diff --git a/llama/baseline-qwen3.6-2026-09-04.md b/llama/baseline-qwen3.6-2026-09-04.md
new file mode 100644
index 0000000..3d9f8b6
--- /dev/null
+++ b/llama/baseline-qwen3.6-2026-09-04.md
@@ -0,0 +1,50 @@
+# Baseline — qwen3.6-35b-a3b (2026-09-04)
+
+## Build / host
+- llama-server `b210-50f068fff`, Strix Halo 395 (32 threads), 125 GiB unified
+- Router: `MODELS_MAX=1`, port 9931, sleep-idle 900s. Model was `sleeping`, woke on first request.
+
+## Config snapshot (models.ini [qwen3.6-35b-a3b])
+- weights `Qwen3.6-35B-A3B-UD-Q8_K_XL.gguf` (37G) + `mmproj-BF16` (861M)
+- `spec-type draft-mtp`, `spec-draft-n-max 2`
+- template `qwen-fixed-chat_template.jinja` (froggeric-v22.1), `reasoning auto`, `deepseek`, `preserve`
+- sampling: temp 1.0, top-p 0.95, top-k 20, min-p 0.0, presence 1.5, repeat 1.0 (= upstream Thinking/general)
+- unset (defaults): ctx-size → 262144 from GGUF, batch 2048 / ubatch 512, cache-type K/V default, np auto → **4 slots × 262144**, unified KV on, cont-batching on
+
+## Final lock-in (2026-09-04)
+
+`llama/models.ini` `[*]` + `[qwen3.6-35b-a3b]`:
+- `n-gpu-layers = all` (was 99; future-proof for deeper models, global `[*]`)
+- `spec-draft-n-max = 3` (was 2; MTP sweep winner)
+- `ubatch-size = 2048` (was default 512; −21% long-prompt ingest, peaked vs 4096 regression)
+- unchanged: ctx 262144, 4 slots unified KV (F16), flash-attn, sampling Thinking/general, froggeric jinja, KV quant skipped
+
+Verified: child args `--n-gpu-layers all --spec-draft-n-max 3 --ubatch-size 2048`; sanity 71.1 tok/s / 11-11 accepted.
+
+## Tests (all via POST /v1/chat/completions, model=qwen3.6-35b-a3b, stream=false)
+
+| # | Task | prompt_n | prompt tok/s | gen tok/s | wall | draft accepted/total | Notes |
+|---|------|----------|--------------|-----------|------|----------------------|-------|
+| T1 | "one sentence: speculative decoding", max 128 | 19 | 151 | 60.9 | 2.2s | — | thinking model; short replies mostly land in `reasoning_content` |
+| T2 | math + `<\|think_medium\|>`, max 512 | 40 | 266 | 68.0 | 7.7s | — | ok |
+| T3 | tool-use (get_weather/Paris), max 256 | 382 | 646 | 59.8 | 2.0s | — | correct `tool_calls` emitted |
+| T4 | codebase dump ~38KB chars / 44 files, max 256 | 12431 | 835 | 64.7 | 16.3s | 57/64 (0.89 this req) | correct answer (top-level dirs + llama/ contents), `reasoning_content` empty (auto skipped thinking) |
+| T5 | "Say hi in one sentence", max 64 | 16 | 147 | 62.1 | ~1.2s | 38/48 (0.79 this req) | **truncated inside `<think>`** — content empty, reasoning only. Expected: thinking models need generous `max_tokens`; 64 cuts the answer off. |
+
+## Server counters (after T5, /metrics?model=qwen3.6-35b-a3b)
+- prompt avg 471 tok/s, gen avg 65.1 tok/s
+- spec totals: 2727 drafted / 1968 accepted = **0.72**; pos0 1088, pos1 880 (cond. pos1 ≈ 0.81)
+- `n_tokens_max` observed 36163 (historical, pre-baseline)
+- `requests_deferred` 0
+- Prior journal (overnight task 477): prompt 276 tok/s, gen 46 tok/s, acceptance 0.55, mean len 2.10 — slower/longer-context than today's short tests
+
+## Memory
+- `MemoryCurrent` 42.5G, `MemoryPeak` 63.7G; system 49G used / 75G avail
+- Headroom for full-262k KV (~21.5G at F16) looks fine on paper; not yet tested at the limit
+
+## Observations / follow-ups
+1. Sampling default matches upstream Thinking/general — keep; precise-coding + instruct presets belong as **per-request overrides** from pi, not in models.ini.
+2. `max_tokens` guidance for pi: thinking responses need room (T5 shows 64 tokens = answer never leaves `<think>`). Suggest ≥1024 default for agentic calls.
+3. Tool-use path works with custom jinja (T3). Deeper agentic loop (multi-turn + truncation warnings) not yet tested.
+4. Long-context validated to 12k prompt tokens only. Full 262k soak + compaction-concurrency test still open (deferred, lower priority per user).
+5. Next tuning candidates: ubatch/batch for big prompts, `spec-draft-n-max` 2→3→4 sweep, cache-type-k/v if 262k-soak pressures memory.
diff --git a/llama/models.ini b/llama/models.ini
index adadf13..43668d1 100644
--- a/llama/models.ini
+++ b/llama/models.ini
@@ -8,7 +8,7 @@ version = 1
 ; Global defaults shared by every model instance (overridable per-model)
 ; ---------------------------------------------------------------------------
 [*]
-n-gpu-layers  = 99
+n-gpu-layers  = all
 flash-attn    = on
 jinja         = true
 metrics       = true ; Prometheus /metrics endpoint (per-model: /metrics?model=...)
@@ -21,7 +21,8 @@ load-on-startup    = true
 model              = /home/spencer/.cache/models/Qwen3.6-35B-A3B-UD-Q8_K_XL.gguf
 mmproj             = /home/spencer/.cache/models/Qwen3.6-35B-A3B-mmproj-BF16.gguf
 spec-type          = draft-mtp
-spec-draft-n-max   = 2
+spec-draft-n-max   = 3
+ubatch-size        = 2048
 chat-template-file = /home/spencer/.config/llama.cpp/qwen-fixed-chat_template.jinja
 reasoning-format   = deepseek
 reasoning          = auto
diff --git a/llama/prometheus/README.md b/llama/prometheus/README.md
new file mode 100644
index 0000000..06bbb0b
--- /dev/null
+++ b/llama/prometheus/README.md
@@ -0,0 +1,21 @@
+# Short-term Prometheus for llama-server (rootless podman)
+
+- Container: `prometheus-short`, host networking, web UI on 127.0.0.1:9090
+- Config: `prometheus.yml` (this dir), bind-mounted read-only +SELinux :z
+- TSDB data: `~/.local/share/prometheus-data` (7d retention)
+- Scrapes router `127.0.0.1:9931/metrics?model=qwen3.6-35b-a3b` every 10s
+- Router requires `?model=` per job (400 without it)
+
+Manage:
+    podman start prometheus-short | podman stop prometheus-short | podman logs -f prometheus-short
+    # config edits: podman exec prometheus-short kill -HUP 1   (or restart)
+
+Useful queries (qwen3.6 job):
+    increase(llamacpp:prompt_tokens_seconds[...])   # prompt tok/s windowed
+    llmacpp:predicted_tokens_seconds                 # gen tok/s gauge
+    sum(increase(llamacpp:spec_decode_num_accepted_tokens_total[5m]))
+      / sum(increase(llamacpp:spec_decode_num_draft_tokens_total[5m]))   # acceptance
+    sum(llamacpp:spec_decode_num_accepted_tokens_per_pos_total{position="1"})
+      / sum(llamacpp:spec_decode_num_accepted_tokens_per_pos_total{position="0"})  # pos1 conditional
+
+NOTE: rootless container does NOT auto-restart on machine reboot; restart manually with `podman start prometheus-short` (or promote to a quadlet user service if it outlives the short-term window).
diff --git a/llama/prometheus/prometheus.yml b/llama/prometheus/prometheus.yml
new file mode 100644
index 0000000..ca8e662
--- /dev/null
+++ b/llama/prometheus/prometheus.yml
@@ -0,0 +1,25 @@
+# Short-term Prometheus scrape config for llama-server router (user service, port 9931).
+# One scrape job per model instance; the router serves per-model metrics via ?model=.
+# Stored in the dotfiles repo for reproducibility alongside tuning notes.
+global:
+  scrape_interval: 10s
+  evaluation_interval: 30s
+
+rule_files: []
+
+scrape_configs:
+  # qwen3.6-35b-a3b (current tuning target)
+  - job_name: llama-qwen3.6-35b-a3b
+    metrics_path: /metrics
+    params:
+      model: ["qwen3.6-35b-a3b"]
+    static_configs:
+      - targets: ["127.0.0.1:9931"]
+        labels:
+          model: "qwen3.6-35b-a3b"
+
+  # NOTE: the router returns 400 for /metrics without a model param,
+  # so every job must pin ?model=. User-side aggregate is not exposed.
+
+  # Add other models later by copying the qwen3.6 job and changing params.model
+  # e.g. qwen3.8-27b, qwen3.8-flash-next, qwen3-coder-next-80b-a3b
\ No newline at end of file
diff --git a/llama/results-envelope-100k-2026-09-04.md b/llama/results-envelope-100k-2026-09-04.md
new file mode 100644
index 0000000..e9109a1
--- /dev/null
+++ b/llama/results-envelope-100k-2026-09-04.md
@@ -0,0 +1,34 @@
+# Envelope soak — 92k-token prompt on qwen3.6-35b-a3b (2026-09-04)
+
+Single 92,095-token prompt (llama.cpp `src/**` source dump, 47 files, 361 KB) sent to the live router instance (n-max 3, 4 slots, unified KV, ubatch 512, KV F16 default).
+
+## Results
+- Prompt processing: 92,095 tokens in 200.2 s → **460 tok/s** (vs 813 tok/s at 14.4k prompt — long prompts roughly halve prompt throughput)
+- TTFT ≈ 200 s (prompt-bound; generation negligible by comparison)
+- Generation: 192 tokens @ **45.8 tok/s** — long-context decode is slower than short-ctx ~60-72 (KV attention cost + unified pool)
+- Draft acceptance at depth: **123/202 = 0.61** (vs 0.76-0.84 short-ctx; matches overnight 0.55 @ 36k — acceptance degrades with ctx depth, expected for MTP)
+- Prompt was completely cached-cold (`cached_total 0`), `n_tokens_max = 92286`
+- Answer truncated inside ` thinking` (max_tokens 192 too small — harness artifact, not a defect; pi sends 262144)
+
+## Memory
+- No OOM; RSS +0.7 GB during the soak (48→49 G used of 125 G; 75 G available). KV for 92k ≈ 7.5 GB F16, absorbed by the unified pool.
+- Extrapolation to full 262,144: ~21.5 GB F16 KV → peak ~62-66 G RSS. **Comfortable — KV cache quantization NOT needed**; keep F16 for max long-context recall.
+
+## Decisions
+- **ubatch-size 2048 (lock-in)** — see sweep table below; peaked at 2048, regressed at 4096.
+- KV quant skipped (no pressure).
+- Acceptance-at-depth 0.61 is inherent to MTP at long ctx; n-max 3 remains the best setting.
+
+## ubatch sweep (same 92,095-token prompt, restart per variant)
+
+| ubatch | batch  | prompt tok/s | gen tok/s | draft acc |
+|--------|--------|--------------|-----------|-----------|
+| 512 (default) | 2048 (default) | 460.1 | 45.8 | 0.61 |
+| 1024   | 2048   | 519.7 (+13%) | 47.9 | 0.65 |
+| **2048** | 2048 | **555.5 (+21%)** | 59.7 | 0.93 |
+| 4096   | 4096   | 524.8 (−6% vs 2048) | 60.3 | 0.95 |
+
+- Optimal at 2048; 4096 regresses (larger physical batches overshoot GPU efficiency on this APU).
+- `batch-size` left at default 2048 (physical cap is ubatch).
+- Final child args verified: `--spec-draft-n-max 3 --flash-attn on --n-gpu-layers 99 --ubatch-size 2048`.
+- Post-lock sanity: 72.4 tok/s short chat, 11/11 accepted.
\ No newline at end of file
diff --git a/llama/results-spec-sweep-2026-09-04.md b/llama/results-spec-sweep-2026-09-04.md
new file mode 100644
index 0000000..94ff290
--- /dev/null
+++ b/llama/results-spec-sweep-2026-09-04.md
@@ -0,0 +1,24 @@
+# Spec-draft sweep — qwen3.6-35b-a3b (2026-09-04)
+
+Same-day A/B, same build (b210), same battery (3× short-chat, 3× reason, 3× tool, 1× ~14.4k-token codebase Q&A), `draft-mtp` fixed, only `spec-draft-n-max` varied. Acceptance/mean-len from fresh `/metrics` counters (child restarted per variant → totals = battery-only).
+
+| n-max | acceptance | mean draft len | chat gps | reason gps | tool gps | code gps |
+|-------|-----------|----------------|----------|------------|----------|----------|
+| 2     | 0.836     | 2.00           | 61       | 67         | 61       | 65.2     |
+| **3** | 0.760     | **3.00**       | 62       | **72**     | 61       | **71.7** |
+| 4     | 0.674     | 3.98           | 56       | 74.5       | 62       | 59.9     |
+
+## Decision: keep `spec-draft-n-max = 3`
+
+- n-max 3 posts full 3-token draft chains (mean len 3.00) with best agentic (code) and reasoning throughput.
+- n-max 4 over-drafts: acceptance falls to 0.67, code gen drops ~12 tok/s vs n-max 3 despite full 4-token chains — MTP useful depth exceeded for this model.
+- n-max 2 under-drafts (mean 2.00): less per-step throughput.
+
+## Locks / verification
+- `models.ini` `[qwen3.6-35b-a3b]`: `spec-type=draft-mtp`, `spec-draft-n-max=3` (qwen3.8-27b stanza untouched at 4).
+- Router restarted; child args verified via `/v1/models`: `--spec-draft-n-max 3 --spec-type draft-mtp`.
+- Post-restart sanity: gen 70.7 tok/s, 11/11 accepted.
+
+## Notes
+- Counters reset on child restart; Prometheus scrape deltas across restarts are invalid — always read `/metrics?model=...` fresh per test window.
+- Prior overnight run (36k ctx, n-max 2): acceptance 0.55 / mean 2.10 — long-context acceptance is lower than short-ctx; keep that in mind for the 262k soak.
\ No newline at end of file