diff --git a/docs/claude/research-tree.md b/docs/claude/research-tree.md index c7456a368..b86f79d6c 100644 --- a/docs/claude/research-tree.md +++ b/docs/claude/research-tree.md @@ -34,6 +34,7 @@ Rules: | P3 | O1 with size-matched K = 8 partitions | — | — | PRUNED (eligibility 23 %; lesson: control cut area as a covariate, do not match it) | Log (newest first): +- 2026-11-18: **S2 last-fill diagnostic (descriptive, post hoc, no gate; MoldAstra spec 05344bac; result 98519a6c).** 552 / 204. Full-node peak class A / H2: near the true peak 153 / 129; inside the true last-1 % tail but away from the peak 350 / 350; *outside the true tail (far)* 49 / 71 (component-weighted 9.4 % / 16.4 %); exact tie at another true peak 0 / 2. The γ > 2ε certificate holds for 0 / 552 in every combination (errors are too large relative to peak margins for the sufficient condition, so it is uninformative). **Reading:** H2's last-fill loss is mostly *more far-from-tail picks* plus fewer near-peak picks, not exact-tie flips; my hypothesis (b) (metric fragility) is at most a partial explanation. For about 63 % of parts both models land inside the true late region but not at its peak, so the set-based endpoint matters. This supports item c (location-aware objective) as designed; nothing is decided by this diagnostic. - 2026-11-18: **Phase 3 item 1 (data sufficiency, v2) running.** Design revised from e24fff7c: six fresh fits (25 % / 50 % at 200 epochs; 50 % and 100 % each at U = 32k and 2U = 64k steps with their own complete schedules), 100 % 200-epoch reused from S1; U vs 2U = joint budget+schedule change. Pre-run review PASS (MeltAssist cbd67318); cap 120,000 GPU-s. S2 published (2338dfe9) with independent terminal audit PASS (d4999e4a); folds 1-4 alone: field Δ -.00495, CI [-.00765, -.00248]. - 2026-11-18: **H2 CONFIRMED on field (pre-registered pooled rule), last fill WORSE.** 552 / 204 test parts, pre-score review PASS 9c163702; postfit launched by TRON as a labelled one-time exception (MeltAstra unreachable). Field H2 .06579 vs A .07083: Δ -.00504, CI [-.00754, -.00283] (about 7 % relative); vs B13-E1b Δ -.00391, CI [-.00631, -.00164]. τ +.0035, CI [+.0010, +.0062]. **Last fill H2 .4566 vs A .3767: Δ +.080, CI [+.007, +.156] (worse).** Boundary: the pooled rule was registered after fold 0 was seen, so fold 0 is not independent; folds 1-4 alone to be reported descriptively. Reading: the learned representation is the lever for the field (better ordering, lower error), but it puts the last-filled point in the wrong place more often than A. Phase 3 is triggered (data sufficiency → seed 1 → c last-fill in H2 → b → seed 2 / ensemble); c is now the priority question. - 2026-11-17: **Gate calibration v2 FAIL_OUTCOME_EQUIVALENCE** (pre-registered, review PASS before scoring). Label-derived minus LOG gates, 545 paired parts / 202 components: field A +.00083, CI [+.00039, +.00134]; E1b +.00077, CI [+.00035, +.00125] (both exceed ±.001); last/Rg CIs exceed ±.010. Cost 94,261 / 104,639. **Consequence:** the v2 fresh holdout is BLOCKED_INPUT for gate-dependent methods; the shift is of the same size as the effects we test (E1b gain .0011). The A and E1b shifts are similar, so a paired method difference *might* cancel, but that is an observation from this result, not a pre-registered test, and will not be used. Route forward: real gate metadata for v2 (question to Yanqi), or a different fresh source. Phase-3 item c's fresh-family stage is blocked accordingly.