Iteration 31 — STABLE screen S-10: 7/7, the fourth perfect screen

Iteration 31 — STABLE screen S-10 7/7 PASS B-07 THREE PASSERS AWAY

2026-07-09 · thirty-first firing · ledger row 36 · screening batch S-10 · campaign branch loop/matrix-harvest-campaign

In plain words: the depth-phase cohort passed intact — the fourth consecutive perfect screen, and by now the pattern is a law of the campaign: when instruments are harvested against the screen's own bar, the screen finds only survivors (implemented-first entries arrive with their trust anchors already attached). One passer deserves special mention: the refutation battery is anchor-native by design — its placebo refuter literally IS a permutation null and its add-a-confounder refuter IS a planted known-answer, so the trust-anchor question answers itself. The Andrews tests passed carrying their graveyard-fence flag forward to the lab's real-data gate. The implication enumerator passed with a composition note: its battery of small tests plugs into the FDR layer already in the catalog — instruments beginning to interlock. Seven passers puts the pending STABLE queue at seven — three short of the third STABLE handoff, which one mini-harvest will close.

The seven passers

id	Instrument	M1 anchor
H-096	CD-NOD	kernel-CI p-values + block shuffle
H-099	Andrews sup-Wald	sup-F critical values; boneyard-fence + heavy-tail flags travel
H-101	DoWhy refutation battery	anchor-native by design — refuters ARE nulls/known-answers
H-102	E-value	closed-form bound + observed-covariate benchmarking
H-103	J-PCMCI+	per-link p-values (lineage precedent)
H-104	Robustness value (sensemakr)	closed-form bounds + benchmarking practice
H-105	Implication battery (dagitty)	per-implication p-values + in-house FDR layer (composition note)

Instrument

M1 anchor

H-096

CD-NOD

kernel-CI p-values + block shuffle

H-099

Andrews sup-Wald

sup-F critical values; boneyard-fence + heavy-tail flags travel

H-101

DoWhy refutation battery

anchor-native by design — refuters ARE nulls/known-answers

H-102

E-value

closed-form bound + observed-covariate benchmarking

H-103

J-PCMCI+

per-link p-values (lineage precedent)

H-104

Robustness value (sensemakr)

closed-form bounds + benchmarking practice

H-105

Implication battery (dagitty)

per-implication p-values + in-house FDR layer (composition note)

Coverage counters after this firing

Boundary	Minimum	Verified	Screened-pass	Fed
STABLE	100	43	30 (23 fed + 7 pending)	23
FRAGILE↔CONDITIONAL	60	14	12	12
CONDITIONAL-scoping	40	14	12	12
FAIL	40	13	10	10
EXPIRY	40	16	11	11
cross-cutting	20	5	0	0
Total	300	105	75	68

Boundary

Minimum

Verified

Screened-pass

Fed

STABLE

100

30 (23 fed + 7 pending)

FRAGILE↔CONDITIONAL

CONDITIONAL-scoping

FAIL

EXPIRY

cross-cutting

Total

300

105

▶ Next iteration

Iteration 32 = STABLE mini-harvest #4 (implemented-first; 3+ passers close B-07). Veins: R invariance remainder (e.g., `CompareCausalNetworks`-style aggregation diagnostics, tram/tramicp — transformation-model ICP, a 2023+ implemented line), py-why `pywhy-graphs` implication utilities, and the GROW-variant residue. After B-07: cross-cutting completion (15-gap, cheapest minimum), then FRAGILE↔COND depth.

Standing: firecrawl host still down (5 revisits parked) · git-HTTPS transient timeouts recurring (pushes retried successfully so far).

Steering point: no directive = STABLE mini-harvest next firing. Directives are ledger-recorded verbatim before taking effect.

Machine-readable record: ledger row 36 + catalog §S-10. WebSearch: blocked this firing (recorded). Pages are append-only.