diff --git a/docs/evals/2026-09-07-slice16-diagnostic.md b/docs/evals/2026-09-07-slice16-diagnostic.md index 19d37a3..1d0d68c 100644 --- a/docs/evals/2026-09-07-slice16-diagnostic.md +++ b/docs/evals/2026-09-07-slice16-diagnostic.md @@ -243,4 +243,4 @@ This confirms the six-way softmax competition is a significant contributor. ## 11. Commit hash for diagnostic tooling -(to be filled after commit) +`774c066` diff --git a/docs/evals/CLAUDE.md b/docs/evals/CLAUDE.md index bcedad3..69c45a6 100644 --- a/docs/evals/CLAUDE.md +++ b/docs/evals/CLAUDE.md @@ -48,6 +48,7 @@ A pair in `docs/routing.md` went stale unnoticed. Its source predated the | [The CPT+SFT Qwen3-1.7B routes better and cannot hold a sentence](2026-08-19-maven-model-cpt-sft.md) | live | | [Linear-head experiment on e5-small (slice 14 baseline, 136 examples)](2026-09-07-linear-e5-router-experiment.md) | superseded | | [Expanded corpus linear-head experiment (3025 examples, 93 seed families)](2026-09-07-expanded-corpus-linear-head-experiment.md) | live | +| [Slice 16 diagnostic: action/non-action boundary analysis](2026-09-07-slice16-diagnostic.md) | live | `docs/routing.md` holds the arm table these feed. Cite from there, not from here.