Phase 2 ESL verdict + Step 5 decision gate

[pixiiidust]

What to build

The Phase 2 ESL verdict and the project's steering decision (HITL — maintainer makes the final call).

Produce the one-page report from plan §15 and state explicitly which §8 outcome holds: ESL = SC × DD works / is partially useful / does not compress the surface well — backed by the §8 decision rule (monotonicity, bucket variance, boundary alignment vs contours, ranking quality vs SC alone/DD alone/token count), on both label and realized ESL. Run the weighted fallback SC^a × DD^b only if simple ESL fails. The issue closes when the maintainer records the Step 5 decision: proceed to ToolDAG / weighted calibration / different metric / adjust grid / fix harness.

Reference: docs/plan.md §8, §13 Steps 4–5, §15.

Acceptance criteria

• One-page report per the §15 template committed
• Verdict states one of the three §8 outcomes with supporting evidence
• Equal-ESL cell comparison included on both label and realized ESL
• Weighted ESL run only if simple ESL failed
• Maintainer's Step 5 decision recorded in a comment on this issue

Blocked by

• Run the 200-episode StatefulPuzzle grid + Phase 1 report #6

[pixiiidust]

Step 5 decision recorded (maintainer, 2026-07-04): Option C — calibrate and transfer in parallel.

1. Run the weighted calibration ESL_weighted = SC^a × DD^b on the existing 240 deduplicated Phase 1 episodes (analysis only, zero new spend). Expect a > b. Report whether any (a,b) beats SC+DD and token count on the §8 diagnostics, on both label and realized coordinates.
2. In parallel, run the locked ToolDAG-B DD-sweep trigger gate (issue ToolDAG transfer run (Phase 3) #8, configs/tooldag_b_gate_dd_sweep.json, ~$1–3). Full Phase 3 grid only on a PASS.
3. This issue closes when the §15 one-page verdict report is committed with the calibration results folded in.

Phase 1 evidence: docs/reports/phase1-scdd-surface.md + docs/reports/phase1-report.html. Headline basis: SC×DD is not a clean multiplicative law (Spearman 0.781 vs SC+DD 0.806; equal-ESL spread up to 0.90 at the boundary) but DD genuinely bends the boundary at SC ≥ 20 — the §8 'partially useful, needs calibration' outcome.

[pixiiidust]

Closing — all acceptance criteria met.

• §15 one-page verdict committed: docs/reports/esl-verdict.md (commit 5fa535d), calibration results folded in.
• §8 outcome stated with evidence: partially useful, needs calibration — simple SC×DD beaten by SC+DD/token count on Phase 1 (0.781 vs 0.806/0.857); weighted fit SC¹×DD^0.48 reaches 0.858, tying token count. Equal-ESL comparison on both label and realized coordinates (spread 0.00 at ESL≤30, up to 0.90 at the boundary; realized = label except the saturated (5,6) cell).
• Weighted ESL run only after simple ESL failed (PR feat: weighted ESL calibration SC^a x DD^b (#7) #18).
• Step 5 decision recorded above (Option C, 2026-07-04) — both tracks executed: calibration merged, ToolDAG-B gate passed, Phase 3 grid complete.

Full-project report: docs/reports/esl-full-report.html (+ PDF).