Iteration 3 — First M0/M1 screen batch: 10 pass, feed threshold triggered

Iteration 3 — First M0/M1 screen batch 10 PASS · 6 OUT FEED THRESHOLD TRIGGERED

2026-07-04 · third firing · ledger row 5 · screening batch S-01 · campaign branch loop/matrix-harvest-campaign

In plain words: harvesting paused for one firing; interrogation began. Every catalogued STABLE-boundary instrument faced the two paper-questions the certification lab poses first: M0 — does your evidence actually move one of our four feature-status boundaries? and M1 — can your readings be checked against anything (a null to shuffle, an error bound, a known answer)? The default answer is "out" — an instrument must earn its way in. Ten of sixteen earned it. Six were shown the door with the reason on the record: three theory papers (kept as citation anchors, not gauges), one raw score with no way to catch it lying, one meta-diagnostic aimed at our environment definitions rather than our features, and one that needs the ability to run experiments — which markets do not grant. Ten passers means the first handoff batch to the admission lab is now due — the campaign's first deliverable, next firing.

The screen (M0/M1, verbatim gate definitions, default = OUT)

16 screenable STABLE entries │ ▼ M0: name exactly ONE boundary + the failure mode you detect there 4 OUT (OFF-GOAL): H-010, H-020, H-026 (theory anchors — retained as citations, not gauges) H-024 (environment meta-diagnostic, not feature evidence) 1 OUT (SUBSTRATE-HOPELESS): H-022 Active ICP — needs experiments; markets are observation-only │ ▼ M1: name a matched null / error bound / ground-truth anchor, constructible on OUR data 1 OUT (POLICY): H-009 IRM penalty — raw score, no anchor (bootci_w/stab_freq precedent) │ ▼ 10 SCREENED-M0M1-PASS ──► feed threshold (≥10) TRIGGERED → batch B-01 next firing

The ten passers (and their trust anchors)

id	Instrument	M1 anchor	Note
H-001	GROW e-statistics (group invariance)	e-value error bound (Ville)	already admission candidate A — flagged so the lab can skip
H-002	Koning e-values (exchangeability/invariance)	e-value error bound	candidate A family — flagged
H-004	PCMCI+	per-link p-values, autocorrelation-aware + block-shuffle null	autocorrelation-native
H-016	seqICP	type-I-controlled test (JASA)	time-structured environments, no labels needed
H-017	Stabilized Regression	env-permutation null + duplicate-feature known-answer	anchor constructed, not paper-native — M2 must confirm collapse
H-018	Causal Dantzig	asymptotic CIs (+ HAC/block here)	hidden-confounder angle
H-019	Invariant Probabilistic Prediction	proper-scoring-rule test (paper-native)	distributional invariance — tails included
H-021	EILLS	selection-consistency bound + env-permutation null	borderline — noted for M2
H-023	Invariant Ancestry Search	FWER control (paper-native)	higher-power weak readout
H-025	HSIC-X instrument validity	HSIC block-permutation null	guards the anchors themselves — a slot no stack gate covers

Instrument

M1 anchor

Note

H-001

GROW e-statistics (group invariance)

e-value error bound (Ville)

already admission candidate A — flagged so the lab can skip

H-002

Koning e-values (exchangeability/invariance)

e-value error bound

candidate A family — flagged

H-004

PCMCI+

per-link p-values, autocorrelation-aware + block-shuffle null

autocorrelation-native

H-016

seqICP

type-I-controlled test (JASA)

time-structured environments, no labels needed

H-017

Stabilized Regression

env-permutation null + duplicate-feature known-answer

anchor constructed, not paper-native — M2 must confirm collapse

H-018

Causal Dantzig

asymptotic CIs (+ HAC/block here)

hidden-confounder angle

H-019

Invariant Probabilistic Prediction

proper-scoring-rule test (paper-native)

distributional invariance — tails included

H-021

EILLS

selection-consistency bound + env-permutation null

borderline — noted for M2

H-023

Invariant Ancestry Search

FWER control (paper-native)

higher-power weak readout

H-025

HSIC-X instrument validity

HSIC block-permutation null

guards the anchors themselves — a slot no stack gate covers

Coverage counters after this firing

Boundary	Minimum	Verified	Screened-pass
STABLE	100	19	10
FRAGILE↔CONDITIONAL	60	0	0
CONDITIONAL-scoping	40	1	0
FAIL	40	0	0
EXPIRY	40	5	0
cross-cutting	20	1	0
Total	300	26	10

Boundary

Minimum

Verified

Screened-pass

Fed

STABLE

100

FRAGILE↔CONDITIONAL

CONDITIONAL-scoping

FAIL

EXPIRY

cross-cutting

Total

300

▶ Next iteration

Iteration 4 = publish feed batch B-01 (the rule fires it: STABLE screened-pass reached 10). The batch goes into FRONTIER-FEED.md cheapest-kill-first — likely order: the paper-native-anchor tests first (H-016 seqICP, H-023 IAS, H-019 IPP), then the constructed-anchor entries (H-017, H-021), then the heavier machinery (H-004 PCMCI+, H-018, H-025), with H-001/H-002 listed last under their "already candidate A" flag. This is a merge-worthy milestone per the LOOP-PROMPT operator notes ("first feed batch published") — after it lands, consider merging scaffold PR #581 and campaign PR #582 so the admission loop can consume the batch from main.

Steering point: no directive = B-01 publishes next firing. Directives (reorder the batch, hold the feed, resume harvesting instead) are ledger-recorded verbatim before taking effect.

Machine-readable record: ledger row 5 + catalog §Screening batch S-01. Pages are append-only.