Matrix harvest and status-grounding loop

HALTED2026-07-04

An automated literature-sweeping loop ran 32 times to collect and paper-screen published statistical tests that could tell whether a data feature stays independent from one market period to the next; it reached 108 vetted candidates against a planned 300 and handed 68 of them to the sister evaluation lab before the operator told it to stop.

Why it carries this status

Lifecycle, not result. This says where the audit sits in its process — never whether what it found was good.

Operator, typed into the loop session during firing 32 (verbatim): "I want you to stop the loop." Effect: the 20-minute cron schedule (job 86a2f02e, launched at row 7) is CANCELLED; firing 32 was completed and shipped cleanly before halting (no uncommitted mid-slice state — the iteration-4 lesson applied).

findings/evolution/audits/2026-07-04-matrix-harvest-status-grounding-loop/LEDGER.md

2026-08-17. Derived from folder evidence, then adversarially challenged; the challenge pass was upheld.

Blocked on The operator (nasimubd). The 20-minute /loop cron was cancelled mid-campaign by an explicit operator instruction; LEDGER row 38 records the resume path ('re-launch per LOOP-PROMPT.md'). H0 §7 additionally leaves the operator the choice to resume, convert to slow standing-refill, or close the campaign. A secondary blocker is the firecrawl host on littleblack being off the tailnet, which parks 5 NO-IMPL revisits.

The verdict

The audit's own conclusion, reproduced in full from the source below. Not a summary — this is the document, rendered. Links inside it that point at unpublished files are shown as plain text rather than as links that would 404 here.

verdict.md — plain-English bottom line

Status (2026-07-04): SCAFFOLDED, NOT YET FIRED.

The certification lab next door (the admission loop) has proven it can judge instruments — but its pipeline of candidates is thin, and the STABLE-ORTHOGONAL status still has no instrument at all. This campaign is the supply chain: it will systematically sweep the literature (20 method families), verify every source, check every implementation, throw out duplicates of what we already trust or already killed, and stack 300–400 vetted candidates — STABLE first, but covering every status — into feed batches the admission lab works through at its own pace.

Nothing here issues verdicts. Kills and admissions stay with the admission loop. This loop's product is a deep, honest, deduplicated queue — so the lab never idles for lack of candidates again.

source: findings/evolution/audits/2026-07-04-matrix-harvest-status-grounding-loop/verdict.md

Still owed 9

Operator decision owed: resume the loop, convert it to a slow standing-refill cadence, or close the campaign (H0 §7 'park' clause).
Catalogue stands at 108 of the ≥300 minimum; every per-boundary minimum is unmet (STABLE 46/100, FRAGILE↔COND 14/60, COND-scoping 14/40, FAIL 13/40, EXPIRY 16/40, cross-cutting 5/20).
7 STABLE screened-pass entries are queued but unfed — feed batch B-07 needs 3 more passers.
3 catalogued entries (H-106, H-107, H-108) are still unscreened; mini-screen S-11 is the queued next slice.
4 of the 20 method families remain unswept (families 3 conformal, 8 conditional-independence and 10 resampling are the named ones).
5 no-implementation revisits (H-088, H-091, H-097, H-098, H-100) are parked pending the firecrawl host returning to the tailnet.
H-015 (Wasserstein-relaxed invariance) is still provenance-pending after repeated failed searches.
Campaign PR #582 was OPEN and mergeable at the stop.
verdict.md still reads 'SCAFFOLDED, NOT YET FIRED' and needs superseding to reflect the 32 firings.

What it claims, and what backs each claim 16

Every row pairs a claim with the file it came from and the verbatim text in that file. The sources sit above the deploy root, so the quote is embedded and the path is printed as text rather than linked — a link would resolve on a laptop and 404 here.

Claim	Evidence
The campaign's job was to source and paper-screen 300–400 candidate measurement instruments covering every feature-status boundary, with the STABLE-orthogonal gap first, and hand screened batches to a separate admission loop. ASSERTED target 300–400 candidate instruments (cap 400)	The candidate-sourcing campaign for the matrix-admission machine: a standing loop that harvests, provenance-verifies, deduplicates, and M0/M1 paper-screens 300–400 candidate matrices (measurement instruments), covering every status boundary — with the STABLE-ORTHOGONAL instrument gap as first priority — and hands screened batches to the admission loop for the real-data gates. findings/evolution/audits/2026-07-04-matrix-harvest-status-grounding-loop/CLAUDE.md
At the moment the campaign was designed, the STABLE ceiling boundary had no instrument at all — the gap that motivated the whole harvest. CONFIRMED STABLE instruments available at start = 0; harvest minimum ≥ 100	\| STABLE ceiling \| VACANT — ICP identification falsified BLIND-GAUGE (sibling row 17); e-value family in entrance exam \| ≥ 100 \| findings/evolution/audits/2026-07-04-matrix-harvest-status-grounding-loop/H0-HARVEST-PRE-REGISTRATION.md
The frozen rulebook set per-boundary harvest minimums summing to a ≥300-entry catalogue: STABLE ≥100, FRAGILE↔CONDITIONAL ≥60, CONDITIONAL-scoping ≥40, FAIL ≥40, expiry/decay ≥40, cross-cutting ≥20. CONFIRMED 6 boundaries; minimums 100/60/40/40/40/20; total ≥300, cap 400	\| Total catalog target \| \| ≥ 300, cap 400 \| findings/evolution/audits/2026-07-04-matrix-harvest-status-grounding-loop/H0-HARVEST-PRE-REGISTRATION.md
The campaign is paper-only by construction: it was forbidden from touching ClickHouse at all, even for a read. CONFIRMED resource budget ≈2 cores / 2 GB / 5 minutes per step	1. No ClickHouse — not even SELECT. No credentials loaded, no `OPENDEVIATIONBAR_MODE=remote` calls. findings/evolution/audits/2026-07-04-matrix-harvest-status-grounding-loop/H0-HARVEST-PRE-REGISTRATION.md
The loop has no verdict authority: it may only catalogue, screen and feed — never admit or exclude an instrument. CONFIRMED 4 terminal states: CATALOGUED, SCREENED-PASS, SCREENED-OUT, FED	This loop never runs real-data gates, never writes to the admission loop's folder/branch/PR, and never issues ADMIT/EXCLUDE verdicts — its terminal states are CATALOGUED, SCREENED-PASS, SCREENED-OUT, and FED. findings/evolution/audits/2026-07-04-matrix-harvest-status-grounding-loop/CLAUDE.md
At the stop, the catalogue held 108 provenance-verified entries of the ≥300 target, 75 of which had passed paper screening and 68 of which had been handed off in six feed batches. MEASURED 108 verified / 75 screened-pass / 68 fed against a 300 minimum; per-boundary verified: STABLE 46, FRAGILE↔COND 14, COND-scoping 14, FAIL 13, EXPIRY 16, cross-cutting 5	\| Total \| 300 \| 108 \| 75 \| 68 \| findings/evolution/audits/2026-07-04-matrix-harvest-status-grounding-loop/CANDIDATE-CATALOG.md
All six status boundaries received at least one feed batch; the sixth batch completed boundary coverage of the status machine. MEASURED 6 batches: B-01 STABLE ×10, B-02 FAIL ×10, B-03 STABLE ×13, B-04 FRAGILE↔COND ×12, B-05 EXPIRY ×11, B-06 COND-scoping ×12 = 68 fed	Feed batch B-06 published — EVERY BOUNDARY IN THE STATUS MACHINE HAS NOW FED. findings/evolution/audits/2026-07-04-matrix-harvest-status-grounding-loop/LEDGER.md
Reported finding #1: the invariance-testing literature has far more published methods than working code — three of five depth-phase candidates died for having no implementation. MEASURED 3 of 5 candidates NO-IMPL in the paper-first depth harvest (n=5)	Depth-phase finding #1 reported (H0 §8): the invariance literature's CODE GAP — 3/5 died NO-IMPL; paper frontier ≫ implemented frontier. findings/evolution/audits/2026-07-04-matrix-harvest-status-grounding-loop/LEDGER.md
Inverting the search order to look for working implementations first eliminated the no-implementation casualties entirely. MEASURED implemented-first 8/8 catalogued (n=8) vs paper-first 2/5 (n=5)	3/3 CATALOGUED — implemented-first now 8/8 across D2–D4 (vs 2/5 paper-first) findings/evolution/audits/2026-07-04-matrix-harvest-status-grounding-loop/LEDGER.md
Reported finding #2: the hosted web-search channel blocks queries conditionally by phrasing — statistics-worded queries are filter-rejected while finance-worded ones pass. MEASURED ~1 WebSearch probe per firing across ~20 firings; majority BLOCKED when statistics-phrased	WebSearch this firing: 1 query, filter-BLOCKED (statistics-phrased; confirms row-9 conditional pattern — finance/regime phrasing passes, statistics-testing phrasing blocks). findings/evolution/audits/2026-07-04-matrix-harvest-status-grounding-loop/LEDGER.md
The pre-registration froze at a specific merge commit, making that SHA the campaign's tamper-evident anchor. CONFIRMED 1 freeze commit: 68596b5a97aa8b78b84e57a2660b33ae3c3ed891	Scaffold PR #581 MERGED 2026-07-07T00:06Z — squash SHA `68596b5a97aa8b78b84e57a2660b33ae3c3ed891` = the frozen H0 pre-registration hash findings/evolution/audits/2026-07-04-matrix-harvest-status-grounding-loop/LEDGER.md
Screening quality improved sharply once the harvest was scoped to the bar the screen actually applies — four consecutive screens passed everything they screened. MEASURED S-06 1/5 pass → S-07 5/5 → S-08 8/8 → S-09 4/4 → S-10 7/7	EXPIRY screen S-07 — 5/5 PASS, the native-guarantee bar vindicated (previous EXPIRY cohort went 1/5; harvesting against the screen's own bar is the cheap lesson of iters 19–21, recorded in the catalog cohort note) findings/evolution/audits/2026-07-04-matrix-harvest-status-grounding-loop/LEDGER.md
When a boundary could not reach the 10-entry feed threshold with quality candidates, the shortfall was reported rather than padded. CONFIRMED 2 recorded shortfalls: EXPIRY 5 of 10 (row 18), CONDITIONAL-scoping 8 of 10 (row 29)	EXPIRY = 5 screened-pass < 10 → NO feed trigger — honest shortfall reported per H0 §8; B-04 path = provenance-completion of the four pending + family-3/5/6/18 harvest. findings/evolution/audits/2026-07-04-matrix-harvest-status-grounding-loop/LEDGER.md
The last rung of the search-channel fallback ladder was never usable: the self-hosted firecrawl host was unreachable on first live test, parking five no-implementation revisits. CONFIRMED 5 parked NO-IMPL revisits (H-088, H-091, H-097, H-098, H-100)	Channel event: firecrawl-on-littleblack UNREACHABLE at its first live test — host absent from tailnet status; the ladder's last rung is theoretical until the host returns; five NO-IMPL revisits stay parked (H-088, H-091, H-097, H-098, H-100). findings/evolution/audits/2026-07-04-matrix-harvest-status-grounding-loop/LEDGER.md
The campaign's own completion criterion — catalogue ≥300 with every boundary minimum met and all 20 families swept — was never satisfied. REFUTED reached 108 of ≥300 (36%); 20 of 20 families touched, 16 swept; 3 entries left unscreened	COMPLETE when: catalog ≥300 with every §2 minimum met AND all 20 families swept at least once AND every entry is in a terminal state (screened or fed) AND all triggered feed batches are published. findings/evolution/audits/2026-07-04-matrix-harvest-status-grounding-loop/H0-HARVEST-PRE-REGISTRATION.md
The folder's root verdict document is stale — it still describes the loop as scaffolded and never fired, which was superseded by 32 recorded firings. REFUTED verdict.md dated 2026-07-04; LEDGER records firings through 2026-07-09 (iteration 32)	Status (2026-07-04): SCAFFOLDED, NOT YET FIRED. findings/evolution/audits/2026-07-04-matrix-harvest-status-grounding-loop/verdict.md

The audit folder 9 markdown files

Source of record: findings/evolution/audits/2026-07-04-matrix-harvest-status-grounding-loop/ — not published, so these are listed rather than linked.

File	Role
`CANDIDATE-CATALOG.md`	Append-only catalogue of the 108 provenance-verified instruments plus every screening and feed state transition.
`CLAUDE.md`	Hub: what the campaign is, the two-campaign division of labour, spoke map, conventions.
`FAMILY-TAXONOMY.md`	The 20 method families mapped to status boundaries, with harvest leads and a standing dedup watchlist.
`FRONTIER-FEED.md`	Append-only handoff batches B-01..B-06 — the only interface to the sibling admission loop.
`H0-HARVEST-PRE-REGISTRATION.md`	The frozen rulebook — objective, per-boundary minimums, entry schema, screening rules, feed rule, resource limits, stop shape.
`LEDGER.md`	Append-only iteration ledger, one row per firing, including the operator NOTE rows and the stop directive.
`LOOP-PROMPT.md`	Canonical /loop prompt (SSoT, laptop edition) and the resume instructions.
`SOTA-PROVENANCE.md`	Where the 'stable orthogonal tomorrow' idea comes from (ICP lineage), what the project already covers, and the recorded search-filter constraint.
`verdict.md`	Plain-English bottom line — written at scaffold time, now stale relative to the ledger.

Dashboard twin

findings/dashboard/campaigns/2026-07-04-matrix-harvest-status-grounding-loop/

Generated by findings/dashboard/build_audits.py from findings/evolution/audits/2026-07-04-matrix-harvest-status-grounding-loop/AUDIT_LEDGER.json — never hand-edited. Each quote was verified to occur in the file named beside it when the ledger was written.