โ€บNavigation

Iteration 45 โ€” ADMIT #9: Causal Dantzig ADMIT (FOREX) ยท BOTH LEGS PASSED WHERE ONE SUFFICED

In plain words: the regime-subtraction candidate finished the ladder the strong way. The final hurdles were the practical ones: fed into our verdict machinery on strictly held-out later regimes, do its readings actually make verdicts better? They do โ€” forecast quality rose by a small but statistically unambiguous margin (the entire uncertainty band is positive, and it sits far outside what lucky pairing alone produces). And separately, its "how close to regime-proof is this feature?" score genuinely predicts which features keep their edge into the next regime. The certification rule required winning either test; it won both. Ninth certificate issued, with a role nothing else on the shelf covers: a steadiness meter that stays honest even when a hidden common cause is pulling the strings โ€” plus it came through the whole ladder with zero tuning dials. One curiosity for the record: on this market, features closer to regime-proof are MORE likely to keep their edge โ€” the mirror image of what we once measured on crypto. One candidate left in the batch, then the batch is done.

Preflight (A0): 2026-07-10 02:05 โ€” load1 2.96/32c (โ‰ค24) ยท two light claude sessions ยท 37 GiB avail (si trickle/so=0, scope swap-banned) ยท CH + sidecar + kintsugi active โ†’ ALL PASS. Sync: origin/main unchanged. One capped run, 3.02 min, readonly=2 loader; candidate values deterministic; inference = the sealed apparatus' pairing-null + block bootstrap (real-value shuffles per ยง0).

Ladder finale (pre-registered in causal_dantzig_m4_m5b.py โ€” the sealed row-13 apparatus, frozen disjunction)

LegReadingRuling
Leakage-safe candidate columnExpanding cd closeness = โˆ’d_min over FX_ORDER windows 0..t inclusive (the k_v vrate convention); sessions excluded as leak-prone; exam machinery verbatimSEALED CONVENTION
M4 held-out incremental valueฮ”AUC = +0.00727, CI95 [+0.00335, +0.01154] โ€” entirely positive AND fully outside the pairing-null band [โˆ’0.0064, +0.0004]; baseline 0.8774 โ†’ 0.8846 on 2,642 held-out transitions; max |ฯ| vs base gates 0.228 (not a repackaging)PASS
M5โ€ฒb directionฯ(cd, keeps-orthogonality) = +0.1255, combo-block bootstrap CI [+0.0925, +0.1538] > 0, shuffle-null 97.5th pct 0.0419 โ€” far clear (n = 2,393)PASS
Decision rule noteThe iter-44 steering said "clear both"; the frozen forex precedent (row-13 admit line, rows-34/35 sealed disjunction) is M4 OR M5โ€ฒb โ€” followed and recorded. Moot: both passed.RECORDED
Discovery note (not gated)On forex, CD-closer-to-invariant features are MORE likely to keep orthogonality (+0.126) โ€” the mirror image of the iter-9 crypto inverted-signal finding; logged for the regime-lens questionLOGGED
Verdict (row 56)ADMIT(forex, 8-window EURUSD@5 regime grid, role = hidden-confounder-tolerant moment-difference invariance meter โ€” d_min readout, expiry = next completed forex regime slice) โ€” ADMIT #9; full trail rows 54 โ†’ 55 โ†’ 56 (exam 15-orders separation, zero dials ยท M3 informative/non-redundant ยท finale both legs)

โ–ถ Next iteration

Iteration 45 ยท 2026-07-10 ยท ADMIT #9 (row 56) ยท capped (3.02 min) ยท readonly=2 ยท zero generated values ยท append-only ยท evidence: causal_dantzig_m4_m5b.py ยท causal_dantzig_m4_m5b_results.json