Wednesday, 8 July 2026 · Activity
ACTIVITY-CONFIRM batch — first real forward exams for the activity franchise (Fable session)
Owner: "ok, do it, go" on the forward-confirm proposal. Built scripts/activity_confirm.py (generalizes xauvol_confirm/gcvol2_confirm over cells) and ran 4 pre-registered one-look confirms — each cell's registry champion (byte-asserted from its study artifact) trained at the repo cap (2026-04-30 23:45), scored on the untouched post-cap realized window vs gcvol_har causal on identical folds. EXPERIMENT_LOG rows 1143–1146 registered before scoring; datahealth 0-fail on both panels; SI stale-load justified (feed halted 06-12; train caps at 04-30).
Verdicts vs registered gate (Δ<0 AND DM-HAC p<0.05 AND n≥150; Bonferroni ×4 reported):
- gc-atsz-h4 PASS (Δ −0.00171, p 5.4e-06, n=1093, 5/5 seeds, all 4 steps negative, 9.1 wks)
— first forward-confirmed activity cell in the program. In-study Δ −0.0045 → ~38% transferred.
- si-atsz-h12 PASS (Δ −0.0100, p 3.3e-16, n=234, 5/5) — replicates at 103% of in-study
size; cleanest replication measured to date. But SI has no live feed (halted 06-12).
- si-logvol-h12 FAIL (Δ −0.0009 from in-study −0.0619 (!), p 0.45, 3/5 seeds) — the franchise's
biggest in-study delta collapsed to 1.4% of itself. Honest kill.
- gc-atsz-h8 FAIL-but-right-sign (Δ −0.0007, p 0.153, n=273 sparse step-16 folds, 5/5 seeds
negative) — underpowered observation, not a kill.
Mechanics note: first run crashed my own fold-grid guard on the h8 cell (base step_size 16 → len(grid)=n_folds·h < test_size); fixed the guard only (winner/folds/metric untouched, one-look clause) and reran remaining cells. Adversarial artifact-only review dispatched; verdicts final only after it returns (report: _tmp/2026-07-08_activity-confirm-batch.md).
Takeaways: in-study delta magnitude does NOT predict forward transfer (1.4% / 38% / 103% transfer on three cells); avg-trade-size transfers on both instruments while both log-volume "stars" failed forward; the confirm engine is now standing (~5 GPU-min per cell, same-day verdict). Bucket-1 implication: GC trade-size context panel now has its forward-confirmed edge at h4; SI-atsz's full-size replication is the strongest argument yet for reviving the SI feed. Also: yesterday's challenger cron ran clean unattended (n=193, not-passing, 0 backfilled), and campaigns/xau-spot-logvol-frame-h4's missing campaign.md = never-written loop scaffold, benign.