diff --git a/roadmap.md b/roadmap.md
index 24582db3..7cab9089 100644
--- a/roadmap.md
+++ b/roadmap.md
@@ -149,6 +149,40 @@ category's pass rate stops improving under the repair loop.
 
 Exit: a measured pass rate per category, with the failure taxonomy Phase 0 calls for.
 
+Measured (width equation, the bootstrap order's first category): the seam
+(`crates/layout/src/width.rs` -- `WidthInputs`/`WidthOutputs`/`WidthResolver`) isolates
+CSS 2.1 10.3.3's auto-margin balancing and over-constraint resolution behind one
+plain-data function, depending on neither `dom` nor Stylo, so a candidate crate compiles
+against it alone (`crates/generation` -- `candidate_crate::write`).
+`bin/ground_truth_cache.rs` caches the 227-case corpus's engine dumps to a gitignored
+directory; every step after that reads the cache and never starts a browser
+(`docs/infrastructure.md` -> "Verification tier"). A hand-written discrimination proof
+(`bin/generation_loop.rs`) ran four candidates through the permanent regression set gate
+(`roadmap.md` -> "Re-verify"): the baseline hand rule 94/227 passed (seeding the
+regression set, since nothing had been admitted before), a candidate implementing auto
+margins and direction-aware over-constraint resolution 112/227 and admitted, and two
+seeded-wrong candidates -- one that centers every box regardless of whether its margins
+are actually `auto`, one that recalculates the wrong margin under over-constraint -- 0/227
+each and rejected, both losing every one of the 94 previously-passing cases. The
+admitted candidate's gain over baseline is exactly the `MarginAutoLeft` (0/10 to 10/10)
+and `MarginAutoBoth` (0/8 to 8/8) categories; `MarginAutoRight` and `OverConstrained`
+were already correct under baseline by construction (the categories' own recipes put
+the auto or over-constrained margin on the side whose used value the corpus's geometry
+never observes). Percentage resolution, margin collapsing and min/max clamping
+categories remain at 0 passed, outside this seam's scope. Provenance record:
+`corpus/candidates/provenance/correct.provenance.json` (gitignored, not committed --
+regenerate with `bin/generation_loop.rs`).
+
+A small laptop-model run (not the rig; `GENERA_MODEL`, an OpenAI-compatible endpoint,
+`model::RemoteCandidateSource`) against the same seam and gate did not produce a
+candidate: the endpoint answers `GET /v1/models` (reporting the target model as
+`"loaded"`) but every `POST /v1/chat/completions` attempted -- including a one-word,
+`max_tokens: 10` request sent directly, outside this project's own client -- did not
+return within 120 seconds, and three retries with backoff each exhausted a 600-second
+timeout before `bin/real_model_run.rs` gave up. Router-level reachability and
+completion-level responsiveness are different failure modes, and only the first one
+held.
+
 Throughout: the determinism check runs from the first generated algorithm, and the
 constraint space is complete even though nothing caches yet.