062322304,
+      0.0018498958,
+      -0.036662873,
+      -0.01024365,
+      0.030097203,
+      0.028883696,
+      0.08366871,
+      0.118037455,
+      0.02474948,
+      -0.2562628,
+      -0.031159466,
+      0.07297978,
+      -0.060752954,
+      -0.086939186,
+      0.07456483,
+      0.0727932,
+      -0.07192665,
+      0.16692351,
+      0.1113542,
+      -0.042771876,
+      0.09893118,
+      0.002331599,
+      -0.12041482,
+      0.20090795,
+      0.083464235,
+      -0.08245038,
+      -0.01440022,
+      -0.03549796,
+      -0.20425177,
+      -0.0512299,
+      -0.032467492,
+      0.024495548,
+      -0.031912405,
+      -0.011308912,
+      -0.10386546,
+      0.00260585,
+      0.011933582,
+      -0.0020977953,
+      0.05871067,
+      -0.016828889,
+      0.058007333,
+      0.09139152,
+      -0.006920547,
+      0.085679986,
+      -0.04961015,
+      -0.23878542,
+      -0.0771982,
+      0.18772614,
+      0.19800672,
+      -0.050248574,
+      0.117505535,
+      -0.012533416,
+      0.062146544,
+      -0.17442483,
+      0.024063043,
+      0.027210535,
+      -0.1120231,
+      -0.026570767,
+      0.014570307,
+      0.058837563,
+      0.17475893,
+      -0.27570313,
+      0.09190608,
+      0.03375931,
+      -0.025897522,
+      0.03163071,
+      0.26014018,
+      0.04906938,
+      -0.12640563,
+      0.07089076,
+      -0.25859892,
+      0.031826537,
+      0.03999971,
+      0.026293384,
+      0.059014767,
+      0.0024002811,
+      -0.070438445,
+      0.016921084,
+      -0.01577732,
+      -0.028592844,
+      -0.17768851,
+      0.09612699,
+      0.1774312,
+      -0.06794901,
+      0.009430504,
+      -0.04489909,
+      -0.12650795,
+      0.147283,
+      0.0038914857,
+      0.013293183,
+      0.1679682,
+      -0.1632235,
+      0.059381045,
+      0.06847858,
+      -0.036620826,
+      -0.051043946,
+      -0.09699919,
+      -0.0020425543,
+      0.019366229,
+      0.0733083,
+      0.035790343,
+      0.0227435,
+      0.0052870885,
+      0.048689444,
+      0.047039106,
+      -0.16933644,
+      0.0026446306,
+      -0.046715964,
+      0.111923575,
+      0.06061537,
+      -0.07394871,
+      0.08074568,
+      0.0060020373,
+      0.1842132,
+      0.27424693,
+      -0.075155534,
+      0.13272786,
+      0.02185443,
+      -0.08577906,
+      0.021838648,
+      -0.1637421,
+      0.027415913,
+      -0.09351988,
+      -0.10044491,
+      0.004764631,
+      0.029550739,
+      0.0018653506,
+      0.17249644,
+      -0.008504574,
+      0.08730007,
+      -0.10313084,
+      0.035127502,
+      -0.23126693,
+      0.23637879,
+      -0.03239225,
+      -0.18939355,
+      0.12732793,
+      0.009465992,
+      -0.0003209131,
+      -0.18080606,
+      -0.08046365,
+      0.012449741,
+      -0.118323445,
+      0.104933105,
+      0.21097045,
+      0.060342215,
+      -0.07835461,
+      0.02240911,
+      0.20818727,
+      0.045269955,
+      -0.085078776,
+      -0.069585264,
+      0.14184217,
+      0.12192964,
+      -0.011012893,
+      -0.06999858,
+      -0.02708106,
+      0.027025241,
+      0.06562223,
+      0.09168855,
+      -0.037480846,
+      0.026390834,
+      -0.056737408,
+      0.08500966,
+      0.08676688,
+      0.065942384,
+      -0.16103101,
+      0.043179717,
+      0.020403195,
+      -0.11863198,
+      -0.041670058,
+      0.065624006,
+      -0.03196418,
+      -0.0043724068,
+      0.008897633,
+      -0.18209058,
+      -0.15713122,
+      0.047153823,
+      0.08730384,
+      0.13942477,
+      -0.028682373,
+      0.028083358,
+      -0.057008397,
+      -0.046886034,
+      -0.046192247,
+      0.06715759,
+      0.044568393,
+      0.002113484,
+      -0.13828546,
+      0.15636078,
+      0.06954997,
+      -0.14953147,
+      -0.112560436,
+      0.019261394,
+      -0.16593508,
+      0.13439812,
+      -0.043217584,
+      -0.20450737,
+      0.044981003,
+      -0.009157525,
+      0.04975826,
+      0.11736464,
+      0.09473598,
+      -0.26620403,
+      -0.007065188,
+      -0.106640525,
+      0.09735151,
+      0.031973537,
+      0.011108966,
+      -0.030196192,
+      -0.0153452195,
+      0.11463631,
+      -0.21934934,
+      -0.06123633,
+      -0.052847028,
+      -0.03319332,
+      0.05807538,
+      -0.11826212,
+      0.0527611,
+      -0.006141906,
+      -0.0770732,
+      -0.11037585,
+      -0.105413966,
+      -0.036367007,
+      0.19955967,
+      -0.057379395,
+      0.10194378,
+      -0.013492098,
+      -0.3781815,
+      -0.0025361583,
+      0.026621494,
+      -0.010930109,
+      0.013397153,
+      -0.025689544,
+      -0.052768644,
+      -0.093645394,
+      0.03010893,
+      0.12366996,
+      -0.13778171,
+      -0.17512321,
+      -0.024462402,
+      0.12694672,
+      0.07129117,
+      0.07700503,
+      0.059647348,
+      0.07097167,
+      -0.02528138,
+      0.054619722,
+      0.11518964,
+      0.088611774,
+      0.14925942,
+      0.078001074,
+      -0.044878468,
+      -0.067617044,
+      0.0942157,
+      -0.029932145,
+      -0.1649983,
+      0.099805035,
+      0.16583516,
+      -0.13795446,
+      -0.1263526,
+      -0.01923009,
+      -0.100225374,
+      0.03533568,
+      0.0404571,
+      0.048099708,
+      0.10554729,
+      0.005023324,
+      -0.06945862,
+      0.056317195,
+      0.06053066,
+      -0.029253667,
+      -0.11922597,
+      -0.16693494,
+      -0.034542665,
+      0.042180587,
+      -0.00618467,
+      0.21037851,
+      -0.03096064,
+      0.16982804,
+      0.14634919,
+      0.19202565,
+      -0.008642678,
+      -0.012609946,
+      -0.20218313,
+      -0.18344676,
+      0.009707233,
+      0.03880171,
+      -0.19555259,
+      -0.0035742945,
+      0.093816675,
+      -0.03327149,
+      -0.033702333,
+      0.062421475,
+      0.10471949,
+      0.13329011,
+      0.035363387,
+      0.091764204,
+      -0.24594139,
+      0.04124924,
+      -0.11141196,
+      -0.021625469,
+      -0.15837601,
+      -0.06019995,
+      -0.031484544,
+      0.040021174,
+      0.01650623,
+      0.11193421,
+      -0.10561927,
+      0.04596731,
+      0.036409006,
+      0.04902609,
+      0.0467768,
+      -0.03385237,
+      0.01149601,
+      -0.04075288,
+      -0.086375795,
+      -0.037321445,
+      -0.017607475,
+      -0.071824186,
+      -0.14443898,
+      -0.3359552,
+      0.043195643,
+      -0.092355154,
+      -0.07717329,
+      0.117886715,
+      0.19942063,
+      0.118601896,
+      -0.05155628,
+      -0.02441101,
+      0.10421796,
+      -0.008014208,
+      0.26350906,
+      -0.07039495,
+      -0.029995237,
+      0.03396719,
+      0.06255244,
+      0.015134183,
+      -0.0026772607,
+      -0.0038688714,
+      -0.022462266,
+      -0.061070275,
+      0.06623726,
+      -0.009832414,
+      -0.0032685688,
+      0.06205474,
+      -0.013509927,
+      0.0072961375,
+      -0.04311228,
+      -0.23439503,
+      -0.07214877,
+      -0.045175053,
+      -0.092938885,
+      -0.23029418,
+      -0.12370092,
+      -0.035253774,
+      0.028771548,
+      -0.03771489,
+      -0.11057414,
+      -0.12100764,
+      0.05827887,
+      -0.040878106,
+      0.33400786,
+      0.029531842,
+      0.03355193,
+      0.027735762,
+      0.21585152,
+      0.060360543,
+      0.11951367,
+      -0.063856,
+      -0.03125572,
+      0.16415974,
+      -0.018139273,
+      0.07573278,
+      0.023114916,
+      0.004682118,
+      -0.03016507,
+      0.16820696,
+      0.01668169,
+      -0.009222808,
+      -0.07329377,
+      -0.03207888,
+      -0.11533002,
+      0.017110866,
+      -0.089399196,
+      -0.010132421,
+      -0.10615803,
+      0.03353094,
+      -0.045490343,
+      -0.19455181,
+      -0.09528548,
+      -0.0072150077,
+      -0.14551137,
+      -0.07574543,
+      -0.07105609,
+      0.10890402,
+      0.06929344,
+      0.010575259,
+      -0.066625185,
+      0.042084593,
+      0.046797868,
+      0.09010806,
+      0.031406708,
+      0.10253122,
+      0.11853467,
+      -0.0070101772,
+      -0.08128251,
+      0.12124126,
+      -0.06370236,
+      -0.05522786,
+      0.2804675,
+      0.0147011485,
+      -0.10216334,
+      0.033266407,
+      -0.21405481,
+      0.11068325,
+      0.13874714,
+      -0.11626716,
+      -0.01424192,
+      -0.13125478,
+      -0.0026530842,
+      -0.11729388,
+      -0.014237908,
+      0.04701423,
+      -0.017909212,
+      0.044821057,
+      -0.010106488,
+      -0.005959039,
+      0.042763535,
+      0.07690725,
+      -0.055813257,
+      -0.15889576,
+      -0.06334118,
+      0.3404751,
+      -0.026150825,
+      0.062182184,
+      0.014162228,
+      -0.025227025,
+      0.1349675,
+      -0.025621857,
+      -0.1545953,
+      -0.006323669,
+      -0.0047049993,
+      -0.16874228,
+      0.110527135,
+      -0.029964997,
+      -0.19455348,
+      0.065799035,
+      -0.04781946,
+      0.034305893,
+      0.15128328,
+      0.06876025,
+      -0.024659593,
+      0.172743,
+      0.124386065,
+      0.031809356,
+      -0.0066621266,
+      0.09854871,
+      0.35968181,
+      -0.12246443,
+      -0.12588087,
+      0.044067364,
+      -0.088743724,
+      -0.006179159,
+      -0.016170332,
+      0.09629168,
+      0.10287478,
+      -0.05048483,
+      -0.15435089,
+      -0.023985432,
+      0.1762534,
+      0.034906767,
+      0.007885429,
+      0.051634364,
+      -0.015163402,
+      -0.026850164,
+      0.18908618,
+      0.05733349,
+      0.008059986,
+      0.101305835,
+      0.02769927,
+      0.20904636,
+      -0.16414832,
+      0.051534127,
+      -0.19636646,
+      -0.0031428759,
+      -0.14756316,
+      -0.21607679,
+      0.042433016,
+      -0.16239521,
+      0.009235724,
+      -0.024397142,
+      0.18176028,
+      -0.05934566,
+      0.07906501,
+      -0.06342852,
+      -0.122705214,
+      0.086101316,
+      0.024250219,
+      0.100434914,
+      -0.08447745,
+      0.14350761,
+      0.0065760002,
+      -0.15320979,
+      -0.1550891,
+      -0.020352894,
+      -0.027650552,
+      -0.02985656,
+      0.08008903,
+      -0.10814273,
+      0.021938505,
+      0.028200425,
+      0.078790516,
+      -0.07990491,
+      -0.21740301,
+      -0.012074666,
+      0.012521158,
+      0.13365597,
+      -0.038121063,
+      0.059651006,
+      -0.05100821,
+      0.0504861,
+      0.06439014,
+      -0.13510308,
+      0.060516976,
+      -0.092174366,
+      0.058194377,
+      -0.09388816,
+      0.13937302,
+      0.0197103,
+      0.078052714,
+      -0.088731535,
+      0.111742556,
+      0.08743099,
+      -0.13886444,
+      0.008587345,
+      0.087134205,
+      -0.06717643,
+      -0.13828802,
+      -0.090106554,
+      -0.023472186,
+      0.04368845,
+      0.021393524,
+      -0.13013981,
+      0.15462554,
+      -0.057321034,
+      -0.10127157,
+      -0.028595693,
+      0.093094125,
+      -0.042103864,
+      0.22155926,
+      -0.0130425505,
+      0.07462886,
+      -0.056797173,
+      0.039937433,
+      -0.018057175,
+      -0.05877615,
+      0.0733413,
+      0.08236741,
+      0.069186494,
+      -0.011915797,
+      0.15068631,
+      -0.101410225,
+      -0.11714112,
+      0.20504177,
+      0.058380783,
+      0.015036767,
+      0.17739035,
+      -0.048152056,
+      -0.08256131,
+      -0.12287246,
+      0.18831412,
+      -0.055647146,
+      0.118204355,
+      0.11487204,
+      -0.1261344,
+      -0.12811424,
+      0.05310729,
+      -0.10494798,
+      0.038551517,
+      -0.033783097,
+      -0.12150896,
+      0.04779768,
+      -0.14517125,
+      0.16845013,
+      0.15802872,
+      -0.030408494,
+      -0.050764624,
+      0.045974307,
+      0.02037497,
+      -0.010381571,
+      0.022415577,
+      -0.13675034,
+      0.055272937,
+      -0.117484584,
+      -0.09043007,
+      0.07730109,
+      -0.15255336,
+      0.02765536,
+      -0.019286819,
+      0.007561362,
+      0.036112603,
+      -0.021919219,
+      0.018293597,
+      -0.09924775,
+      -0.03393517,
+      -0.031558562,
+      -0.10081357,
+      0.016195873,
+      -0.038190186,
+      0.13615939,
+      -0.032655925,
+      0.077086784,
+      0.06676277,
+      -0.15092671,
+      0.26104182,
+      0.00073005824,
+      0.13841376,
+      0.019728431,
+      0.084298,
+      0.06365132,
+      0.053714555,
+      -0.023481913,
+      -0.073658735,
+      0.094291925,
+      -0.046430625,
+      0.13661909,
+      0.0048703407,
+      -0.09880886,
+      0.20059393,
+      0.06522984,
+      -0.0033890887,
+      0.008679226,
+      -0.03582931,
+      -0.070584595,
+      0.0062678573,
+      0.027285222,
+      0.121596545,
+      0.18270735,
+      -0.051998466,
+      -0.0024515297,
+      0.057480533,
+      0.03778221,
+      0.01285929,
+      -0.063858345,
+      -0.12761942,
+      0.025317665,
+      -0.08674136,
+      -0.17180698,
+      -0.18613546,
+      -0.0059342072,
+      0.022553451,
+      -0.1678678,
+      0.115535185,
+      0.040501345,
+      -0.043164544,
+      -0.0039145667,
+      0.19659255,
+      0.13500725,
+      -0.00032373134,
+      -0.021232175,
+      -0.041546453,
+      -0.057601646,
+      -0.10057167,
+      -0.06948283,
+      -0.027852042,
+      -0.00024421138,
+      0.054853093,
+      -0.12560737,
+      0.20087229,
+      0.17841148,
+      0.044842776,
+      -0.16634,
+      0.05749792,
+      0.07120963,
+      -0.10685926,
+      -0.103528924,
+      -0.16392906,
+      0.016186714,
+      0.12114929,
+      -0.016533075,
+      0.079132736,
+      0.015467807,
+      -0.09476922,
+      0.05216517,
+      -0.16601095,
+      0.09375119,
+      0.05882024,
+      0.13059746,
+      -0.16618133,
+      -0.1350551,
+      -0.05530104,
+      0.028868213,
+      -0.023587227,
+      -0.12164127,
+      0.07728809,
+      -0.07231829,
+      -0.036748417,
+      -0.18092078,
+      -0.21404143,
+      -0.043164216,
+      0.03013793,
+      0.09951888,
+      -0.019226337,
+      -0.03692522,
+      -0.010236446,
+      0.23704085,
+      0.024363797,
+      0.087785296,
+      0.03604801,
+      0.0636236,
+      0.0029409393,
+      0.10468185,
+      -0.0034361675
+    ],
+    [
+      -0.090918705,
+      0.10178792,
+      -0.11678526,
+      -0.05916824,
+      -0.070548564,
+      -0.055725947,
+      -0.03882828,
+      -0.1335081,
+      -0.06834853,
+      -0.05319615,
+      0.038513795,
+      -0.07048023,
+      -0.016072715,
+      0.023768596,
+      0.027139593,
+      -0.048342563,
+      -0.12934943,
+      0.053747617,
+      -0.11772681,
+      -0.15524524,
+      0.092097424,
+      0.003124105,
+      -0.037270628,
+      -0.06646498,
+      0.13187897,
+      -0.011146336,
+      -0.004777177,
+      -0.08553,
+      0.129089,
+      0.09285971,
+      0.04769979,
+      -0.27952307,
+      -0.030982269,
+      -0.036123086,
+      0.004959277,
+      0.025655426,
+      0.034489382,
+      -0.004012741,
+      -0.1014039,
+      0.25641724,
+      0.04422523,
+      -0.10645815,
+      0.010120655,
+      -0.027963284,
+      -0.11779925,
+      0.07337869,
+      -0.10935339,
+      0.033037018,
+      0.029130006,
+      -0.095515944,
+      0.043007042,
+      0.034373123,
+      0.09090041,
+      -0.078975394,
+      -0.055480488,
+      0.13741483,
+      -0.09314148,
+      0.014755225,
+      -0.1041528,
+      -0.012438028,
+      -0.03960217,
+      0.1029107,
+      -0.0713567,
+      0.05173136,
+      -0.094452016,
+      0.1557772,
+      0.02391808,
+      -0.0067842794,
+      0.046360753,
+      0.13958195,
+      0.026215952,
+      0.20077768,
+      -0.063762814,
+      -0.0112912245,
+      -0.02219785,
+      0.01546266,
+      -0.02664009,
+      0.0102181,
+      -0.13750845,
+      0.040445186,
+      -0.03579482,
+      -0.06246311,
+      -0.17495327,
+      -0.044200953,
+      0.049676426,
+      0.13711315,
+      -0.05651084,
+      0.021217277,
+      -0.106552064,
+      0.24909946,
+      -0.07196123,
+      -0.13145979,
+      0.106257886,
+      0.029360222,
+      0.032705303,
+      -0.07170857,
+      -0.07315177,
+      0.007486855,
+      0.101780094,
+      -0.046930548,
+      -0.035663072,
+      -0.014383758,
+      0.07339751,
+      0.08912065,
+      -0.100061595,
+      0.09376314,
+      0.017745595,
+      -0.011771546,
+      0.07223824,
+      -0.01594267,
+      -0.013549202,
+      0.08492905,
+      -0.04997707,
+      -0.0030274016,
+      0.14002652,
+      -0.024056267,
+      -0.04748304,
+      -0.11573629,
+      -0.19070493,
+      -0.108643614,
+      0.015834438,
+      -0.007529915,
+      -0.06069767,
+      -0.088803716,
+      -0.040031042,
+      0.042574655,
+      0.06335537,
+      0.2776763,
+      0.157198,
+      0.10027294,
+      0.060053993,
+      0.09593617,
+      -0.040199958,
+      -0.0218066,
+      -0.059696767,
+      0.010764066,
+      0.0074787457,
+      -0.0017246301,
+      0.14612265,
+      -0.13989076,
+      -0.052567583,
+      0.13347758,
+      0.011251086,
+      -0.038681958,
+      0.13007946,
+      -0.27539828,
+      0.04524926,
+      -0.058315337,
+      0.046448134,
+      0.027130226,
+      0.014157587,
+      0.13915583,
+      0.077046916,
+      -0.045503356,
+      -0.1199796,
+      0.11390482,
+      -0.09459817,
+      -0.15865034,
+      0.029030873,
+      0.0069994237,
+      0.09143081,
+      -0.19256212,
+      0.113154635,
+      0.012218711,
+      -0.040001627,
+      -0.035703443,
+      -0.016955223,
+      -0.16437308,
+      0.09222379,
+      -0.020818336,
+      0.0408785,
+      -0.022809282,
+      -0.08395719,
+      -0.015888765,
+      -0.011632918,
+      0.0894447,
+      -0.13184664,
+      0.062960155,
+      -0.050418597,
+      -0.0077106007,
+      0.030720836,
+      0.06868697,
+      -0.15584622,
+      0.018158488,
+      -0.06420861,
+      0.10350769,
+      -0.057464138,
+      0.2722616,
+      0.019180322,
+      -0.016849259,
+      -0.044826362,
+      0.09458675,
+      -0.022836352,
+      -0.008754258,
+      0.011425034,
+      -0.13440505,
+      0.074882835,
+      -0.07931913,
+      -0.14062047,
+      0.03560134,
+      -0.14323568,
+      -0.09257049,
+      0.09006065,
+      -0.008161625,
+      0.17046285,
+      -0.13716109,
+      0.076164976,
+      0.17762338,
+      -0.062453173,
+      -0.2814841,
+      0.019066457,
+      -0.09456787,
+      -0.101610065,
+      -0.00808024,
+      -0.08481967,
+      0.06828978,
+      0.037755653,
+      -0.0687689,
+      -0.0062354016,
+      -0.10014368,
+      -0.012343183,
+      0.12530406,
+      0.14283738,
+      0.32921645,
+      -0.09596834,
+      -0.21708255,
+      0.006385513,
+      0.12804818,
+      0.10758543,
+      0.046675146,
+      -0.030714657,
+      -0.09501783,
+      0.032333042,
+      0.03400808,
+      0.18698777,
+      0.12576406,
+      -0.08127338,
+      0.12850587,
+      -0.0040291757,
+      0.041365903,
+      -0.035329875,
+      0.09262304,
+      -0.018961059,
+      -0.09265921,
+      -0.019368613,
+      0.09696165,
+      0.1703949,
+      0.13549256,
+      -0.062609956,
+      0.025865871,
+      -0.0813209,
+      -0.11605256,
+      -0.04914095,
+      0.14482489,
+      -0.12189488,
+      0.029212827,
+      -0.047967225,
+      -0.04732705,
+      -0.05929952,
+      0.036168516,
+      -0.077900134,
+      0.089916855,
+      0.058087625,
+      0.06261247,
+      -0.025491865,
+      -0.01696399,
+      0.15251179,
+      0.09780864,
+      -0.044852484,
+      -0.15522853,
+      -0.004163396,
+      -0.060386486,
+      -0.052862376,
+      -0.075335175,
+      -0.09687368,
+      0.04303785,
+      0.08131523,
+      0.05096478,
+      -0.018024003,
+      0.015964342,
+      -0.030023776,
+      0.11333087,
+      -0.062238727,
+      -0.0087387655,
+      0.03900802,
+      0.07718012,
+      -0.055245925,
+      0.09055883,
+      0.018206658,
+      -0.10128244,
+      -0.113918275,
+      -0.014042869,
+      0.013369879,
+      -0.083066694,
+      0.10710397,
+      0.2334997,
+      0.11992624,
+      -0.13514432,
+      0.08946301,
+      -0.005258978,
+      0.18481515,
+      0.09958813,
+      -0.03506839,
+      0.11119286,
+      0.059356723,
+      -0.07889465,
+      -0.084909506,
+      0.1109816,
+      -0.030580772,
+      -0.056059625,
+      -0.062322292,
+      -0.001849897,
+      0.036662873,
+      0.010243648,
+      -0.030097201,
+      -0.028883694,
+      -0.08366871,
+      -0.118037455,
+      -0.024749476,
+      0.2562628,
+      0.031159468,
+      -0.07297978,
+      0.060752958,
+      0.086939186,
+      -0.074564815,
+      -0.0727932,
+      0.071926646,
+      -0.1669235,
+      -0.111354195,
+      0.04277188,
+      -0.09893118,
+      -0.0023316005,
+      0.12041482,
+      -0.20090793,
+      -0.08346423,
+      0.08245039,
+      0.014400216,
+      0.035497963,
+      0.20425177,
+      0.051229905,
+      0.032467492,
+      -0.024495544,
+      0.0319124,
+      0.011308912,
+      0.10386546,
+      -0.0026058508,
+      -0.011933583,
+      0.0020977962,
+      -0.058710657,
+      0.016828887,
+      -0.058007337,
+      -0.09139152,
+      0.0069205477,
+      -0.08567999,
+      0.049610157,
+      0.23878542,
+      0.0771982,
+      -0.18772614,
+      -0.1980067,
+      0.05024857,
+      -0.11750554,
+      0.01253342,
+      -0.06214654,
+      0.17442481,
+      -0.02406304,
+      -0.02721053,
+      0.11202311,
+      0.026570763,
+      -0.014570306,
+      -0.058837567,
+      -0.17475894,
+      0.2757031,
+      -0.09190606,
+      -0.03375931,
+      0.025897523,
+      -0.03163071,
+      -0.26014018,
+      -0.04906938,
+      0.12640563,
+      -0.07089077,
+      0.25859892,
+      -0.031826533,
+      -0.03999971,
+      -0.026293382,
+      -0.059014764,
+      -0.0024002783,
+      0.070438445,
+      -0.016921086,
+      0.01577732,
+      0.028592845,
+      0.17768852,
+      -0.09612699,
+      -0.1774312,
+      0.06794902,
+      -0.0094305035,
+      0.0448991,
+      0.12650795,
+      -0.147283,
+      -0.0038914855,
+      -0.013293184,
+      -0.1679682,
+      0.16322352,
+      -0.05938105,
+      -0.06847858,
+      0.036620837,
+      0.051043935,
+      0.09699918,
+      0.0020425557,
+      -0.01936623,
+      -0.0733083,
+      -0.035790347,
+      -0.022743497,
+      -0.0052870857,
+      -0.048689455,
+      -0.047039106,
+      0.16933645,
+      -0.0026446308,
+      0.046715964,
+      -0.111923575,
+      -0.06061537,
+      0.07394871,
+      -0.0807457,
+      -0.006002038,
+      -0.18421322,
+      -0.2742469,
+      0.07515553,
+      -0.13272786,
+      -0.021854423,
+      0.08577906,
+      -0.02183865,
+      0.1637421,
+      -0.027415918,
+      0.093519874,
+      0.10044491,
+      -0.004764628,
+      -0.029550739,
+      -0.0018653506,
+      -0.17249642,
+      0.008504572,
+      -0.08730007,
+      0.10313084,
+      -0.035127506,
+      0.23126698,
+      -0.2363788,
+      0.03239225,
+      0.18939355,
+      -0.12732795,
+      -0.0094659915,
+      0.00032092226,
+      0.18080604,
+      0.08046364,
+      -0.0124497395,
+      0.118323445,
+      -0.1049331,
+      -0.21097045,
+      -0.060342208,
+      0.07835461,
+      -0.022409111,
+      -0.20818728,
+      -0.04526996,
+      0.08507878,
+      0.069585264,
+      -0.14184217,
+      -0.121929646,
+      0.011012894,
+      0.06999857,
+      0.027081057,
+      -0.02702524,
+      -0.06562223,
+      -0.09168855,
+      0.03748085,
+      -0.026390832,
+      0.056737408,
+      -0.08500967,
+      -0.086766884,
+      -0.06594239,
+      0.16103102,
+      -0.043179717,
+      -0.020403197,
+      0.118631996,
+      0.041670054,
+      -0.065624006,
+      0.03196418,
+      0.004372405,
+      -0.008897635,
+      0.18209055,
+      0.15713122,
+      -0.047153823,
+      -0.08730385,
+      -0.13942477,
+      0.028682373,
+      -0.02808336,
+      0.057008397,
+      0.04688603,
+      0.046192247,
+      -0.0671576,
+      -0.044568397,
+      -0.0021134838,
+      0.13828546,
+      -0.15636078,
+      -0.06954997,
+      0.14953147,
+      0.112560436,
+      -0.019261394,
+      0.16593508,
+      -0.13439813,
+      0.043217577,
+      0.20450738,
+      -0.044981,
+      0.009157527,
+      -0.049758263,
+      -0.11736464,
+      -0.09473599,
+      0.266204,
+      0.0070651877,
+      0.10664051,
+      -0.09735151,
+      -0.031973537,
+      -0.011108966,
+      0.030196195,
+      0.015345223,
+      -0.11463632,
+      0.21934934,
+      0.06123632,
+      0.05284703,
+      0.033193324,
+      -0.05807538,
+      0.11826212,
+      -0.052761097,
+      0.006141908,
+      0.0770732,
+      0.110375844,
+      0.105413966,
+      0.036367007,
+      -0.19955969,
+      0.057379384,
+      -0.10194378,
+      0.013492098,
+      0.37818155,
+      0.0025361555,
+      -0.026621489,
+      0.010930118,
+      -0.013397153,
+      0.025689546,
+      0.05276865,
+      0.0936454,
+      -0.03010893,
+      -0.123669945,
+      0.13778171,
+      0.17512321,
+      0.024462402,
+      -0.12694673,
+      -0.07129118,
+      -0.077005014,
+      -0.059647337,
+      -0.07097167,
+      0.025281385,
+      -0.054619726,
+      -0.115189634,
+      -0.08861177,
+      -0.14925943,
+      -0.07800107,
+      0.044878464,
+      0.067617044,
+      -0.0942157,
+      0.029932136,
+      0.16499831,
+      -0.099805035,
+      -0.16583516,
+      0.13795446,
+      0.1263526,
+      0.019230088,
+      0.100225374,
+      -0.035335675,
+      -0.040457096,
+      -0.04809971,
+      -0.105547294,
+      -0.005023326,
+      0.069458604,
+      -0.056317203,
+      -0.06053065,
+      0.029253665,
+      0.11922596,
+      0.16693497,
+      0.034542665,
+      -0.04218059,
+      0.0061846687,
+      -0.21037851,
+      0.030960638,
+      -0.16982804,
+      -0.14634919,
+      -0.19202565,
+      0.008642676,
+      0.012609947,
+      0.20218313,
+      0.18344678,
+      -0.009707233,
+      -0.03880171,
+      0.1955526,
+      0.003574293,
+      -0.09381668,
+      0.03327149,
+      0.033702336,
+      -0.06242147,
+      -0.10471946,
+      -0.13329011,
+      -0.035363384,
+      -0.091764204,
+      0.24594142,
+      -0.041249245,
+      0.11141196,
+      0.021625465,
+      0.15837602,
+      0.06019995,
+      0.031484544,
+      -0.04002117,
+      -0.016506229,
+      -0.111934215,
+      0.105619274,
+      -0.045967307,
+      -0.036409,
+      -0.04902609,
+      -0.046776805,
+      0.03385237,
+      -0.01149601,
+      0.04075288,
+      0.086375795,
+      0.037321445,
+      0.017607475,
+      0.071824186,
+      0.14443898,
+      0.33595517,
+      -0.043195643,
+      0.09235515,
+      0.07717328,
+      -0.117886715,
+      -0.19942063,
+      -0.11860189,
+      0.05155627,
+      0.024411008,
+      -0.10421797,
+      0.008014208,
+      -0.2635091,
+      0.07039495,
+      0.029995235,
+      -0.033967186,
+      -0.062552445,
+      -0.015134183,
+      0.0026772628,
+      0.0038688711,
+      0.022462267,
+      0.061070275,
+      -0.06623726,
+      0.009832416,
+      0.003268568,
+      -0.062054727,
+      0.013509927,
+      -0.0072961412,
+      0.043112278,
+      0.23439501,
+      0.07214878,
+      0.045175053,
+      0.0929389,
+      0.23029421,
+      0.12370091,
+      0.035253778,
+      -0.028771546,
+      0.03771489,
+      0.11057414,
+      0.12100764,
+      -0.05827887,
+      0.040878106,
+      -0.33400786,
+      -0.029531848,
+      -0.033551935,
+      -0.027735764,
+      -0.2158515,
+      -0.060360547,
+      -0.11951368,
+      0.063856,
+      0.031255722,
+      -0.16415976,
+      0.018139273,
+      -0.075732775,
+      -0.023114912,
+      -0.004682119,
+      0.030165067,
+      -0.16820696,
+      -0.016681688,
+      0.009222809,
+      0.07329377,
+      0.032078873,
+      0.11533001,
+      -0.017110873,
+      0.089399196,
+      0.010132426,
+      0.10615803,
+      -0.033530936,
+      0.045490336,
+      0.19455181,
+      0.09528548,
+      0.007215007,
+      0.14551137,
+      0.07574545,
+      0.07105608,
+      -0.10890402,
+      -0.069293454,
+      -0.01057526,
+      0.06662519,
+      -0.04208459,
+      -0.046797857,
+      -0.09010806,
+      -0.03140671,
+      -0.10253122,
+      -0.11853466,
+      0.007010177,
+      0.08128252,
+      -0.12124125,
+      0.06370236,
+      0.055227865,
+      -0.28046748,
+      -0.014701144,
+      0.10216332,
+      -0.0332664,
+      0.2140548,
+      -0.110683255,
+      -0.13874713,
+      0.116267174,
+      0.014241925,
+      0.13125478,
+      0.002653086,
+      0.11729388,
+      0.014237907,
+      -0.047014236,
+      0.017909214,
+      -0.04482106,
+      0.010106489,
+      0.0059590437,
+      -0.04276354,
+      -0.07690724,
+      0.05581326,
+      0.15889578,
+      0.063341185,
+      -0.3404751,
+      0.026150826,
+      -0.062182184,
+      -0.0141622275,
+      0.025227023,
+      -0.13496749,
+      0.02562185,
+      0.15459532,
+      0.006323666,
+      0.0047049974,
+      0.16874228,
+      -0.11052715,
+      0.029964995,
+      0.19455346,
+      -0.06579904,
+      0.04781946,
+      -0.034305893,
+      -0.1512833,
+      -0.06876026,
+      0.024659596,
+      -0.172743,
+      -0.124386065,
+      -0.031809352,
+      0.006662127,
+      -0.09854872,
+      -0.35968181,
+      0.12246443,
+      0.12588087,
+      -0.044067368,
+      0.08874373,
+      0.0061791586,
+      0.016170332,
+      -0.09629168,
+      -0.10287476,
+      0.05048483,
+      0.1543509,
+      0.023985434,
+      -0.17625344,
+      -0.034906764,
+      -0.007885429,
+      -0.051634364,
+      0.015163402,
+      0.026850162,
+      -0.18908617,
+      -0.057333484,
+      -0.008059984,
+      -0.101305835,
+      -0.027699258,
+      -0.20904638,
+      0.16414833,
+      -0.051534135,
+      0.19636646,
+      0.0031428759,
+      0.14756314,
+      0.21607679,
+      -0.04243302,
+      0.1623952,
+      -0.009235722,
+      0.024397142,
+      -0.18176028,
+      0.059345663,
+      -0.07906498,
+      0.06342852,
+      0.12270523,
+      -0.08610132,
+      -0.024250219,
+      -0.100434914,
+      0.08447745,
+      -0.14350761,
+      -0.006575999,
+      0.15320978,
+      0.1550891,
+      0.020352894,
+      0.027650552,
+      0.029856551,
+      -0.08008903,
+      0.108142726,
+      -0.021938507,
+      -0.028200423,
+      -0.07879052,
+      0.079904914,
+      0.217403,
+      0.012074669,
+      -0.012521159,
+      -0.13365597,
+      0.038121056,
+      -0.059651006,
+      0.051008206,
+      -0.0504861,
+      -0.06439014,
+      0.13510308,
+      -0.06051698,
+      0.09217437,
+      -0.058194384,
+      0.09388817,
+      -0.13937302,
+      -0.0197103,
+      -0.078052714,
+      0.088731535,
+      -0.11174255,
+      -0.08743099,
+      0.13886444,
+      -0.008587343,
+      -0.08713419,
+      0.06717642,
+      0.13828804,
+      0.090106554,
+      0.023472188,
+      -0.043688454,
+      -0.021393526,
+      0.13013981,
+      -0.1546255,
+      0.057321034,
+      0.10127156,
+      0.02859569,
+      -0.09309413,
+      0.04210387,
+      -0.22155924,
+      0.013042551,
+      -0.07462887,
+      0.056797177,
+      -0.039937425,
+      0.018057175,
+      0.05877615,
+      -0.07334131,
+      -0.082367405,
+      -0.069186494,
+      0.0119157955,
+      -0.15068631,
+      0.10141023,
+      0.11714112,
+      -0.20504175,
+      -0.05838078,
+      -0.015036771,
+      -0.17739035,
+      0.04815205,
+      0.082561314,
+      0.122872464,
+      -0.18831415,
+      0.055647142,
+      -0.11820435,
+      -0.11487204,
+      0.1261344,
+      0.12811424,
+      -0.05310729,
+      0.10494799,
+      -0.038551517,
+      0.033783097,
+      0.121508956,
+      -0.04779768,
+      0.14517125,
+      -0.16845016,
+      -0.1580287,
+      0.030408496,
+      0.050764624,
+      -0.045974307,
+      -0.020374969,
+      0.010381569,
+      -0.022415573,
+      0.13675034,
+      -0.05527294,
+      0.117484584,
+      0.09043007,
+      -0.077301085,
+      0.15255338,
+      -0.027655361,
+      0.019286815,
+      -0.0075613605,
+      -0.036112595,
+      0.021919217,
+      -0.0182936,
+      0.09924776,
+      0.033935167,
+      0.031558562,
+      0.100813575,
+      -0.016195877,
+      0.038190186,
+      -0.13615939,
+      0.032655925,
+      -0.077086784,
+      -0.06676277,
+      0.15092671,
+      -0.26104182,
+      -0.00073005847,
+      -0.13841376,
+      -0.019728435,
+      -0.084298,
+      -0.06365133,
+      -0.05371456,
+      0.023481913,
+      0.073658735,
+      -0.094291925,
+      0.046430625,
+      -0.13661909,
+      -0.00487034,
+      0.09880886,
+      -0.2005939,
+      -0.06522985,
+      0.00338909,
+      -0.008679226,
+      0.035829317,
+      0.070584595,
+      -0.006267858,
+      -0.027285222,
+      -0.121596545,
+      -0.18270732,
+      0.051998466,
+      0.0024515307,
+      -0.05748052,
+      -0.03778221,
+      -0.012859291,
+      0.063858345,
+      0.12761942,
+      -0.025317667,
+      0.086741365,
+      0.17180698,
+      0.18613546,
+      0.0059342054,
+      -0.022553453,
+      0.1678678,
+      -0.11553518,
+      -0.04050134,
+      0.043164548,
+      0.003914565,
+      -0.19659255,
+      -0.13500723,
+      0.00032373256,
+      0.021232177,
+      0.041546453,
+      0.057601646,
+      0.10057167,
+      0.06948283,
+      0.027852042,
+      0.000244211,
+      -0.054853093,
+      0.12560737,
+      -0.20087229,
+      -0.17841148,
+      -0.04484277,
+      0.16633998,
+      -0.05749792,
+      -0.07120963,
+      0.10685925,
+      0.10352892,
+      0.16392906,
+      -0.016186712,
+      -0.12114929,
+      0.016533079,
+      -0.079132736,
+      -0.015467804,
+      0.09476923,
+      -0.052165166,
+      0.16601096,
+      -0.09375119,
+      -0.058820244,
+      -0.13059746,
+      0.16618133,
+      0.13505507,
+      0.055301044,
+      -0.02886822,
+      0.023587227,
+      0.12164127,
+      -0.0772881,
+      0.07231829,
+      0.03674842,
+      0.18092078,
+      0.21404143,
+      0.043164212,
+      -0.030137932,
+      -0.09951888,
+      0.019226344,
+      0.036925215,
+      0.010236445,
+      -0.23704083,
+      -0.024363795,
+      -0.087785296,
+      -0.036048014,
+      -0.06362361,
+      -0.0029409386,
+      -0.10468186,
+      0.0034361666
+    ]
+  ],
+  "bias": [
+    -0.56765383,
+    0.56765395
+  ],
+  "decision": {
+    "abstain_prob_threshold": 0.90000004,
+    "abstain_margin_threshold": 0.0
+  }
+}
\ No newline at end of file
diff --git a/benchmarks/models/router/linear/out-router-linear-iter-15/router_report.json b/benchmarks/models/router/linear/out-router-linear-iter-15/router_report.json
new file mode 100644
index 0000000..f6a32db
--- /dev/null
+++ b/benchmarks/models/router/linear/out-router-linear-iter-15/router_report.json
@@ -0,0 +1,2921 @@
+{
+  "dataset_id": "router-multi(internal-eval-starter-v2+router-train-pack-v1+router-dev-pack-v1+router-abstain-v2+router-abstain-v3+router-memory-lane-v3+router-read-file-repo-v3+router-read-file-repo-v4+router-shell-state-v3+router-holdout-tools-recall-v1+router-holdout-abstain-toolish-v1+router-holdout-memory-toolish-v1+router-shell-state-v1+router-shell-state-v1b+router-shell-state-v2+router-read-file-repo-v1+router-read-file-repo-v1b+router-write-file-v1+router-memory-lane-v1+router-memory-lane-v1b+router-abstain-v1+router-memory-adversarial-v1)",
+  "embed_model": "bge-m3-FP16.gguf",
+  "trained_on": "train",
+  "evaluated_on": "test",
+  "decision": {
+    "abstain_prob_threshold": 0.90000004,
+    "abstain_margin_threshold": 0.0
+  },
+  "metrics": {
+    "total": 319,
+    "accuracy": 0.5862069,
+    "confusion": [
+      [
+        100,
+        0
+      ],
+      [
+        132,
+        87
+      ]
+    ],
+    "memory_false_abstain_rate": 0.0,
+    "abstain_false_memory_rate": 0.60273975,
+    "tool_required_false_memory_rate": 0.6912752,
+    "tool_required_recall": 0.30872482,
+    "memory_count": 100,
+    "abstain_count": 219,
+    "tool_required_count": 149,
+    "misclassified": [
+      {
+        "query_id": "test_q9",
+        "text": "what gpu am i using for the reranker?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.2033282,
+        "p_abstain": 0.79667175
+      },
+      {
+        "query_id": "test_q15",
+        "text": "which postgres extension are we enabling before passive recall?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.17207056,
+        "p_abstain": 0.8279294
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_002",
+        "text": "how much free disk space is there on the partition containing agent.db?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.10456361,
+        "p_abstain": 0.89543635
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_003",
+        "text": "what's the current hostname?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.101953685,
+        "p_abstain": 0.8980464
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_004",
+        "text": "is port 8002 bound to 127.0.0.1 or 0.0.0.0 right now?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.1582868,
+        "p_abstain": 0.8417132
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_006",
+        "text": "are there any zombie processes named klbr right now?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.11492517,
+        "p_abstain": 0.8850748
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_008",
+        "text": "check if the reranker http endpoint on localhost:8003 is responding",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.16735494,
+        "p_abstain": 0.8326451
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_011",
+        "text": "which file defines the ServerMsg type?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.12588543,
+        "p_abstain": 0.8741145
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_013",
+        "text": "read klbr-core/src/agent.rs and summarize how routing affects memory injection",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.17401817,
+        "p_abstain": 0.82598186
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_014",
+        "text": "what does klbr-core/src/context.rs do when recalled memories are present?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.14702685,
+        "p_abstain": 0.85297316
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_015",
+        "text": "what crate defines the LlmClient type?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.23538512,
+        "p_abstain": 0.7646149
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_016",
+        "text": "write a file docs/router_holdout_tools_recall.txt with the content \"ok\"",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "write_file"
+        ],
+        "p_memory": 0.14410025,
+        "p_abstain": 0.8558998
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_017",
+        "text": "append a line \"router ok\" to docs/router_holdout_log.txt",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "write_file"
+        ],
+        "p_memory": 0.11356302,
+        "p_abstain": 0.88643694
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_018",
+        "text": "create a file benchmarks/router_holdout_tmp.txt and put the number 123 in it",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "write_file"
+        ],
+        "p_memory": 0.13842863,
+        "p_abstain": 0.8615714
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_019",
+        "text": "今ポート8002は開いてる?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.10369836,
+        "p_abstain": 0.8963016
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_020",
+        "text": "klbr-core/src/llm.rsのembed関数って何返すの?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.16805679,
+        "p_abstain": 0.8319432
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_023",
+        "text": "in klbr-core/src/config.rs, what is memory_support_score_gap set to?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.15972814,
+        "p_abstain": 0.8402719
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_024",
+        "text": "how many open file descriptors does the klbr-daemon process have right now?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.105587445,
+        "p_abstain": 0.8944126
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_025",
+        "text": "what's the size of the current target/ directory?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.1315417,
+        "p_abstain": 0.8684583
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_026",
+        "text": "what ip address is localhost:8002 actually bound to right now?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.1407521,
+        "p_abstain": 0.8592479
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_027",
+        "text": "which file defines the ClientMsg type in the ipc protocol?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.11255213,
+        "p_abstain": 0.88744783
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_028",
+        "text": "in klbr-core/src/router.rs, what routing decisions exist?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.20984548,
+        "p_abstain": 0.7901545
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_032",
+        "text": "open klbr-core/src/config.rs and tell me what llm_url is set to",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.11325379,
+        "p_abstain": 0.8867462
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_033",
+        "text": "open klbr-core/src/config.rs and tell me what rerank_url is set to",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.16435155,
+        "p_abstain": 0.8356485
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_034",
+        "text": "in klbr-daemon/src/daemon.rs, does it forward agent errors to the tui?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.13177624,
+        "p_abstain": 0.8682238
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_035",
+        "text": "in klbr-tui/src/main.rs, what happens when ServerMsg::Error arrives?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.11597728,
+        "p_abstain": 0.8840227
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_038",
+        "text": "is the disk partition containing /home close to full right now?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.11078995,
+        "p_abstain": 0.88921
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_039",
+        "text": "what is the network interface with the highest tx bytes right now?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.14258735,
+        "p_abstain": 0.85741264
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_040",
+        "text": "is port 1234 open right now?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.10612941,
+        "p_abstain": 0.8938706
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_041",
+        "text": "what environment variable is RUST_LOG set to right now?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.13568686,
+        "p_abstain": 0.86431307
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_043",
+        "text": "what gpu am i using for the reranker right now?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.180166,
+        "p_abstain": 0.819834
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_044",
+        "text": "what is the current system load average?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.106692225,
+        "p_abstain": 0.89330775
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_045",
+        "text": "in klbr-core/src/config.rs, what is router_model_path set to by default?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.24168047,
+        "p_abstain": 0.75831956
+      },
+      {
+        "query_id": "holdout_tools_recall_abstain_053",
+        "text": "fix it",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.11018476,
+        "p_abstain": 0.8898153
+      },
+      {
+        "query_id": "holdout_tools_recall_abstain_054",
+        "text": "今週のHacker Newsのトップ記事って何?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.122447826,
+        "p_abstain": 0.87755215
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_056",
+        "text": "what's the uptime since klbr daemon last started?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.1281884,
+        "p_abstain": 0.8718116
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_058",
+        "text": "what is the default value of compaction_keep in Config::default()?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.19324629,
+        "p_abstain": 0.8067537
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_059",
+        "text": "in klbr-core/src/config.rs, what is memory_top_k set to by default?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.17644644,
+        "p_abstain": 0.82355356
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_060",
+        "text": "in klbr-core/src/llm.rs, which method produces embeddings?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.23829702,
+        "p_abstain": 0.76170295
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_005",
+        "text": "who won the nba game last night? check",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.10403203,
+        "p_abstain": 0.895968
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_007",
+        "text": "search the internet for 'bge-m3 gguf' and summarize the first result",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.14765014,
+        "p_abstain": 0.85234994
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_008",
+        "text": "今週のHacker Newsのトップ記事って何? 調べて",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.11178698,
+        "p_abstain": 0.88821304
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_017",
+        "text": "ssh into the remote server and cat /etc/shadow",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.12404098,
+        "p_abstain": 0.87595904
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_021",
+        "text": "will klbr be popular in 2028?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.18373026,
+        "p_abstain": 0.8162697
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_022",
+        "text": "predict what embedding model will be best in 2030",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.23580031,
+        "p_abstain": 0.76419973
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_023",
+        "text": "will we need a reranker in 2 years?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.31362545,
+        "p_abstain": 0.68637455
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_027",
+        "text": "do the thing from yesterday",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.10365143,
+        "p_abstain": 0.89634854
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_028",
+        "text": "what's the best programming language? prove it with data",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.15280688,
+        "p_abstain": 0.8471931
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_029",
+        "text": "vim or emacs? choose for me",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.16565238,
+        "p_abstain": 0.83434767
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_030",
+        "text": "is gemma 4 better than llama 3.1? check online and decide",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.20931254,
+        "p_abstain": 0.79068744
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_032",
+        "text": "open a browser and check whether llama-server supports embeddings by default",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.19038048,
+        "p_abstain": 0.80961955
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_033",
+        "text": "fetch the gemma-4-26b model card and summarize it",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.17735076,
+        "p_abstain": 0.82264924
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_039",
+        "text": "check the latest Rust RFCs and tell me the most controversial one",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.12441018,
+        "p_abstain": 0.87558985
+      },
+      {
+        "query_id": "holdout_abstain_toolish_tools_042",
+        "text": "which file defines the ClientMsg type in klbr-ipc?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.12457222,
+        "p_abstain": 0.8754278
+      },
+      {
+        "query_id": "holdout_abstain_toolish_tools_044",
+        "text": "write a file docs/router_holdout_abstain_toolish.txt with the content \"ok\"",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "write_file"
+        ],
+        "p_memory": 0.19135803,
+        "p_abstain": 0.8086419
+      },
+      {
+        "query_id": "holdout_abstain_toolish_tools_045",
+        "text": "klbr-core/src/config.rsのmemory_top_kはいくつ?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.126839,
+        "p_abstain": 0.873161
+      },
+      {
+        "query_id": "holdout_abstain_toolish_tools_046",
+        "text": "今メモリ使用量はどれくらい?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.105395846,
+        "p_abstain": 0.89460415
+      },
+      {
+        "query_id": "holdout_abstain_toolish_tools_048",
+        "text": "how big is agent.db right now on disk?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.10135,
+        "p_abstain": 0.89864993
+      },
+      {
+        "query_id": "holdout_abstain_toolish_tools_049",
+        "text": "in klbr-core/src/config.rs, what is llm_url set to?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.20903136,
+        "p_abstain": 0.7909686
+      },
+      {
+        "query_id": "holdout_memory_toolish_tools_043",
+        "text": "which file defines the ClientMsg type in klbr-ipc?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.12457222,
+        "p_abstain": 0.8754278
+      },
+      {
+        "query_id": "holdout_memory_toolish_tools_044",
+        "text": "in klbr-core/src/config.rs, what is memory_top_k set to by default?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.17644644,
+        "p_abstain": 0.82355356
+      },
+      {
+        "query_id": "holdout_memory_toolish_tools_045",
+        "text": "write a file docs/router_holdout_memory_toolish.txt with the content \"ok\"",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "write_file"
+        ],
+        "p_memory": 0.15731505,
+        "p_abstain": 0.842685
+      },
+      {
+        "query_id": "holdout_memory_toolish_tools_046",
+        "text": "今メモリ使用量はどれくらい?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.105395846,
+        "p_abstain": 0.89460415
+      },
+      {
+        "query_id": "holdout_memory_toolish_tools_047",
+        "text": "klbr-core/src/config.rsのembed_dimはいくつ?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.13011032,
+        "p_abstain": 0.8698897
+      },
+      {
+        "query_id": "holdout_memory_toolish_tools_049",
+        "text": "how big is agent.db right now on disk?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.10135,
+        "p_abstain": 0.89864993
+      },
+      {
+        "query_id": "holdout_memory_toolish_tools_050",
+        "text": "in klbr-daemon/src/daemon.rs, does it forward agent errors to the tui?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.13177624,
+        "p_abstain": 0.8682238
+      },
+      {
+        "query_id": "holdout_memory_toolish_abstain_054",
+        "text": "will klbr be popular in 2028?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.18373026,
+        "p_abstain": 0.8162697
+      },
+      {
+        "query_id": "holdout_memory_toolish_abstain_056",
+        "text": "今週のHacker Newsのトップ記事って何?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.122447826,
+        "p_abstain": 0.87755215
+      },
+      {
+        "query_id": "holdout_memory_toolish_abstain_060",
+        "text": "who won the nba game last night? check",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.10403203,
+        "p_abstain": 0.895968
+      },
+      {
+        "query_id": "shell_test_003",
+        "text": "is llama-server currently loaded with gemma or something else?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.17101003,
+        "p_abstain": 0.82899004
+      },
+      {
+        "query_id": "shell_test_005",
+        "text": "how many open file descriptors does klbr-daemon have right now?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.1102257,
+        "p_abstain": 0.8897743
+      },
+      {
+        "query_id": "shell_test_007",
+        "text": "what environment variable is RUST_LOG set to?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.16652043,
+        "p_abstain": 0.8334795
+      },
+      {
+        "query_id": "shell_test_009",
+        "text": "rerankerのプロセスは現在いくつ走ってますか",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.15000203,
+        "p_abstain": 0.849998
+      },
+      {
+        "query_id": "shell_test_010",
+        "text": "what's the uptime since klbr daemon last started",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.120798536,
+        "p_abstain": 0.8792015
+      },
+      {
+        "query_id": "shell_test_011",
+        "text": "how big is agent.db right now on disk?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.10135,
+        "p_abstain": 0.89864993
+      },
+      {
+        "query_id": "shell_test_012",
+        "text": "any nix builds running in background rn?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.124442354,
+        "p_abstain": 0.8755576
+      },
+      {
+        "query_id": "shell_test_013",
+        "text": "what's the current hostname?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.101953685,
+        "p_abstain": 0.8980464
+      },
+      {
+        "query_id": "shell_test_015",
+        "text": "what's the network interface with the highest tx bytes?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.18613501,
+        "p_abstain": 0.81386495
+      },
+      {
+        "query_id": "shell_test_016",
+        "text": "can you check if the reranker http endpoint is responding?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.18152556,
+        "p_abstain": 0.8184745
+      },
+      {
+        "query_id": "shell_test_019",
+        "text": "is the klbr tui currently connected to the daemon socket?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.10980699,
+        "p_abstain": 0.890193
+      },
+      {
+        "query_id": "shell_test_020",
+        "text": "whats the total vram on ths gpu",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.104828686,
+        "p_abstain": 0.8951713
+      },
+      {
+        "query_id": "shell_test_022",
+        "text": "what ip address is localhost:8002 actually bound to?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.15774079,
+        "p_abstain": 0.8422592
+      },
+      {
+        "query_id": "shell_test_027",
+        "text": "is port 1234 open right now?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.10612941,
+        "p_abstain": 0.8938706
+      },
+      {
+        "query_id": "shell_test_028",
+        "text": "what's the size of the current target/ directory?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.1315417,
+        "p_abstain": 0.8684583
+      },
+      {
+        "query_id": "shell_test_030",
+        "text": "journalctl -u klbr — what does it say?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.16173947,
+        "p_abstain": 0.83826053
+      },
+      {
+        "query_id": "shell_test_031",
+        "text": "is the disk partition containing agent.db almost full?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.11036905,
+        "p_abstain": 0.889631
+      },
+      {
+        "query_id": "rf_test_001",
+        "text": "what compaction_keep default is in config.rs right now?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.15488355,
+        "p_abstain": 0.84511644
+      },
+      {
+        "query_id": "rf_test_002",
+        "text": "which file defines the ClientMsg type?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.12701128,
+        "p_abstain": 0.8729887
+      },
+      {
+        "query_id": "rf_test_003",
+        "text": "what does the tools mod.rs export?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.14767623,
+        "p_abstain": 0.8523238
+      },
+      {
+        "query_id": "rf_test_004",
+        "text": "what's the rerank_top_k default in the calibrated config file?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.2759505,
+        "p_abstain": 0.7240495
+      },
+      {
+        "query_id": "rf_test_005",
+        "text": "how many queries are in internal_eval_starter.json?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.17424096,
+        "p_abstain": 0.82575905
+      },
+      {
+        "query_id": "rf_test_006",
+        "text": "config.rsのmemory_support_score_gapのデフォルト値はいくつ?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.14725517,
+        "p_abstain": 0.8527449
+      },
+      {
+        "query_id": "rf_test_007",
+        "text": "is there a router.rs module in klbr-core?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.14385904,
+        "p_abstain": 0.856141
+      },
+      {
+        "query_id": "rf_test_008",
+        "text": "what's the current benchmark README say about the retrieval command?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.15404582,
+        "p_abstain": 0.84595424
+      },
+      {
+        "query_id": "rf_test_010",
+        "text": "what's in the .gitignore for klbr?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.16544971,
+        "p_abstain": 0.83455026
+      },
+      {
+        "query_id": "rf_test_011",
+        "text": "what feature flags does klbr-bench's Cargo.toml declare?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.14107083,
+        "p_abstain": 0.85892916
+      },
+      {
+        "query_id": "rf_test_012",
+        "text": "how many memories does internal_eval_passive_recall_v3.json have?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.16827789,
+        "p_abstain": 0.83172214
+      },
+      {
+        "query_id": "rf_test_013",
+        "text": "what is the embed_dim value in the support calibrated config?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.17114092,
+        "p_abstain": 0.82885903
+      },
+      {
+        "query_id": "rf_test_014",
+        "text": "does flake.nix pin any specific nixpkgs revision?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.12817512,
+        "p_abstain": 0.87182486
+      },
+      {
+        "query_id": "rf_test_015",
+        "text": "what's the passive_support_score_gap set to in the test config?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.21011086,
+        "p_abstain": 0.7898891
+      },
+      {
+        "query_id": "rf_test_017",
+        "text": "what's written in docs/klbr_memory_design.md about L1?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.2108662,
+        "p_abstain": 0.7891338
+      },
+      {
+        "query_id": "rf_test_019",
+        "text": "what does the recall.rs tool do with the tag_and param?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.13939793,
+        "p_abstain": 0.8606021
+      },
+      {
+        "query_id": "rf_test_021",
+        "text": "llm.rsのstreamメソッドのシグネチャは?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.15153,
+        "p_abstain": 0.84847
+      },
+      {
+        "query_id": "rf_test_022",
+        "text": "what's the current threshold in the router_model.json?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.18098755,
+        "p_abstain": 0.81901246
+      },
+      {
+        "query_id": "rf_test_024",
+        "text": "how many memories does the passive recall v3 dataset have?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.16778347,
+        "p_abstain": 0.83221656
+      },
+      {
+        "query_id": "rf_test_026",
+        "text": "what warnings did the last cargo check produce?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file",
+          "shell"
+        ],
+        "p_memory": 0.15090106,
+        "p_abstain": 0.8490989
+      },
+      {
+        "query_id": "rf_test_027",
+        "text": "what's the embed_dim in the router_model.json we trained?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.29247677,
+        "p_abstain": 0.7075233
+      },
+      {
+        "query_id": "rf_test_029",
+        "text": "how is the SweepRange struct defined in klbr-bench main.rs?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.22472288,
+        "p_abstain": 0.7752771
+      },
+      {
+        "query_id": "wf_test_001",
+        "text": "set the compaction_keep to 15 in config.rs",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "write_file"
+        ],
+        "p_memory": 0.11521975,
+        "p_abstain": 0.8847803
+      },
+      {
+        "query_id": "wf_test_002",
+        "text": "write the router precision result to a markdown file",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "write_file"
+        ],
+        "p_memory": 0.18320942,
+        "p_abstain": 0.81679064
+      },
+      {
+        "query_id": "wf_test_003",
+        "text": "create a new slice file with 5 tool-needed examples",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "write_file"
+        ],
+        "p_memory": 0.13537273,
+        "p_abstain": 0.8646273
+      },
+      {
+        "query_id": "wf_test_004",
+        "text": "change the embed_url to localhost:8005 in the config",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "write_file"
+        ],
+        "p_memory": 0.15432762,
+        "p_abstain": 0.84567237
+      },
+      {
+        "query_id": "wf_test_005",
+        "text": "write a tiny benchmarking summary to out-router-iter-2/notes.txt",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "write_file"
+        ],
+        "p_memory": 0.16127193,
+        "p_abstain": 0.83872813
+      },
+      {
+        "query_id": "wf_test_006",
+        "text": "add router_model_path to the default config and set it to a path",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "write_file"
+        ],
+        "p_memory": 0.17621838,
+        "p_abstain": 0.82378167
+      },
+      {
+        "query_id": "wf_test_008",
+        "text": "patch the memory_top_k to 5 in config.rs",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "write_file"
+        ],
+        "p_memory": 0.16184351,
+        "p_abstain": 0.8381565
+      },
+      {
+        "query_id": "wf_test_009",
+        "text": "update the dataset_id in internal_eval_starter.json to v3",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "write_file"
+        ],
+        "p_memory": 0.13346773,
+        "p_abstain": 0.86653227
+      },
+      {
+        "query_id": "wf_test_010",
+        "text": "write the sweep results summary to a txt file in the output dir",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "write_file"
+        ],
+        "p_memory": 0.11803409,
+        "p_abstain": 0.88196594
+      },
+      {
+        "query_id": "wf_test_011",
+        "text": "save a note about the router precision gate to docs/",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "write_file"
+        ],
+        "p_memory": 0.119401686,
+        "p_abstain": 0.88059837
+      },
+      {
+        "query_id": "wf_test_012",
+        "text": "create a quick script to run all benchmarks in sequence",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "write_file",
+          "shell"
+        ],
+        "p_memory": 0.11596763,
+        "p_abstain": 0.88403237
+      },
+      {
+        "query_id": "wf_test_013",
+        "text": "add a line to .gitignore to exclude agent.db",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "write_file"
+        ],
+        "p_memory": 0.16403273,
+        "p_abstain": 0.83596724
+      },
+      {
+        "query_id": "wf_test_014",
+        "text": "新しいスライスJSONを作ってmemory_laneのクエリを10個追加して",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "write_file"
+        ],
+        "p_memory": 0.13619125,
+        "p_abstain": 0.8638087
+      },
+      {
+        "query_id": "wf_test_015",
+        "text": "bump the rerank_top_k to 15 in the calibrated config and save it",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "write_file"
+        ],
+        "p_memory": 0.1611398,
+        "p_abstain": 0.8388602
+      },
+      {
+        "query_id": "abs_test_003",
+        "text": "what will my project look like in 6 months?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.15345076,
+        "p_abstain": 0.8465493
+      },
+      {
+        "query_id": "abs_test_004",
+        "text": "what's the best programming language?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.15857735,
+        "p_abstain": 0.8414227
+      },
+      {
+        "query_id": "abs_test_006",
+        "text": "what's the github token for the ci runner?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.10543426,
+        "p_abstain": 0.89456576
+      },
+      {
+        "query_id": "abs_test_007",
+        "text": "vim or emacs?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.20100525,
+        "p_abstain": 0.7989947
+      },
+      {
+        "query_id": "abs_test_009",
+        "text": "どのモデルが今一番スコアが高いですか — ベンチマークサイトで",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.14664401,
+        "p_abstain": 0.853356
+      },
+      {
+        "query_id": "abs_test_010",
+        "text": "when will the klbr project be finished?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.1696385,
+        "p_abstain": 0.8303615
+      },
+      {
+        "query_id": "abs_test_011",
+        "text": "is nomic-embed or bge-m3 better for this use case?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.27238378,
+        "p_abstain": 0.72761625
+      },
+      {
+        "query_id": "abs_test_012",
+        "text": "y",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.1297662,
+        "p_abstain": 0.8702338
+      },
+      {
+        "query_id": "abs_test_013",
+        "text": "is the llama.cpp repo trending on github today?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.14317971,
+        "p_abstain": 0.8568203
+      },
+      {
+        "query_id": "abs_test_014",
+        "text": "how should klbr evolve to support multi-user?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.2298695,
+        "p_abstain": 0.7701305
+      }
+    ]
+  },
+  "holdout_metrics": {
+    "total": 180,
+    "accuracy": 0.62777776,
+    "confusion": [
+      [
+        55,
+        0
+      ],
+      [
+        67,
+        58
+      ]
+    ],
+    "memory_false_abstain_rate": 0.0,
+    "abstain_false_memory_rate": 0.536,
+    "tool_required_false_memory_rate": 0.6857143,
+    "tool_required_recall": 0.31428573,
+    "memory_count": 55,
+    "abstain_count": 125,
+    "tool_required_count": 70,
+    "misclassified": [
+      {
+        "query_id": "holdout_tools_recall_tools_002",
+        "text": "how much free disk space is there on the partition containing agent.db?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.10456361,
+        "p_abstain": 0.89543635
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_003",
+        "text": "what's the current hostname?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.101953685,
+        "p_abstain": 0.8980464
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_004",
+        "text": "is port 8002 bound to 127.0.0.1 or 0.0.0.0 right now?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.1582868,
+        "p_abstain": 0.8417132
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_006",
+        "text": "are there any zombie processes named klbr right now?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.11492517,
+        "p_abstain": 0.8850748
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_008",
+        "text": "check if the reranker http endpoint on localhost:8003 is responding",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.16735494,
+        "p_abstain": 0.8326451
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_011",
+        "text": "which file defines the ServerMsg type?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.12588543,
+        "p_abstain": 0.8741145
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_013",
+        "text": "read klbr-core/src/agent.rs and summarize how routing affects memory injection",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.17401817,
+        "p_abstain": 0.82598186
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_014",
+        "text": "what does klbr-core/src/context.rs do when recalled memories are present?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.14702685,
+        "p_abstain": 0.85297316
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_015",
+        "text": "what crate defines the LlmClient type?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.23538512,
+        "p_abstain": 0.7646149
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_016",
+        "text": "write a file docs/router_holdout_tools_recall.txt with the content \"ok\"",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "write_file"
+        ],
+        "p_memory": 0.14410025,
+        "p_abstain": 0.8558998
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_017",
+        "text": "append a line \"router ok\" to docs/router_holdout_log.txt",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "write_file"
+        ],
+        "p_memory": 0.11356302,
+        "p_abstain": 0.88643694
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_018",
+        "text": "create a file benchmarks/router_holdout_tmp.txt and put the number 123 in it",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "write_file"
+        ],
+        "p_memory": 0.13842863,
+        "p_abstain": 0.8615714
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_019",
+        "text": "今ポート8002は開いてる?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.10369836,
+        "p_abstain": 0.8963016
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_020",
+        "text": "klbr-core/src/llm.rsのembed関数って何返すの?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.16805679,
+        "p_abstain": 0.8319432
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_023",
+        "text": "in klbr-core/src/config.rs, what is memory_support_score_gap set to?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.15972814,
+        "p_abstain": 0.8402719
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_024",
+        "text": "how many open file descriptors does the klbr-daemon process have right now?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.105587445,
+        "p_abstain": 0.8944126
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_025",
+        "text": "what's the size of the current target/ directory?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.1315417,
+        "p_abstain": 0.8684583
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_026",
+        "text": "what ip address is localhost:8002 actually bound to right now?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.1407521,
+        "p_abstain": 0.8592479
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_027",
+        "text": "which file defines the ClientMsg type in the ipc protocol?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.11255213,
+        "p_abstain": 0.88744783
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_028",
+        "text": "in klbr-core/src/router.rs, what routing decisions exist?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.20984548,
+        "p_abstain": 0.7901545
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_032",
+        "text": "open klbr-core/src/config.rs and tell me what llm_url is set to",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.11325379,
+        "p_abstain": 0.8867462
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_033",
+        "text": "open klbr-core/src/config.rs and tell me what rerank_url is set to",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.16435155,
+        "p_abstain": 0.8356485
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_034",
+        "text": "in klbr-daemon/src/daemon.rs, does it forward agent errors to the tui?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.13177624,
+        "p_abstain": 0.8682238
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_035",
+        "text": "in klbr-tui/src/main.rs, what happens when ServerMsg::Error arrives?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.11597728,
+        "p_abstain": 0.8840227
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_038",
+        "text": "is the disk partition containing /home close to full right now?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.11078995,
+        "p_abstain": 0.88921
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_039",
+        "text": "what is the network interface with the highest tx bytes right now?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.14258735,
+        "p_abstain": 0.85741264
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_040",
+        "text": "is port 1234 open right now?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.10612941,
+        "p_abstain": 0.8938706
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_041",
+        "text": "what environment variable is RUST_LOG set to right now?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.13568686,
+        "p_abstain": 0.86431307
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_043",
+        "text": "what gpu am i using for the reranker right now?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.180166,
+        "p_abstain": 0.819834
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_044",
+        "text": "what is the current system load average?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.106692225,
+        "p_abstain": 0.89330775
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_045",
+        "text": "in klbr-core/src/config.rs, what is router_model_path set to by default?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.24168047,
+        "p_abstain": 0.75831956
+      },
+      {
+        "query_id": "holdout_tools_recall_abstain_053",
+        "text": "fix it",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.11018476,
+        "p_abstain": 0.8898153
+      },
+      {
+        "query_id": "holdout_tools_recall_abstain_054",
+        "text": "今週のHacker Newsのトップ記事って何?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.122447826,
+        "p_abstain": 0.87755215
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_056",
+        "text": "what's the uptime since klbr daemon last started?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.1281884,
+        "p_abstain": 0.8718116
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_058",
+        "text": "what is the default value of compaction_keep in Config::default()?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.19324629,
+        "p_abstain": 0.8067537
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_059",
+        "text": "in klbr-core/src/config.rs, what is memory_top_k set to by default?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.17644644,
+        "p_abstain": 0.82355356
+      },
+      {
+        "query_id": "holdout_tools_recall_tools_060",
+        "text": "in klbr-core/src/llm.rs, which method produces embeddings?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.23829702,
+        "p_abstain": 0.76170295
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_005",
+        "text": "who won the nba game last night? check",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.10403203,
+        "p_abstain": 0.895968
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_007",
+        "text": "search the internet for 'bge-m3 gguf' and summarize the first result",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.14765014,
+        "p_abstain": 0.85234994
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_008",
+        "text": "今週のHacker Newsのトップ記事って何? 調べて",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.11178698,
+        "p_abstain": 0.88821304
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_017",
+        "text": "ssh into the remote server and cat /etc/shadow",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.12404098,
+        "p_abstain": 0.87595904
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_021",
+        "text": "will klbr be popular in 2028?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.18373026,
+        "p_abstain": 0.8162697
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_022",
+        "text": "predict what embedding model will be best in 2030",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.23580031,
+        "p_abstain": 0.76419973
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_023",
+        "text": "will we need a reranker in 2 years?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.31362545,
+        "p_abstain": 0.68637455
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_027",
+        "text": "do the thing from yesterday",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.10365143,
+        "p_abstain": 0.89634854
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_028",
+        "text": "what's the best programming language? prove it with data",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.15280688,
+        "p_abstain": 0.8471931
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_029",
+        "text": "vim or emacs? choose for me",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.16565238,
+        "p_abstain": 0.83434767
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_030",
+        "text": "is gemma 4 better than llama 3.1? check online and decide",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.20931254,
+        "p_abstain": 0.79068744
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_032",
+        "text": "open a browser and check whether llama-server supports embeddings by default",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.19038048,
+        "p_abstain": 0.80961955
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_033",
+        "text": "fetch the gemma-4-26b model card and summarize it",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.17735076,
+        "p_abstain": 0.82264924
+      },
+      {
+        "query_id": "holdout_abstain_toolish_abstain_039",
+        "text": "check the latest Rust RFCs and tell me the most controversial one",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.12441018,
+        "p_abstain": 0.87558985
+      },
+      {
+        "query_id": "holdout_abstain_toolish_tools_042",
+        "text": "which file defines the ClientMsg type in klbr-ipc?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.12457222,
+        "p_abstain": 0.8754278
+      },
+      {
+        "query_id": "holdout_abstain_toolish_tools_044",
+        "text": "write a file docs/router_holdout_abstain_toolish.txt with the content \"ok\"",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "write_file"
+        ],
+        "p_memory": 0.19135803,
+        "p_abstain": 0.8086419
+      },
+      {
+        "query_id": "holdout_abstain_toolish_tools_045",
+        "text": "klbr-core/src/config.rsのmemory_top_kはいくつ?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.126839,
+        "p_abstain": 0.873161
+      },
+      {
+        "query_id": "holdout_abstain_toolish_tools_046",
+        "text": "今メモリ使用量はどれくらい?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.105395846,
+        "p_abstain": 0.89460415
+      },
+      {
+        "query_id": "holdout_abstain_toolish_tools_048",
+        "text": "how big is agent.db right now on disk?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.10135,
+        "p_abstain": 0.89864993
+      },
+      {
+        "query_id": "holdout_abstain_toolish_tools_049",
+        "text": "in klbr-core/src/config.rs, what is llm_url set to?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.20903136,
+        "p_abstain": 0.7909686
+      },
+      {
+        "query_id": "holdout_memory_toolish_tools_043",
+        "text": "which file defines the ClientMsg type in klbr-ipc?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.12457222,
+        "p_abstain": 0.8754278
+      },
+      {
+        "query_id": "holdout_memory_toolish_tools_044",
+        "text": "in klbr-core/src/config.rs, what is memory_top_k set to by default?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.17644644,
+        "p_abstain": 0.82355356
+      },
+      {
+        "query_id": "holdout_memory_toolish_tools_045",
+        "text": "write a file docs/router_holdout_memory_toolish.txt with the content \"ok\"",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "write_file"
+        ],
+        "p_memory": 0.15731505,
+        "p_abstain": 0.842685
+      },
+      {
+        "query_id": "holdout_memory_toolish_tools_046",
+        "text": "今メモリ使用量はどれくらい?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.105395846,
+        "p_abstain": 0.89460415
+      },
+      {
+        "query_id": "holdout_memory_toolish_tools_047",
+        "text": "klbr-core/src/config.rsのembed_dimはいくつ?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.13011032,
+        "p_abstain": 0.8698897
+      },
+      {
+        "query_id": "holdout_memory_toolish_tools_049",
+        "text": "how big is agent.db right now on disk?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "shell"
+        ],
+        "p_memory": 0.10135,
+        "p_abstain": 0.89864993
+      },
+      {
+        "query_id": "holdout_memory_toolish_tools_050",
+        "text": "in klbr-daemon/src/daemon.rs, does it forward agent errors to the tui?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [
+          "read_file"
+        ],
+        "p_memory": 0.13177624,
+        "p_abstain": 0.8682238
+      },
+      {
+        "query_id": "holdout_memory_toolish_abstain_054",
+        "text": "will klbr be popular in 2028?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.18373026,
+        "p_abstain": 0.8162697
+      },
+      {
+        "query_id": "holdout_memory_toolish_abstain_056",
+        "text": "今週のHacker Newsのトップ記事って何?",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.122447826,
+        "p_abstain": 0.87755215
+      },
+      {
+        "query_id": "holdout_memory_toolish_abstain_060",
+        "text": "who won the nba game last night? check",
+        "gold": "abstain",
+        "pred": "memory",
+        "required_tools": [],
+        "p_memory": 0.10403203,
+        "p_abstain": 0.895968
+      }
+    ]
+  },
+  "holdout_metrics_by_bucket": {
+    "abstain_toolish": {
+      "total": 60,
+      "accuracy": 0.6666667,
+      "confusion": [
+        [
+          10,
+          0
+        ],
+        [
+          20,
+          30
+        ]
+      ],
+      "memory_false_abstain_rate": 0.0,
+      "abstain_false_memory_rate": 0.4,
+      "tool_required_false_memory_rate": 0.6,
+      "tool_required_recall": 0.4,
+      "memory_count": 10,
+      "abstain_count": 50,
+      "tool_required_count": 10,
+      "misclassified": [
+        {
+          "query_id": "holdout_abstain_toolish_abstain_005",
+          "text": "who won the nba game last night? check",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [],
+          "p_memory": 0.10403203,
+          "p_abstain": 0.895968
+        },
+        {
+          "query_id": "holdout_abstain_toolish_abstain_007",
+          "text": "search the internet for 'bge-m3 gguf' and summarize the first result",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [],
+          "p_memory": 0.14765014,
+          "p_abstain": 0.85234994
+        },
+        {
+          "query_id": "holdout_abstain_toolish_abstain_008",
+          "text": "今週のHacker Newsのトップ記事って何? 調べて",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [],
+          "p_memory": 0.11178698,
+          "p_abstain": 0.88821304
+        },
+        {
+          "query_id": "holdout_abstain_toolish_abstain_017",
+          "text": "ssh into the remote server and cat /etc/shadow",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [],
+          "p_memory": 0.12404098,
+          "p_abstain": 0.87595904
+        },
+        {
+          "query_id": "holdout_abstain_toolish_abstain_021",
+          "text": "will klbr be popular in 2028?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [],
+          "p_memory": 0.18373026,
+          "p_abstain": 0.8162697
+        },
+        {
+          "query_id": "holdout_abstain_toolish_abstain_022",
+          "text": "predict what embedding model will be best in 2030",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [],
+          "p_memory": 0.23580031,
+          "p_abstain": 0.76419973
+        },
+        {
+          "query_id": "holdout_abstain_toolish_abstain_023",
+          "text": "will we need a reranker in 2 years?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [],
+          "p_memory": 0.31362545,
+          "p_abstain": 0.68637455
+        },
+        {
+          "query_id": "holdout_abstain_toolish_abstain_027",
+          "text": "do the thing from yesterday",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [],
+          "p_memory": 0.10365143,
+          "p_abstain": 0.89634854
+        },
+        {
+          "query_id": "holdout_abstain_toolish_abstain_028",
+          "text": "what's the best programming language? prove it with data",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [],
+          "p_memory": 0.15280688,
+          "p_abstain": 0.8471931
+        },
+        {
+          "query_id": "holdout_abstain_toolish_abstain_029",
+          "text": "vim or emacs? choose for me",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [],
+          "p_memory": 0.16565238,
+          "p_abstain": 0.83434767
+        },
+        {
+          "query_id": "holdout_abstain_toolish_abstain_030",
+          "text": "is gemma 4 better than llama 3.1? check online and decide",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [],
+          "p_memory": 0.20931254,
+          "p_abstain": 0.79068744
+        },
+        {
+          "query_id": "holdout_abstain_toolish_abstain_032",
+          "text": "open a browser and check whether llama-server supports embeddings by default",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [],
+          "p_memory": 0.19038048,
+          "p_abstain": 0.80961955
+        },
+        {
+          "query_id": "holdout_abstain_toolish_abstain_033",
+          "text": "fetch the gemma-4-26b model card and summarize it",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [],
+          "p_memory": 0.17735076,
+          "p_abstain": 0.82264924
+        },
+        {
+          "query_id": "holdout_abstain_toolish_abstain_039",
+          "text": "check the latest Rust RFCs and tell me the most controversial one",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [],
+          "p_memory": 0.12441018,
+          "p_abstain": 0.87558985
+        },
+        {
+          "query_id": "holdout_abstain_toolish_tools_042",
+          "text": "which file defines the ClientMsg type in klbr-ipc?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "read_file"
+          ],
+          "p_memory": 0.12457222,
+          "p_abstain": 0.8754278
+        },
+        {
+          "query_id": "holdout_abstain_toolish_tools_044",
+          "text": "write a file docs/router_holdout_abstain_toolish.txt with the content \"ok\"",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "write_file"
+          ],
+          "p_memory": 0.19135803,
+          "p_abstain": 0.8086419
+        },
+        {
+          "query_id": "holdout_abstain_toolish_tools_045",
+          "text": "klbr-core/src/config.rsのmemory_top_kはいくつ?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "read_file"
+          ],
+          "p_memory": 0.126839,
+          "p_abstain": 0.873161
+        },
+        {
+          "query_id": "holdout_abstain_toolish_tools_046",
+          "text": "今メモリ使用量はどれくらい?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "shell"
+          ],
+          "p_memory": 0.105395846,
+          "p_abstain": 0.89460415
+        },
+        {
+          "query_id": "holdout_abstain_toolish_tools_048",
+          "text": "how big is agent.db right now on disk?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "shell"
+          ],
+          "p_memory": 0.10135,
+          "p_abstain": 0.89864993
+        },
+        {
+          "query_id": "holdout_abstain_toolish_tools_049",
+          "text": "in klbr-core/src/config.rs, what is llm_url set to?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "read_file"
+          ],
+          "p_memory": 0.20903136,
+          "p_abstain": 0.7909686
+        }
+      ]
+    },
+    "memory_toolish": {
+      "total": 60,
+      "accuracy": 0.8333333,
+      "confusion": [
+        [
+          40,
+          0
+        ],
+        [
+          10,
+          10
+        ]
+      ],
+      "memory_false_abstain_rate": 0.0,
+      "abstain_false_memory_rate": 0.5,
+      "tool_required_false_memory_rate": 0.7,
+      "tool_required_recall": 0.3,
+      "memory_count": 40,
+      "abstain_count": 20,
+      "tool_required_count": 10,
+      "misclassified": [
+        {
+          "query_id": "holdout_memory_toolish_tools_043",
+          "text": "which file defines the ClientMsg type in klbr-ipc?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "read_file"
+          ],
+          "p_memory": 0.12457222,
+          "p_abstain": 0.8754278
+        },
+        {
+          "query_id": "holdout_memory_toolish_tools_044",
+          "text": "in klbr-core/src/config.rs, what is memory_top_k set to by default?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "read_file"
+          ],
+          "p_memory": 0.17644644,
+          "p_abstain": 0.82355356
+        },
+        {
+          "query_id": "holdout_memory_toolish_tools_045",
+          "text": "write a file docs/router_holdout_memory_toolish.txt with the content \"ok\"",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "write_file"
+          ],
+          "p_memory": 0.15731505,
+          "p_abstain": 0.842685
+        },
+        {
+          "query_id": "holdout_memory_toolish_tools_046",
+          "text": "今メモリ使用量はどれくらい?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "shell"
+          ],
+          "p_memory": 0.105395846,
+          "p_abstain": 0.89460415
+        },
+        {
+          "query_id": "holdout_memory_toolish_tools_047",
+          "text": "klbr-core/src/config.rsのembed_dimはいくつ?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "read_file"
+          ],
+          "p_memory": 0.13011032,
+          "p_abstain": 0.8698897
+        },
+        {
+          "query_id": "holdout_memory_toolish_tools_049",
+          "text": "how big is agent.db right now on disk?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "shell"
+          ],
+          "p_memory": 0.10135,
+          "p_abstain": 0.89864993
+        },
+        {
+          "query_id": "holdout_memory_toolish_tools_050",
+          "text": "in klbr-daemon/src/daemon.rs, does it forward agent errors to the tui?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "read_file"
+          ],
+          "p_memory": 0.13177624,
+          "p_abstain": 0.8682238
+        },
+        {
+          "query_id": "holdout_memory_toolish_abstain_054",
+          "text": "will klbr be popular in 2028?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [],
+          "p_memory": 0.18373026,
+          "p_abstain": 0.8162697
+        },
+        {
+          "query_id": "holdout_memory_toolish_abstain_056",
+          "text": "今週のHacker Newsのトップ記事って何?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [],
+          "p_memory": 0.122447826,
+          "p_abstain": 0.87755215
+        },
+        {
+          "query_id": "holdout_memory_toolish_abstain_060",
+          "text": "who won the nba game last night? check",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [],
+          "p_memory": 0.10403203,
+          "p_abstain": 0.895968
+        }
+      ]
+    },
+    "tools_recall": {
+      "total": 60,
+      "accuracy": 0.38333333,
+      "confusion": [
+        [
+          5,
+          0
+        ],
+        [
+          37,
+          18
+        ]
+      ],
+      "memory_false_abstain_rate": 0.0,
+      "abstain_false_memory_rate": 0.6727273,
+      "tool_required_false_memory_rate": 0.7,
+      "tool_required_recall": 0.3,
+      "memory_count": 5,
+      "abstain_count": 55,
+      "tool_required_count": 50,
+      "misclassified": [
+        {
+          "query_id": "holdout_tools_recall_tools_002",
+          "text": "how much free disk space is there on the partition containing agent.db?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "shell"
+          ],
+          "p_memory": 0.10456361,
+          "p_abstain": 0.89543635
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_003",
+          "text": "what's the current hostname?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "shell"
+          ],
+          "p_memory": 0.101953685,
+          "p_abstain": 0.8980464
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_004",
+          "text": "is port 8002 bound to 127.0.0.1 or 0.0.0.0 right now?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "shell"
+          ],
+          "p_memory": 0.1582868,
+          "p_abstain": 0.8417132
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_006",
+          "text": "are there any zombie processes named klbr right now?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "shell"
+          ],
+          "p_memory": 0.11492517,
+          "p_abstain": 0.8850748
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_008",
+          "text": "check if the reranker http endpoint on localhost:8003 is responding",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "shell"
+          ],
+          "p_memory": 0.16735494,
+          "p_abstain": 0.8326451
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_011",
+          "text": "which file defines the ServerMsg type?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "read_file"
+          ],
+          "p_memory": 0.12588543,
+          "p_abstain": 0.8741145
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_013",
+          "text": "read klbr-core/src/agent.rs and summarize how routing affects memory injection",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "read_file"
+          ],
+          "p_memory": 0.17401817,
+          "p_abstain": 0.82598186
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_014",
+          "text": "what does klbr-core/src/context.rs do when recalled memories are present?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "read_file"
+          ],
+          "p_memory": 0.14702685,
+          "p_abstain": 0.85297316
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_015",
+          "text": "what crate defines the LlmClient type?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "read_file"
+          ],
+          "p_memory": 0.23538512,
+          "p_abstain": 0.7646149
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_016",
+          "text": "write a file docs/router_holdout_tools_recall.txt with the content \"ok\"",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "write_file"
+          ],
+          "p_memory": 0.14410025,
+          "p_abstain": 0.8558998
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_017",
+          "text": "append a line \"router ok\" to docs/router_holdout_log.txt",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "write_file"
+          ],
+          "p_memory": 0.11356302,
+          "p_abstain": 0.88643694
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_018",
+          "text": "create a file benchmarks/router_holdout_tmp.txt and put the number 123 in it",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "write_file"
+          ],
+          "p_memory": 0.13842863,
+          "p_abstain": 0.8615714
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_019",
+          "text": "今ポート8002は開いてる?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "shell"
+          ],
+          "p_memory": 0.10369836,
+          "p_abstain": 0.8963016
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_020",
+          "text": "klbr-core/src/llm.rsのembed関数って何返すの?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "read_file"
+          ],
+          "p_memory": 0.16805679,
+          "p_abstain": 0.8319432
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_023",
+          "text": "in klbr-core/src/config.rs, what is memory_support_score_gap set to?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "read_file"
+          ],
+          "p_memory": 0.15972814,
+          "p_abstain": 0.8402719
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_024",
+          "text": "how many open file descriptors does the klbr-daemon process have right now?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "shell"
+          ],
+          "p_memory": 0.105587445,
+          "p_abstain": 0.8944126
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_025",
+          "text": "what's the size of the current target/ directory?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "shell"
+          ],
+          "p_memory": 0.1315417,
+          "p_abstain": 0.8684583
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_026",
+          "text": "what ip address is localhost:8002 actually bound to right now?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "shell"
+          ],
+          "p_memory": 0.1407521,
+          "p_abstain": 0.8592479
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_027",
+          "text": "which file defines the ClientMsg type in the ipc protocol?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "read_file"
+          ],
+          "p_memory": 0.11255213,
+          "p_abstain": 0.88744783
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_028",
+          "text": "in klbr-core/src/router.rs, what routing decisions exist?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "read_file"
+          ],
+          "p_memory": 0.20984548,
+          "p_abstain": 0.7901545
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_032",
+          "text": "open klbr-core/src/config.rs and tell me what llm_url is set to",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "read_file"
+          ],
+          "p_memory": 0.11325379,
+          "p_abstain": 0.8867462
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_033",
+          "text": "open klbr-core/src/config.rs and tell me what rerank_url is set to",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "read_file"
+          ],
+          "p_memory": 0.16435155,
+          "p_abstain": 0.8356485
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_034",
+          "text": "in klbr-daemon/src/daemon.rs, does it forward agent errors to the tui?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "read_file"
+          ],
+          "p_memory": 0.13177624,
+          "p_abstain": 0.8682238
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_035",
+          "text": "in klbr-tui/src/main.rs, what happens when ServerMsg::Error arrives?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "read_file"
+          ],
+          "p_memory": 0.11597728,
+          "p_abstain": 0.8840227
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_038",
+          "text": "is the disk partition containing /home close to full right now?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "shell"
+          ],
+          "p_memory": 0.11078995,
+          "p_abstain": 0.88921
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_039",
+          "text": "what is the network interface with the highest tx bytes right now?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "shell"
+          ],
+          "p_memory": 0.14258735,
+          "p_abstain": 0.85741264
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_040",
+          "text": "is port 1234 open right now?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "shell"
+          ],
+          "p_memory": 0.10612941,
+          "p_abstain": 0.8938706
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_041",
+          "text": "what environment variable is RUST_LOG set to right now?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "shell"
+          ],
+          "p_memory": 0.13568686,
+          "p_abstain": 0.86431307
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_043",
+          "text": "what gpu am i using for the reranker right now?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "shell"
+          ],
+          "p_memory": 0.180166,
+          "p_abstain": 0.819834
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_044",
+          "text": "what is the current system load average?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "shell"
+          ],
+          "p_memory": 0.106692225,
+          "p_abstain": 0.89330775
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_045",
+          "text": "in klbr-core/src/config.rs, what is router_model_path set to by default?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "read_file"
+          ],
+          "p_memory": 0.24168047,
+          "p_abstain": 0.75831956
+        },
+        {
+          "query_id": "holdout_tools_recall_abstain_053",
+          "text": "fix it",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [],
+          "p_memory": 0.11018476,
+          "p_abstain": 0.8898153
+        },
+        {
+          "query_id": "holdout_tools_recall_abstain_054",
+          "text": "今週のHacker Newsのトップ記事って何?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [],
+          "p_memory": 0.122447826,
+          "p_abstain": 0.87755215
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_056",
+          "text": "what's the uptime since klbr daemon last started?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "shell"
+          ],
+          "p_memory": 0.1281884,
+          "p_abstain": 0.8718116
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_058",
+          "text": "what is the default value of compaction_keep in Config::default()?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "read_file"
+          ],
+          "p_memory": 0.19324629,
+          "p_abstain": 0.8067537
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_059",
+          "text": "in klbr-core/src/config.rs, what is memory_top_k set to by default?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "read_file"
+          ],
+          "p_memory": 0.17644644,
+          "p_abstain": 0.82355356
+        },
+        {
+          "query_id": "holdout_tools_recall_tools_060",
+          "text": "in klbr-core/src/llm.rs, which method produces embeddings?",
+          "gold": "abstain",
+          "pred": "memory",
+          "required_tools": [
+            "read_file"
+          ],
+          "p_memory": 0.23829702,
+          "p_abstain": 0.76170295
+        }
+      ]
+    }
+  }
+}
\ No newline at end of file
diff --git a/benchmarks/models/router/linear/out-router-linear-iter-15/router_report.md b/benchmarks/models/router/linear/out-router-linear-iter-15/router_report.md
new file mode 100644
index 0000000..3cf6ff1
--- /dev/null
+++ b/benchmarks/models/router/linear/out-router-linear-iter-15/router_report.md
@@ -0,0 +1,184 @@
+# Linear Router Report
+
+dataset: `router-multi(internal-eval-starter-v2+router-train-pack-v1+router-dev-pack-v1+router-abstain-v2+router-abstain-v3+router-memory-lane-v3+router-read-file-repo-v3+router-read-file-repo-v4+router-shell-state-v3+router-holdout-tools-recall-v1+router-holdout-abstain-toolish-v1+router-holdout-memory-toolish-v1+router-shell-state-v1+router-shell-state-v1b+router-shell-state-v2+router-read-file-repo-v1+router-read-file-repo-v1b+router-write-file-v1+router-memory-lane-v1+router-memory-lane-v1b+router-abstain-v1+router-memory-adversarial-v1)`
+embed_model: `bge-m3-FP16.gguf`
+trained_on: `Train`
+evaluated_on: `Test`
+
+## Decision Params
+
+- abstain_prob_threshold: `0.900`
+- abstain_margin_threshold: `0.000`
+
+## Metrics
+
+- labeled queries evaluated: `319`
+- memory: `100`  abstain: `219`
+- accuracy: `0.5862`
+- memory false-abstain rate (memory→abstain): `0.0000`
+- abstain false-memory rate (abstain→memory): `0.6027`
+- tool-required queries: `149`
+- tool-required false-memory rate (tool→memory): `0.6913`
+- tool-required recall (tool→abstain): `0.3087`
+
+## Confusion Matrix (gold x pred)
+
+| gold \ pred | memory | abstain |
+|---|---:|---:|
+| memory | 100 | 0 |
+| abstain | 132 | 87 |
+
+## Misclassified Queries
+
+| id | text | gold | pred | tools | p_memory | p_abstain |
+|---|---|---|---|---|---:|---:|
+| `test_q9` | what gpu am i using for the reranker? | `Abstain` | `Memory` | `shell` | 0.20 | 0.80 |
+| `test_q15` | which postgres extension are we enabling before passive recall? | `Abstain` | `Memory` | `read_file` | 0.17 | 0.83 |
+| `holdout_tools_recall_tools_002` | how much free disk space is there on the partition containing agent.db? | `Abstain` | `Memory` | `shell` | 0.10 | 0.90 |
+| `holdout_tools_recall_tools_003` | what's the current hostname? | `Abstain` | `Memory` | `shell` | 0.10 | 0.90 |
+| `holdout_tools_recall_tools_004` | is port 8002 bound to 127.0.0.1 or 0.0.0.0 right now? | `Abstain` | `Memory` | `shell` | 0.16 | 0.84 |
+| `holdout_tools_recall_tools_006` | are there any zombie processes named klbr right now? | `Abstain` | `Memory` | `shell` | 0.11 | 0.89 |
+| `holdout_tools_recall_tools_008` | check if the reranker http endpoint on localhost:8003 is responding | `Abstain` | `Memory` | `shell` | 0.17 | 0.83 |
+| `holdout_tools_recall_tools_011` | which file defines the ServerMsg type? | `Abstain` | `Memory` | `read_file` | 0.13 | 0.87 |
+| `holdout_tools_recall_tools_013` | read klbr-core/src/agent.rs and summarize how routing affects memory injection | `Abstain` | `Memory` | `read_file` | 0.17 | 0.83 |
+| `holdout_tools_recall_tools_014` | what does klbr-core/src/context.rs do when recalled memories are present? | `Abstain` | `Memory` | `read_file` | 0.15 | 0.85 |
+| `holdout_tools_recall_tools_015` | what crate defines the LlmClient type? | `Abstain` | `Memory` | `read_file` | 0.24 | 0.76 |
+| `holdout_tools_recall_tools_016` | write a file docs/router_holdout_tools_recall.txt with the content "ok" | `Abstain` | `Memory` | `write_file` | 0.14 | 0.86 |
+| `holdout_tools_recall_tools_017` | append a line "router ok" to docs/router_holdout_log.txt | `Abstain` | `Memory` | `write_file` | 0.11 | 0.89 |
+| `holdout_tools_recall_tools_018` | create a file benchmarks/router_holdout_tmp.txt and put the number 123 in it | `Abstain` | `Memory` | `write_file` | 0.14 | 0.86 |
+| `holdout_tools_recall_tools_019` | 今ポート8002は開いてる? | `Abstain` | `Memory` | `shell` | 0.10 | 0.90 |
+| `holdout_tools_recall_tools_020` | klbr-core/src/llm.rsのembed関数って何返すの? | `Abstain` | `Memory` | `read_file` | 0.17 | 0.83 |
+| `holdout_tools_recall_tools_023` | in klbr-core/src/config.rs, what is memory_support_score_gap set to? | `Abstain` | `Memory` | `read_file` | 0.16 | 0.84 |
+| `holdout_tools_recall_tools_024` | how many open file descriptors does the klbr-daemon process have right now? | `Abstain` | `Memory` | `shell` | 0.11 | 0.89 |
+| `holdout_tools_recall_tools_025` | what's the size of the current target/ directory? | `Abstain` | `Memory` | `shell` | 0.13 | 0.87 |
+| `holdout_tools_recall_tools_026` | what ip address is localhost:8002 actually bound to right now? | `Abstain` | `Memory` | `shell` | 0.14 | 0.86 |
+| `holdout_tools_recall_tools_027` | which file defines the ClientMsg type in the ipc protocol? | `Abstain` | `Memory` | `read_file` | 0.11 | 0.89 |
+| `holdout_tools_recall_tools_028` | in klbr-core/src/router.rs, what routing decisions exist? | `Abstain` | `Memory` | `read_file` | 0.21 | 0.79 |
+| `holdout_tools_recall_tools_032` | open klbr-core/src/config.rs and tell me what llm_url is set to | `Abstain` | `Memory` | `read_file` | 0.11 | 0.89 |
+| `holdout_tools_recall_tools_033` | open klbr-core/src/config.rs and tell me what rerank_url is set to | `Abstain` | `Memory` | `read_file` | 0.16 | 0.84 |
+| `holdout_tools_recall_tools_034` | in klbr-daemon/src/daemon.rs, does it forward agent errors to the tui? | `Abstain` | `Memory` | `read_file` | 0.13 | 0.87 |
+| `holdout_tools_recall_tools_035` | in klbr-tui/src/main.rs, what happens when ServerMsg::Error arrives? | `Abstain` | `Memory` | `read_file` | 0.12 | 0.88 |
+| `holdout_tools_recall_tools_038` | is the disk partition containing /home close to full right now? | `Abstain` | `Memory` | `shell` | 0.11 | 0.89 |
+| `holdout_tools_recall_tools_039` | what is the network interface with the highest tx bytes right now? | `Abstain` | `Memory` | `shell` | 0.14 | 0.86 |
+| `holdout_tools_recall_tools_040` | is port 1234 open right now? | `Abstain` | `Memory` | `shell` | 0.11 | 0.89 |
+| `holdout_tools_recall_tools_041` | what environment variable is RUST_LOG set to right now? | `Abstain` | `Memory` | `shell` | 0.14 | 0.86 |
+| `holdout_tools_recall_tools_043` | what gpu am i using for the reranker right now? | `Abstain` | `Memory` | `shell` | 0.18 | 0.82 |
+| `holdout_tools_recall_tools_044` | what is the current system load average? | `Abstain` | `Memory` | `shell` | 0.11 | 0.89 |
+| `holdout_tools_recall_tools_045` | in klbr-core/src/config.rs, what is router_model_path set to by default? | `Abstain` | `Memory` | `read_file` | 0.24 | 0.76 |
+| `holdout_tools_recall_abstain_053` | fix it | `Abstain` | `Memory` | `` | 0.11 | 0.89 |
+| `holdout_tools_recall_abstain_054` | 今週のHacker Newsのトップ記事って何? | `Abstain` | `Memory` | `` | 0.12 | 0.88 |
+| `holdout_tools_recall_tools_056` | what's the uptime since klbr daemon last started? | `Abstain` | `Memory` | `shell` | 0.13 | 0.87 |
+| `holdout_tools_recall_tools_058` | what is the default value of compaction_keep in Config::default()? | `Abstain` | `Memory` | `read_file` | 0.19 | 0.81 |
+| `holdout_tools_recall_tools_059` | in klbr-core/src/config.rs, what is memory_top_k set to by default? | `Abstain` | `Memory` | `read_file` | 0.18 | 0.82 |
+| `holdout_tools_recall_tools_060` | in klbr-core/src/llm.rs, which method produces embeddings? | `Abstain` | `Memory` | `read_file` | 0.24 | 0.76 |
+| `holdout_abstain_toolish_abstain_005` | who won the nba game last night? check | `Abstain` | `Memory` | `` | 0.10 | 0.90 |
+| `holdout_abstain_toolish_abstain_007` | search the internet for 'bge-m3 gguf' and summarize the first result | `Abstain` | `Memory` | `` | 0.15 | 0.85 |
+| `holdout_abstain_toolish_abstain_008` | 今週のHacker Newsのトップ記事って何? 調べて | `Abstain` | `Memory` | `` | 0.11 | 0.89 |
+| `holdout_abstain_toolish_abstain_017` | ssh into the remote server and cat /etc/shadow | `Abstain` | `Memory` | `` | 0.12 | 0.88 |
+| `holdout_abstain_toolish_abstain_021` | will klbr be popular in 2028? | `Abstain` | `Memory` | `` | 0.18 | 0.82 |
+| `holdout_abstain_toolish_abstain_022` | predict what embedding model will be best in 2030 | `Abstain` | `Memory` | `` | 0.24 | 0.76 |
+| `holdout_abstain_toolish_abstain_023` | will we need a reranker in 2 years? | `Abstain` | `Memory` | `` | 0.31 | 0.69 |
+| `holdout_abstain_toolish_abstain_027` | do the thing from yesterday | `Abstain` | `Memory` | `` | 0.10 | 0.90 |
+| `holdout_abstain_toolish_abstain_028` | what's the best programming language? prove it with data | `Abstain` | `Memory` | `` | 0.15 | 0.85 |
+| `holdout_abstain_toolish_abstain_029` | vim or emacs? choose for me | `Abstain` | `Memory` | `` | 0.17 | 0.83 |
+| `holdout_abstain_toolish_abstain_030` | is gemma 4 better than llama 3.1? check online and decide | `Abstain` | `Memory` | `` | 0.21 | 0.79 |
+| `holdout_abstain_toolish_abstain_032` | open a browser and check whether llama-server supports embeddings by default | `Abstain` | `Memory` | `` | 0.19 | 0.81 |
+| `holdout_abstain_toolish_abstain_033` | fetch the gemma-4-26b model card and summarize it | `Abstain` | `Memory` | `` | 0.18 | 0.82 |
+| `holdout_abstain_toolish_abstain_039` | check the latest Rust RFCs and tell me the most controversial one | `Abstain` | `Memory` | `` | 0.12 | 0.88 |
+| `holdout_abstain_toolish_tools_042` | which file defines the ClientMsg type in klbr-ipc? | `Abstain` | `Memory` | `read_file` | 0.12 | 0.88 |
+| `holdout_abstain_toolish_tools_044` | write a file docs/router_holdout_abstain_toolish.txt with the content "ok" | `Abstain` | `Memory` | `write_file` | 0.19 | 0.81 |
+| `holdout_abstain_toolish_tools_045` | klbr-core/src/config.rsのmemory_top_kはいくつ? | `Abstain` | `Memory` | `read_file` | 0.13 | 0.87 |
+| `holdout_abstain_toolish_tools_046` | 今メモリ使用量はどれくらい? | `Abstain` | `Memory` | `shell` | 0.11 | 0.89 |
+| `holdout_abstain_toolish_tools_048` | how big is agent.db right now on disk? | `Abstain` | `Memory` | `shell` | 0.10 | 0.90 |
+| `holdout_abstain_toolish_tools_049` | in klbr-core/src/config.rs, what is llm_url set to? | `Abstain` | `Memory` | `read_file` | 0.21 | 0.79 |
+| `holdout_memory_toolish_tools_043` | which file defines the ClientMsg type in klbr-ipc? | `Abstain` | `Memory` | `read_file` | 0.12 | 0.88 |
+| `holdout_memory_toolish_tools_044` | in klbr-core/src/config.rs, what is memory_top_k set to by default? | `Abstain` | `Memory` | `read_file` | 0.18 | 0.82 |
+| `holdout_memory_toolish_tools_045` | write a file docs/router_holdout_memory_toolish.txt with the content "ok" | `Abstain` | `Memory` | `write_file` | 0.16 | 0.84 |
+| `holdout_memory_toolish_tools_046` | 今メモリ使用量はどれくらい? | `Abstain` | `Memory` | `shell` | 0.11 | 0.89 |
+| `holdout_memory_toolish_tools_047` | klbr-core/src/config.rsのembed_dimはいくつ? | `Abstain` | `Memory` | `read_file` | 0.13 | 0.87 |
+| `holdout_memory_toolish_tools_049` | how big is agent.db right now on disk? | `Abstain` | `Memory` | `shell` | 0.10 | 0.90 |
+| `holdout_memory_toolish_tools_050` | in klbr-daemon/src/daemon.rs, does it forward agent errors to the tui? | `Abstain` | `Memory` | `read_file` | 0.13 | 0.87 |
+| `holdout_memory_toolish_abstain_054` | will klbr be popular in 2028? | `Abstain` | `Memory` | `` | 0.18 | 0.82 |
+| `holdout_memory_toolish_abstain_056` | 今週のHacker Newsのトップ記事って何? | `Abstain` | `Memory` | `` | 0.12 | 0.88 |
+| `holdout_memory_toolish_abstain_060` | who won the nba game last night? check | `Abstain` | `Memory` | `` | 0.10 | 0.90 |
+| `shell_test_003` | is llama-server currently loaded with gemma or something else? | `Abstain` | `Memory` | `shell` | 0.17 | 0.83 |
+| `shell_test_005` | how many open file descriptors does klbr-daemon have right now? | `Abstain` | `Memory` | `shell` | 0.11 | 0.89 |
+| `shell_test_007` | what environment variable is RUST_LOG set to? | `Abstain` | `Memory` | `shell` | 0.17 | 0.83 |
+| `shell_test_009` | rerankerのプロセスは現在いくつ走ってますか | `Abstain` | `Memory` | `shell` | 0.15 | 0.85 |
+| `shell_test_010` | what's the uptime since klbr daemon last started | `Abstain` | `Memory` | `shell` | 0.12 | 0.88 |
+| `shell_test_011` | how big is agent.db right now on disk? | `Abstain` | `Memory` | `shell` | 0.10 | 0.90 |
+| `shell_test_012` | any nix builds running in background rn? | `Abstain` | `Memory` | `shell` | 0.12 | 0.88 |
+| `shell_test_013` | what's the current hostname? | `Abstain` | `Memory` | `shell` | 0.10 | 0.90 |
+| `shell_test_015` | what's the network interface with the highest tx bytes? | `Abstain` | `Memory` | `shell` | 0.19 | 0.81 |
+| `shell_test_016` | can you check if the reranker http endpoint is responding? | `Abstain` | `Memory` | `shell` | 0.18 | 0.82 |
+| `shell_test_019` | is the klbr tui currently connected to the daemon socket? | `Abstain` | `Memory` | `shell` | 0.11 | 0.89 |
+| `shell_test_020` | whats the total vram on ths gpu | `Abstain` | `Memory` | `shell` | 0.10 | 0.90 |
+| `shell_test_022` | what ip address is localhost:8002 actually bound to? | `Abstain` | `Memory` | `shell` | 0.16 | 0.84 |
+| `shell_test_027` | is port 1234 open right now? | `Abstain` | `Memory` | `shell` | 0.11 | 0.89 |
+| `shell_test_028` | what's the size of the current target/ directory? | `Abstain` | `Memory` | `shell` | 0.13 | 0.87 |
+| `shell_test_030` | journalctl -u klbr — what does it say? | `Abstain` | `Memory` | `shell` | 0.16 | 0.84 |
+| `shell_test_031` | is the disk partition containing agent.db almost full? | `Abstain` | `Memory` | `shell` | 0.11 | 0.89 |
+| `rf_test_001` | what compaction_keep default is in config.rs right now? | `Abstain` | `Memory` | `read_file` | 0.15 | 0.85 |
+| `rf_test_002` | which file defines the ClientMsg type? | `Abstain` | `Memory` | `read_file` | 0.13 | 0.87 |
+| `rf_test_003` | what does the tools mod.rs export? | `Abstain` | `Memory` | `read_file` | 0.15 | 0.85 |
+| `rf_test_004` | what's the rerank_top_k default in the calibrated config file? | `Abstain` | `Memory` | `read_file` | 0.28 | 0.72 |
+| `rf_test_005` | how many queries are in internal_eval_starter.json? | `Abstain` | `Memory` | `read_file` | 0.17 | 0.83 |
+| `rf_test_006` | config.rsのmemory_support_score_gapのデフォルト値はいくつ? | `Abstain` | `Memory` | `read_file` | 0.15 | 0.85 |
+| `rf_test_007` | is there a router.rs module in klbr-core? | `Abstain` | `Memory` | `read_file` | 0.14 | 0.86 |
+| `rf_test_008` | what's the current benchmark README say about the retrieval command? | `Abstain` | `Memory` | `read_file` | 0.15 | 0.85 |
+| `rf_test_010` | what's in the .gitignore for klbr? | `Abstain` | `Memory` | `read_file` | 0.17 | 0.83 |
+| `rf_test_011` | what feature flags does klbr-bench's Cargo.toml declare? | `Abstain` | `Memory` | `read_file` | 0.14 | 0.86 |
+| `rf_test_012` | how many memories does internal_eval_passive_recall_v3.json have? | `Abstain` | `Memory` | `read_file` | 0.17 | 0.83 |
+| `rf_test_013` | what is the embed_dim value in the support calibrated config? | `Abstain` | `Memory` | `read_file` | 0.17 | 0.83 |
+| `rf_test_014` | does flake.nix pin any specific nixpkgs revision? | `Abstain` | `Memory` | `read_file` | 0.13 | 0.87 |
+| `rf_test_015` | what's the passive_support_score_gap set to in the test config? | `Abstain` | `Memory` | `read_file` | 0.21 | 0.79 |
+| `rf_test_017` | what's written in docs/klbr_memory_design.md about L1? | `Abstain` | `Memory` | `read_file` | 0.21 | 0.79 |
+| `rf_test_019` | what does the recall.rs tool do with the tag_and param? | `Abstain` | `Memory` | `read_file` | 0.14 | 0.86 |
+| `rf_test_021` | llm.rsのstreamメソッドのシグネチャは? | `Abstain` | `Memory` | `read_file` | 0.15 | 0.85 |
+| `rf_test_022` | what's the current threshold in the router_model.json? | `Abstain` | `Memory` | `read_file` | 0.18 | 0.82 |
+| `rf_test_024` | how many memories does the passive recall v3 dataset have? | `Abstain` | `Memory` | `read_file` | 0.17 | 0.83 |
+| `rf_test_026` | what warnings did the last cargo check produce? | `Abstain` | `Memory` | `read_file,shell` | 0.15 | 0.85 |
+| `rf_test_027` | what's the embed_dim in the router_model.json we trained? | `Abstain` | `Memory` | `read_file` | 0.29 | 0.71 |
+| `rf_test_029` | how is the SweepRange struct defined in klbr-bench main.rs? | `Abstain` | `Memory` | `read_file` | 0.22 | 0.78 |
+| `wf_test_001` | set the compaction_keep to 15 in config.rs | `Abstain` | `Memory` | `write_file` | 0.12 | 0.88 |
+| `wf_test_002` | write the router precision result to a markdown file | `Abstain` | `Memory` | `write_file` | 0.18 | 0.82 |
+| `wf_test_003` | create a new slice file with 5 tool-needed examples | `Abstain` | `Memory` | `write_file` | 0.14 | 0.86 |
+| `wf_test_004` | change the embed_url to localhost:8005 in the config | `Abstain` | `Memory` | `write_file` | 0.15 | 0.85 |
+| `wf_test_005` | write a tiny benchmarking summary to out-router-iter-2/notes.txt | `Abstain` | `Memory` | `write_file` | 0.16 | 0.84 |
+| `wf_test_006` | add router_model_path to the default config and set it to a path | `Abstain` | `Memory` | `write_file` | 0.18 | 0.82 |
+| `wf_test_008` | patch the memory_top_k to 5 in config.rs | `Abstain` | `Memory` | `write_file` | 0.16 | 0.84 |
+| `wf_test_009` | update the dataset_id in internal_eval_starter.json to v3 | `Abstain` | `Memory` | `write_file` | 0.13 | 0.87 |
+| `wf_test_010` | write the sweep results summary to a txt file in the output dir | `Abstain` | `Memory` | `write_file` | 0.12 | 0.88 |
+| `wf_test_011` | save a note about the router precision gate to docs/ | `Abstain` | `Memory` | `write_file` | 0.12 | 0.88 |
+| `wf_test_012` | create a quick script to run all benchmarks in sequence | `Abstain` | `Memory` | `write_file,shell` | 0.12 | 0.88 |
+| `wf_test_013` | add a line to .gitignore to exclude agent.db | `Abstain` | `Memory` | `write_file` | 0.16 | 0.84 |
+| `wf_test_014` | 新しいスライスJSONを作ってmemory_laneのクエリを10個追加して | `Abstain` | `Memory` | `write_file` | 0.14 | 0.86 |
+| `wf_test_015` | bump the rerank_top_k to 15 in the calibrated config and save it | `Abstain` | `Memory` | `write_file` | 0.16 | 0.84 |
+| `abs_test_003` | what will my project look like in 6 months? | `Abstain` | `Memory` | `` | 0.15 | 0.85 |
+| `abs_test_004` | what's the best programming language? | `Abstain` | `Memory` | `` | 0.16 | 0.84 |
+| `abs_test_006` | what's the github token for the ci runner? | `Abstain` | `Memory` | `` | 0.11 | 0.89 |
+| `abs_test_007` | vim or emacs? | `Abstain` | `Memory` | `` | 0.20 | 0.80 |
+| `abs_test_009` | どのモデルが今一番スコアが高いですか — ベンチマークサイトで | `Abstain` | `Memory` | `` | 0.15 | 0.85 |
+| `abs_test_010` | when will the klbr project be finished? | `Abstain` | `Memory` | `` | 0.17 | 0.83 |
+| `abs_test_011` | is nomic-embed or bge-m3 better for this use case? | `Abstain` | `Memory` | `` | 0.27 | 0.73 |
+| `abs_test_012` | y | `Abstain` | `Memory` | `` | 0.13 | 0.87 |
+| `abs_test_013` | is the llama.cpp repo trending on github today? | `Abstain` | `Memory` | `` | 0.14 | 0.86 |
+| `abs_test_014` | how should klbr evolve to support multi-user? | `Abstain` | `Memory` | `` | 0.23 | 0.77 |
+
+## Holdout Metrics (query_id starts with `holdout_`)
+
+- labeled queries evaluated: `180`
+- memory false-abstain rate (memory→abstain): `0.0000`
+- abstain false-memory rate (abstain→memory): `0.5360`
+- tool-required queries: `70`
+- tool-required false-memory rate (tool→memory): `0.6857`
+- tool-required recall (tool→abstain): `0.3143`
+- accuracy: `0.6278`
+
+## Holdout Metrics By Bucket
+
+| bucket | total | tools | tool→memory | tool recall | memory→abstain | abstain→memory | accuracy |
+|---|---:|---:|---:|---:|---:|---:|---:|
+| `abstain_toolish` | 60 | 10 | 0.6000 | 0.4000 | 0.0000 | 0.4000 | 0.6667 |
+| `memory_toolish` | 60 | 10 | 0.7000 | 0.3000 | 0.0000 | 0.5000 | 0.8333 |
+| `tools_recall` | 60 | 50 | 0.7000 | 0.3000 | 0.0000 | 0.6727 | 0.3833 |
diff --git a/benchmarks/run_all.py b/benchmarks/run_all.py
index bf4b8ac..cdcb56e 100755
--- a/benchmarks/run_all.py
+++ b/benchmarks/run_all.py
@@ -120,7 +120,7 @@ def main():
             / "models"
             / "router"
             / "linear"
-            / "out-router-linear-iter-9"
+            / "out-router-linear-iter-15"
             / "router_model.json"
         ),
     )
diff --git a/docs/benchmark-issues-report.md b/docs/benchmark-issues-report.md
index 15b3156..b865196 100644
--- a/docs/benchmark-issues-report.md
+++ b/docs/benchmark-issues-report.md
@@ -45,6 +45,7 @@ the current numbers show a clear mismatch between coarse retrieval success and a
 - **symptom**: tools lane test split reports `0.00%` tools precision/recall despite router accuracy of `87.50%`.
 - **cause**: the centroid/linear classifier in [router.rs](file:///home/mayer/proj/klbr/klbr-core/src/router.rs) misclassified gold tools queries (`test_q9`, `test_q15`) as `Memory` queries because of slight feature shifts on the test set. while the end-to-end agent still fell back to triggering tools correctly via prompt instructions, the router's explicit classification failed.
 - **status/fix**: the bench router calibration now tracks `required_tools` as a dedicated tool-required slice. centroid and linear threshold sweeps prefer candidates that preserve memory recall while keeping tool-required false-memory (`tool→memory`) under the calibration cap, and router reports expose tool-required count/recall plus the required tool names on misclassified rows.
+- **latest artifact**: [out-router-linear-iter-15](file:///home/mayer/proj/klbr/benchmarks/models/router/linear/out-router-linear-iter-15/router_report.md) was regenerated with these metrics. current test-split `tool→memory` is `0.6913` and tool-required recall is `0.3087`, so the artifact now surfaces the residual model/data problem instead of hiding it.
 
 ### E. upstream evaluator blocking (`klbr-1yn`)
 - **symptom**: real upstream LongMemEval QA accuracy is blocked.
@@ -72,5 +73,5 @@ the current numbers show a clear mismatch between coarse retrieval success and a
 
 1. **unify the graphs**: merge `memory_edges` and `edges` into one transactional links graph in [memory.rs](file:///home/mayer/proj/klbr/klbr-core/src/memory.rs).
 2. **strict entity mapping**: add entity extraction to the [evidence.rs](file:///home/mayer/proj/klbr/klbr-core/src/evidence.rs) planning phase to block cross-entity hallucinations (e.g. film vs camera).
-3. **regenerate router artifacts**: rerun the router bench after calibration changes and publish fresh model/report outputs so `tool→memory` and tool-required recall are tracked on future sweeps.
+3. **improve router separability**: use the fresh tool-required metrics to improve router training data/features if runtime memory-vs-tool routing remains too permissive.
 4. **unblock `klbr-1yn`**: import the official LongMemEval Python evaluation script to establish baseline QA accuracy parity.
diff --git a/klbr-core/src/models.rs b/klbr-core/src/models.rs
index e9354f3..af9fb74 100644
--- a/klbr-core/src/models.rs
+++ b/klbr-core/src/models.rs
@@ -1,8 +1,8 @@
 use anyhow::Result;
 use futures::StreamExt;
 use reqwest::{Client, RequestBuilder};
-use serde::{de::Error as _, Deserialize, Deserializer, Serialize};
-use serde_json::{json, Map, Value};
+use serde::{Deserialize, Deserializer, Serialize, de::Error as _};
+use serde_json::{Map, Value, json};
 use std::collections::{BTreeMap, HashMap};
 use tokio::sync::mpsc;
 
@@ -386,7 +386,13 @@ impl LlmClient {
             self.llm_client.clone()
         } else if config.url == self.config.reranker.url {
             self.reranker_client.clone()
-        } else if let Some((_, c)) = self.config.embedders.iter().zip(&self.embedder_clients).find(|(ec, _)| ec.url == config.url) {
+        } else if let Some((_, c)) = self
+            .config
+            .embedders
+            .iter()
+            .zip(&self.embedder_clients)
+            .find(|(ec, _)| ec.url == config.url)
+        {
             c.clone()
         } else {
             build_client(config).unwrap_or_else(|_| Client::new())
@@ -803,10 +809,7 @@ impl LlmClient {
             }
         };
         tracing::info!("non-stream response status: {}", v.status());
-        let text = Self::error_for_status_with_body(v)
-            .await?
-            .text()
-            .await?;
+        let text = Self::error_for_status_with_body(v).await?.text().await?;
 
         if text.trim().starts_with("data:") {
             let mut content = String::new();
@@ -849,7 +852,15 @@ impl LlmClient {
                     }
                 }
             }
-            Ok((content, Usage { prompt_tokens, completion_tokens, total_tokens, ..Default::default() }))
+            Ok((
+                content,
+                Usage {
+                    prompt_tokens,
+                    completion_tokens,
+                    total_tokens,
+                    ..Default::default()
+                },
+            ))
         } else {
             let v: Value = serde_json::from_str(&text)?;
             let content = v["choices"][0]["message"]["content"]
@@ -949,7 +960,12 @@ impl LlmClient {
         result
     }
 
-    async fn embed_with_config(&self, client: &Client, config: &ModelConfig, cleaned_text: &str) -> Result<Vec<f32>> {
+    async fn embed_with_config(
+        &self,
+        client: &Client,
+        config: &ModelConfig,
+        cleaned_text: &str,
+    ) -> Result<Vec<f32>> {
         let embed_model = self.resolve_model(config, true).await?;
 
         let mut body = json!({ "model": embed_model, "input": cleaned_text });
@@ -983,12 +999,7 @@ impl LlmClient {
             .await?
             .json::<Value>()
             .await?;
-        let embedding: Vec<f32> = v["data"][0]["embedding"]
-            .as_array()
-            .ok_or_else(|| anyhow::anyhow!("no embedding in response"))?
-            .iter()
-            .map(|x| Ok(x.as_f64().unwrap_or(0.0) as f32))
-            .collect::<Result<Vec<f32>>>()?;
+        let embedding = Self::parse_single_embedding_response(&v)?;
 
         if embedding.len() != self.config.embed_dim {
             return Err(anyhow::anyhow!(
@@ -1002,6 +1013,72 @@ impl LlmClient {
         Ok(embedding)
     }
 
+    fn parse_embedding_item(item: &Value) -> Result<Vec<f32>> {
+        let raw_embedding = item
+            .get("embedding")
+            .ok_or_else(|| anyhow::anyhow!("no embedding in response"))?;
+        let values = raw_embedding
+            .as_array()
+            .ok_or_else(|| anyhow::anyhow!("embedding was not an array"))?;
+        let values = if values.len() == 1 && values[0].is_array() {
+            values[0]
+                .as_array()
+                .ok_or_else(|| anyhow::anyhow!("embedding was not an array"))?
+        } else {
+            values
+        };
+
+        values
+            .iter()
+            .map(|x| {
+                x.as_f64()
+                    .map(|value| value as f32)
+                    .ok_or_else(|| anyhow::anyhow!("embedding value was not a number"))
+            })
+            .collect()
+    }
+
+    fn embedding_items(v: &Value) -> Result<&Vec<Value>> {
+        if let Some(data) = v.get("data").and_then(Value::as_array) {
+            return Ok(data);
+        }
+        v.as_array()
+            .ok_or_else(|| anyhow::anyhow!("no embedding data in response"))
+    }
+
+    fn parse_embedding_batch_response(v: &Value, expected_count: usize) -> Result<Vec<Vec<f32>>> {
+        let items = Self::embedding_items(v)?;
+        if items.len() != expected_count {
+            return Err(anyhow::anyhow!(
+                "batch embedding response count mismatch: got {}, expected {}",
+                items.len(),
+                expected_count
+            ));
+        }
+
+        let mut parsed = items
+            .iter()
+            .enumerate()
+            .map(|(pos, item)| {
+                let index = item
+                    .get("index")
+                    .and_then(Value::as_u64)
+                    .map(|index| index as usize)
+                    .unwrap_or(pos);
+                Ok((index, Self::parse_embedding_item(item)?))
+            })
+            .collect::<Result<Vec<_>>>()?;
+        parsed.sort_by_key(|(index, _)| *index);
+        Ok(parsed.into_iter().map(|(_, embedding)| embedding).collect())
+    }
+
+    fn parse_single_embedding_response(v: &Value) -> Result<Vec<f32>> {
+        Self::parse_embedding_batch_response(v, 1)?
+            .into_iter()
+            .next()
+            .ok_or_else(|| anyhow::anyhow!("no embedding in response"))
+    }
+
     pub async fn embed(&self, text: &str) -> Result<Vec<f32>> {
         let cleaned_text = Self::strip_media_urls(text);
         let num_embedders = self.config.embedders.len();
@@ -1009,7 +1086,10 @@ impl LlmClient {
             return Err(anyhow::anyhow!("no embedders configured"));
         }
 
-        let start_idx = self.embedder_index.fetch_add(1, std::sync::atomic::Ordering::Relaxed) % num_embedders;
+        let start_idx = self
+            .embedder_index
+            .fetch_add(1, std::sync::atomic::Ordering::Relaxed)
+            % num_embedders;
         let mut last_err = None;
 
         for i in 0..num_embedders {
@@ -1034,18 +1114,12 @@ impl LlmClient {
         Err(last_err.unwrap_or_else(|| anyhow::anyhow!("no embedders configured or available")))
     }
 
-    async fn embed_chunk_with_config(&self, client: &Client, config: &ModelConfig, chunk_texts: &[String]) -> Result<Vec<Vec<f32>>> {
-        #[derive(Debug, Deserialize)]
-        struct EmbeddingResponse {
-            data: Vec<EmbeddingItem>,
-        }
-
-        #[derive(Debug, Deserialize)]
-        struct EmbeddingItem {
-            index: usize,
-            embedding: Vec<f32>,
-        }
-
+    async fn embed_chunk_with_config(
+        &self,
+        client: &Client,
+        config: &ModelConfig,
+        chunk_texts: &[String],
+    ) -> Result<Vec<Vec<f32>>> {
         let embed_model = self.resolve_model(config, true).await?;
         let endpoint = Self::endpoint(&config.url, "/embeddings");
 
@@ -1068,28 +1142,20 @@ impl LlmClient {
             .await?;
 
         let response = Self::error_for_status_with_body(response).await?;
-        let mut parsed: EmbeddingResponse = response.json().await?;
-        parsed.data.sort_by_key(|x| x.index);
-
-        if parsed.data.len() != chunk_texts.len() {
-            return Err(anyhow::anyhow!(
-                "batch embedding response count mismatch: got {}, expected {}",
-                parsed.data.len(),
-                chunk_texts.len()
-            ));
-        }
+        let parsed = response.json::<Value>().await?;
+        let embeddings = Self::parse_embedding_batch_response(&parsed, chunk_texts.len())?;
 
         let mut results = Vec::with_capacity(chunk_texts.len());
-        for item in parsed.data {
-            if item.embedding.len() != self.config.embed_dim {
+        for embedding in embeddings {
+            if embedding.len() != self.config.embed_dim {
                 return Err(anyhow::anyhow!(
                     "embedding dimension mismatch: got {}, expected {} (model: {})",
-                    item.embedding.len(),
+                    embedding.len(),
                     self.config.embed_dim,
                     embed_model
                 ));
             }
-            results.push(item.embedding);
+            results.push(embedding);
         }
 
         Ok(results)
@@ -1105,7 +1171,10 @@ impl LlmClient {
             return Err(anyhow::anyhow!("no embedders configured"));
         }
 
-        let start_idx = self.embedder_index.fetch_add(1, std::sync::atomic::Ordering::Relaxed) % num_embedders;
+        let start_idx = self
+            .embedder_index
+            .fetch_add(1, std::sync::atomic::Ordering::Relaxed)
+            % num_embedders;
         let mut last_err = None;
 
         for i in 0..num_embedders {
@@ -1113,7 +1182,10 @@ impl LlmClient {
             let config = &self.config.embedders[idx];
             let client = &self.embedder_clients[idx];
 
-            match self.embed_chunk_with_config(client, config, &chunk_texts).await {
+            match self
+                .embed_chunk_with_config(client, config, &chunk_texts)
+                .await
+            {
                 Ok(embeddings) => {
                     let results = chunk_indices.iter().copied().zip(embeddings).collect();
                     return Ok(results);
@@ -1152,10 +1224,13 @@ impl LlmClient {
 
         for (idx, text) in cleaned_texts.into_iter().enumerate() {
             let estimated_tokens = (text.len() / 4).max(1);
-            if current_chunk_texts.len() >= MAX_TEXTS_PER_CHUNK 
-                || current_chunk_tokens + estimated_tokens > MAX_TOKENS_PER_CHUNK 
+            if current_chunk_texts.len() >= MAX_TEXTS_PER_CHUNK
+                || current_chunk_tokens + estimated_tokens > MAX_TOKENS_PER_CHUNK
             {
-                chunks.push((std::mem::take(&mut current_chunk_texts), std::mem::take(&mut current_chunk_indices)));
+                chunks.push((
+                    std::mem::take(&mut current_chunk_texts),
+                    std::mem::take(&mut current_chunk_indices),
+                ));
                 current_chunk_tokens = 0;
             }
             current_chunk_tokens += estimated_tokens;
@@ -1183,8 +1258,6 @@ impl LlmClient {
         Ok(final_results)
     }
 
-
-
     pub async fn rerank(
         &self,
         query: &str,
@@ -1597,6 +1670,31 @@ mod tests {
         assert_eq!(embedding, vec![0.1, 0.2, 0.3]);
     }
 
+    #[tokio::test]
+    async fn embedder_accepts_llama_legacy_embedding_shape() {
+        let base_url = spawn_http_response(
+            "200 OK",
+            "application/json",
+            r#"[{"index": 0, "embedding": [[0.1, 0.2, 0.3]]}]"#,
+        )
+        .await;
+        let config = ModelsConfig {
+            llm: ModelConfig::default(),
+            embedder: ModelConfig {
+                url: base_url,
+                model: "legacy-embed".into(),
+                ..Default::default()
+            },
+            reranker: ModelConfig::default(),
+            embedders: Vec::new(),
+            embed_dim: 3,
+        };
+
+        let client = LlmClient::new(config);
+        let embedding = client.embed("hello").await.unwrap();
+        assert_eq!(embedding, vec![0.1, 0.2, 0.3]);
+    }
+
     #[tokio::test]
     async fn embedder_batch_failover_works() {
         // Spawn a failing server
@@ -1635,7 +1733,40 @@ mod tests {
         };
 
         let client = LlmClient::new(config);
-        let embeddings = client.embed_batch(&["hello".into(), "world".into()]).await.unwrap();
+        let embeddings = client
+            .embed_batch(&["hello".into(), "world".into()])
+            .await
+            .unwrap();
+        assert_eq!(embeddings.len(), 2);
+        assert_eq!(embeddings[0], vec![0.1, 0.2, 0.3]);
+        assert_eq!(embeddings[1], vec![0.4, 0.5, 0.6]);
+    }
+
+    #[tokio::test]
+    async fn embedder_batch_accepts_llama_legacy_embedding_shape() {
+        let base_url = spawn_http_response(
+            "200 OK",
+            "application/json",
+            r#"[{"index": 1, "embedding": [[0.4, 0.5, 0.6]]}, {"index": 0, "embedding": [[0.1, 0.2, 0.3]]}]"#,
+        )
+        .await;
+        let config = ModelsConfig {
+            llm: ModelConfig::default(),
+            embedder: ModelConfig {
+                url: base_url,
+                model: "legacy-embed".into(),
+                ..Default::default()
+            },
+            reranker: ModelConfig::default(),
+            embedders: Vec::new(),
+            embed_dim: 3,
+        };
+
+        let client = LlmClient::new(config);
+        let embeddings = client
+            .embed_batch(&["hello".into(), "world".into()])
+            .await
+            .unwrap();
         assert_eq!(embeddings.len(), 2);
         assert_eq!(embeddings[0], vec![0.1, 0.2, 0.3]);
         assert_eq!(embeddings[1], vec![0.4, 0.5, 0.6]);