iteration 34 · 2026-07-24 · detector/lead axis · terminal · ★ final instrument · laptop-drives-bigblack
🛰️ #18 Detector lead-vs-coincide GROUNDED
Does the detector D genuinely lead the event E (moves first, predictive) or merely coincide with it — or worse, lag it? Three independent legs — a lead-lag with a block-bootstrap CI, a transfer-entropy directional test, and an MMD²_u distributional test — all agree, and a partial-transfer-entropy harden defeats the latent-common-cause trap. All 3 gates ground, robust across 3 seeds. This is the last instrument.
★★ CAMPAIGN COMPLETE — 20 / 20 terminal
19 GROUNDED · 1 REJECTED (#12, standard knockoffs don't hold on our data → routed to #13) · 0 parked · 0 unverifiable. The BOUNDED-STOP condition is met; the loop ends here.
1.0 · 0 blk
power · lead-bias
0.0 ×4
all four false-lead nulls
20 / 20
terminal (19 G · 1 R)
The three gates
| Gate (§7 row 18) | Result (3 seeds) | |
|---|
| Admit / power (a true lead) | joint-detect 1.0 (lead-lag ℓ*≥+1 & CI>0 & ΔTE>0,p<.05 & MMD p<.05 & KS) · lead-bias 0.0 bars | PASS |
| False-lead on 4 nulls | coincident 0.0 · lagging 0.0 · noise 0.0 · block-permuted 0.0 | PASS |
| Latent-common-cause Harden | naïve lead+TE fires 1.0 spuriously · partial TE(D→E\|Z) removes it 0.0–0.025 | PASS |
Three verify-before-report fixes — the fixes are the science
- Test at the detected lead lag. The planted lead is lag-2 (Dt→Et+2), but the TE and MMD legs were hard-coded to lag-1 and only saw the weak autocorrelation echo — power was stuck at 0.3. Aligning both legs to the detected ℓ* lifted power to 1.0. (Verifiers must test at the discovered operating point, not a default.)
- Proper two-sample MMD null. Block-permuting the pooled
[hi|lo] concatenation preserved the hi/lo split and inflated the null, so the distributional leg under-fired. A plain label-permutation null fixed it.
- Nonparametric partialling at a strict bar. A linear/coarse conditioning on Z leaks (D,E are monotone-nonlinear in Z); the residual is a within-Z-bin echo that shrinks as 1/bins². Per-Z-bin demeaning at 20 bins, judged at a strict internal p<.01 (the permutation test is itself a 5% test, so "removed" must mean below a stricter floor), collapsed the harden from 0.14 → 0.075 → 0.025 across the fix chain — robust across all seeds.
Campaign patterns (durable)
- PATTERN #2 conservative-calibration — nominal thresholds under-control on real long-memory heavy-tailed data → permutation / higher-percentile block-bootstrap CV (#10,#4,#5,#7,#16,#18).
- Long-memory effective-N floor — power flat in N at frozen operating points → publish a power/N envelope or take an operator ruling (#14,#4).
- Latent-variable models for uncorrelated proxies (#15) · full-admit precondition guards (#16) · test at the detected operating point (#18).
- Real-data-only (03c) throughout — every control is a deterministic transform of real values; no synthetic data anywhere; every GROUNDED verdict carries an evidence artifact.
Only a GROUNDED instrument may later judge a real feature. With 19 grounded, the battery is ready; #12's FDR-selection role is covered by the grounded #11 (CMI) + #13 (DML).