Matrix Harvest & Status-Grounding Loop

Reconciliation — open · HALTED

Asked to catalogue 300 to 400 provenance-verified candidate instruments and feed screened batches to the admission lab. 32 firings delivered 108 catalogued and 68 fed before the operator stopped it at roughly a third of target.

What it measured

Quantity	Value	Meaning
Catalogue coverage	108 of a 300 target	192 entries short
Instruments fed to the lab	68 across 6 batches
Screened-pass entries	75	cleared the paper gates
STABLE boundary	46 verified of a 100 minimum	the campaign's own declared first priority
Cross-cutting boundary	5 verified of 20, 0 fed	the only boundary never fed

How it ran

Firing	What it did
01-02	Family-1 invariance harvest, 11 entries; the arXiv channel recovered
04	First feed batch handed to the admission lab
06	Rulebook frozen at merge, May attic imported, 50 entries
27	Sixth batch published — every status boundary has now been fed
32	Three more entries, then operator stop; schedule cancelled, queue preserved

What it produced

32 iteration pages plus an index carrying the per-boundary coverage scoreboard
The audit folder: candidate catalog, frontier feed, family taxonomy, harvest pre-registration and ledger
PR #582 — merged 2026-07-08

What is outstanding, and who owns it

192 catalogue entries short of the declared 300, with the STABLE boundary — its own stated first priority — at 46 of a 100 minimum. The resume queue is preserved and ordered. Needs the OPERATOR choosing among the three options the campaign itself names: resume, convert to slow-refill cadence, or close. Standing item: the self-hosted crawl host is down with 5 revisits parked. Worth flagging honestly — its consumer lab is itself parked with an 11-item backlog, so resuming is not urgent even though it is unfinished.

Why this status

Stopped mid-scope at roughly a third of target by an explicit operator directive, with a named ordered resume queue and a named decider.

'STOPPED BY OPERATOR 2026-07-09 (iter 32; resume via LOOP-PROMPT.md — queue preserved).' (manifest.json status)

Lifecycle status is process state, not a judgement of the findings. Results are stated as numbers with their uncertainty.

Matrix Harvest & Status-Grounding Loop ACTIVE · HARVESTING STOPPED BY OPERATOR 2026-07-09 · RESUME QUEUE PRESERVED

In plain words: the sibling campaign (the Matrix Admission Paradox Loop) is a certification lab for measurement instruments — but a lab is useless without a deep queue of candidates to interrogate. This loop is the sourcing expedition: it combs the scientific literature for 300–400 candidate instruments ("matrices"), verifies every citation against a real journal or archive (no hearsay), checks that working open-source code exists, throws out anything we already own or already rejected, asks two cheap paper-questions (does it serve one of our four status boundaries? can its readings be checked against anything?), and hands vetted batches to the lab. It runs on the laptop, touches no market data, and never issues verdicts — it feeds the machine that does.

The division of labor — two loops, one direction

THIS LOOP (laptop, paper-only) ADMISSION LOOP (bigblack, compute) harvest leads from 20 method families consumes FRONTIER-FEED batches as → verify provenance (journal DOI / arXiv) Tier-2 refills → runs the real-data → check FOSS implementation exists gates M2–M5 on actual market bars → dedup vs certified stack + boneyard → sole authority for ADMIT / EXCLUDE → tag ONE status boundary → screen M0/M1 → batch ≥10 passers per boundary ──────────► FRONTIER-FEED.md (the only interface)

Coverage scoreboard (recomputed every firing — target ≥300, cap 400)

Status boundary	Minimum	Verified entries	Screened-pass	Fed	Starvation (min − verified)
STABLE ceiling (first priority)	100	46	30	23 (B-01+B-03)	54
FRAGILE ↔ CONDITIONAL split	60	14	12	12 (B-04)	46
CONDITIONAL regime-scoping	40	14	12	12 (B-06)	26
FAIL line	40	13	10	10 (B-02)	27
EXPIRY / live decay	40	16	11	11 (B-05)	24
Cross-cutting (nulls under dependence)	20	5	0	0	15
Total	300	108	75	68	192

Status boundary

Minimum

Verified entries

Screened-pass

Fed

Starvation (min − verified)

STABLE ceiling (first priority)

100

23 (B-01+B-03)

FRAGILE ↔ CONDITIONAL split

12 (B-04)

CONDITIONAL regime-scoping

12 (B-06)

FAIL line

10 (B-02)

EXPIRY / live decay

11 (B-05)

Cross-cutting (nulls under dependence)

Total

300

108

192

Priority rule (frozen, H0 §2): serve the most-starved boundary first; ties → STABLE first, then cheapest family. Machine-readable counters: audit twin → CANDIDATE-CATALOG.md. Iter-2 recount note: seed split corrected to STABLE 8 / EXPIRY 5 (H-003 is EXPIRY per its row) — totals were always right.

Provenance discipline (why the numbers move slowly)

No instrument enters the catalog on reputation. Each row needs a primary source resolved this campaign (arXiv API, Crossref DOI, or publisher page) and a checked open-source implementation (license + activity noted) — or an explicit NONE-FOUND with a date. Hosted web search is filter-blocked for these topics, so the loop walks a fallback ladder: arXiv API → GitHub search → Crossref → PyPI → self-hosted firecrawl. A candidate whose paper cannot be resolved stays a lead, not an entry — quantity honesty beats padded counts.

▶ Loop stopped — resume queue preserved

The operator stopped the loop on 2026-07-09 (iteration 32); the recurring schedule is cancelled. The campaign parks at 108/300 with 68 instruments delivered across every boundary, PR #582 open, and the queue preserved in order: S-11 mini-screen → B-07 (third STABLE batch) → cross-cutting completion → FRAGILE↔COND depth. Standing items: firecrawl host down (5 revisits parked). To resume: paste the LOOP-PROMPT.md block into a repo session — every counter picks up exactly here. Operator options: resume · convert to slow-refill cadence · close the campaign.

Iterations (append-only)

#	Report	In plain words	Status
1	Family-1 invariance harvest: 6 new STABLE entries	Bootstrapped the campaign (worktree, dashboard twin, PR) and verified six invariance-testing instruments end-to-end — journal DOI + open-source check on every one	HARVESTED 6 · CHECKPOINT
2	PMLR leads recovered + the ε-frontier anchor: 5 new STABLE entries	The arXiv channel recovered; the three held-back candidates entered with verified papers, plus an instrument-validity test and the 2025 intractability proof that anchors the ε-approximate-invariance framing. Seed-counter off-by-one corrected on the record.	HARVESTED 5 · FAMILY 1 NEAR-SWEPT
3	First M0/M1 screen batch: 10 pass, 6 out	Every STABLE instrument faced the two paper-gates; ten earned their place with named trust anchors, six went out with reasons on record. The first handoff batch to the admission lab is now due.	10 PASS · 6 OUT · FEED TRIGGERED
4	Feed batch B-01 published: first handoff to the admission lab	The ten vetted STABLE instruments were packaged cheapest-kill-first and placed on the admission lab's intake shelf — the campaign's first deliverable. The interrupted prior firing was completed honestly on the record; the operator's relaunch directives (20m cadence, brute-force sweep, web-search re-probe — still filter-blocked) were ledger-recorded.	B-01 FED · 10 ENTRIES TERMINAL
5	Predecessor-import batch I-01: 12 rows, 8 dedup kills	Twelve instruments from the May catalog re-earned their place under today's rules — fresh provenance, fresh code checks, four-way dedup. STABLE deepened (+4 causal-discovery machines), the FAIL and FRAGILE↔CONDITIONAL boundaries opened their first entries, and eight attic items were rejected at the door with reasons on record. The hosted web-search channel answered a query for the first time in four days — eight new leads parked in the taxonomy.	12 IMPORTED · 8 KILLS · WEBSEARCH CONDITIONAL
6	H0 frozen at merge · branch rebased · import complete: 50 entries	The operator merged the scaffold PR — squash 68596b5a is now the rulebook's permanent fingerprint — and the campaign branch was rebased onto it with every record byte-verified. The dashboard-wipe race was root-caused and fixed by the merge itself. The May attic emptied: twelve more instruments entered (two immediately screened out for having no code anywhere — the rule is merciless). Every boundary now has at least one verified instrument.	H0 FROZEN · 12 IMPORTED · 50/300
7	Family 2 swept: nine anytime-valid monitors, 59 total	The anytime-valid school entered: provable-false-alarm change detectors, forever-honest monitoring bands, regularity-free tests, sequential independence and homogeneity tests, fat-tail-native quantile trackers. Two theory papers refused as padding, on the record. The FAIL boundary's first interrogation cohort is ready.	9 HARVESTED · FAMILY 2 SWEPT · 59/300
8	FAIL screen batch S-02: 10 pass, 1 duplicate out	The redundancy instruments faced the gates — harsher this time, since an incumbent gauge already holds the slot: each passer named dependence the incumbent cannot see. One conditional-independence test went out at the door (slot already occupied; a boneyarded predecessor died there). Second handoff batch due next firing.	10 PASS · 1 OUT · B-02 TRIGGERED
9	Feed batch B-02: the FAIL boundary's handoff	Ten redundancy instruments delivered cheapest-kill-first, with flagged borderline anchors and a four-cousins group note so the lab can kill efficiently. FAIL is the second boundary to complete the full harvest-screen-feed cycle; twenty entries now sit on the lab's shelf.	B-02 FED · FAIL CYCLE COMPLETE
10	DRO diagnostics + the anchor triad: 8 STABLE entries	Robustness instruments that put a number on how much shift a relationship survives, plus three econometric classics that guard the guards (anchor validity, strength, weak-anchor-robust inference). One raw-score entry pre-flagged; one near-duplicate refused. Families 12 and 20 swept.	8 HARVESTED · FAMILIES 12+20 SWEPT · 67/300
11	STABLE screen S-03: 13 pass, Neural Granger out on principle	Thirteen passed round two — causal-structure machines, stopping-safe tests, robustness margins, the anchor triad. Neural Granger went out with a principle attached: an instrument whose answers change between training runs cannot be caught lying, while its deterministic sibling passed for exactly that reason. Third handoff fires next.	13 PASS · 1 OUT · B-03 TRIGGERED
12	Feed batch B-03: STABLE's second handoff — 33 fed	Thirteen more STABLE instruments delivered cheapest-kill-first with every screening flag traveling along — exact-guarantee tests first, the anchor toolkit as a unit, known-risk entries last with risks in writing. Three batches, thirty-three instruments, every screened queue empty.	B-03 FED · 33 TOTAL · QUEUES EMPTY
13	EXPIRY screen S-04: 5 pass, honest shortfall	The live-decay monitors faced the gates: five passed (four with paper-native guarantees), the framework root went out on principle, and four scaffold-era entries were unscreenable — the no-hearsay rule bites the campaign's own early rows. Five is below the handoff threshold; padding forbidden, shortfall reported with the refill path named.	5 PASS · NO FEED · HONEST SHORTFALL
14	First FRAGILE↔CONDITIONAL harvest: 5 entries, MRP resolved	Selection-stability measures (with the correlated-features fix), top-weighted ranking overlap answering the incumbent's design bar, region-resolved dependence, and the web-search MRP lead resolved to a real JPM 2025 paper. arXiv was down all firing — three candidates wait rather than enter unverified.	5 HARVESTED · 72/300 · OUTAGE HONORED
15	arXiv recovered: distance cousins complete + null backbone	The archive paid its debts — kernel, transport, and regime-switching tests entered; three distribution-distance cousins now ship with a grouped-attack note. The stationary bootstrap landed as the citable, implemented backbone for the block-null every entry invokes.	4 HARVESTED · NULL BACKBONE · 76/300
16	FRAGILE↔COND screen S-05: 12 pass on the per-feature bar	The lucky-vs-conditional cohort was interrogated under a sharp principle: evidence must be about a feature, not the panel. Twelve cleared it; two elegant classics (shrinkage intensity, whole-matrix similarity) went out on exactly that principle, retained as context. Fourth handoff fires next — into the lab's hungriest slot.	12 PASS · 2 OUT · B-04 TRIGGERED
17	Feed batch B-04: into the lab's most-starved boundary	Twelve vetted lucky-vs-conditional instruments delivered to the slot the lab's own table marks hungriest (one trusted instrument today). Grouped-cousins notes, constructed-anchor flags, and two explicit beat-the-incumbent warnings travel with the batch. Forty-five instruments across four batches.	B-04 FED · 45 TOTAL · HUNGRIEST SLOT SERVED
18	Provenance debt cleared + null-machinery top-up	Four instruments unscreenable since day one got their citations resolved, unlocking EXPIRY's handoff path. The dependence-correction machinery every entry invokes (HAC, multiplier bootstrap) entered with sources and implementations, plus the offline changepoint segmenter. A wrong-identifier attempt and its recovery both on the ledger.	4 RESOLVED · 3 ROWS · 79/300
19	EXPIRY screen S-06: the debt-cleared cohort dies at the gate	Three of the four finally-cited instruments died at dedup — stack duplicate, weaker sibling of a passed variant, self-predicted collision — and the ensemble proved to be an adaptation system, not an instrument. Only the offline segmenter passed. Two honesty mechanisms composing as designed; the refill bar is now precise.	1 PASS · 4 OUT · KILLS ARE WINS
20	EXPIRY refill under the native-guarantee bar: five mechanisms	Only mechanisms with printed guarantees or stack-absent machinery entered: ADWIN (false-alarm bounds as theorems), the 1954 CUSUM root, density-ratio detection, spectral subspace, matrix-profile segmentation. The always-valid p-value line refused as an ancestor of catalogued machinery.	5 HARVESTED · 84/300 · S-07 → B-05 NEXT
21	EXPIRY screen S-07: 5/5 pass — the bar vindicated	Two firings ago this boundary went one-for-five; the massacre wrote the shopping list, the refill shopped against it, and all five passed. Lesson on the record: the screen's kills are the next harvest's specification. B-05 fires next.	5/5 PASS · B-05 TRIGGERED
22	Feed batch B-05: the decay-watchers ship — 56 fed	Eleven decay-watchers delivered with a baseline job description, M2-confirm flags, and a coordinate-don't-duplicate note on the lab's own designed-but-unbuilt monitor. Five batches, fifty-six instruments across the interface; regime-scoping is the last boundary never harvested for directly.	B-05 FED · 56 TOTAL · ONE BOUNDARY COLD
23	COND-scoping's first harvest: the last cold boundary warms	Seven regime-characterizing instruments entered — tail-dependence limits, the extremogram (built for dependent extremes), crash-vs-rally extremal graphs, exceedance correlation. Two excellent ideas exited NO-IMPL with revisit notes. Sixteen of twenty families now swept; the four holdouts are named.	7 ROWS · 91/300 · 16/20 SWEPT
24	COND-scoping screen S-08: 8/8 pass, gap of two	Second consecutive perfect screen — boundary-scoped harvests kill weak candidates at harvest time. Two squads passed: the declaration layer (FWER/FDR) and the characterization layer (tail structure). Eight is two short of a handoff; the rulebook doesn't round up, and the refill is named.	8/8 PASS · GAP OF 2
25	Family 14 falls: all 20 families touched	The forecast-comparison classics entered, built for dependent errors from day one — including the Giacomini–White conditional test, the regime-scoping question in native statistical form, and the set-valued Model Confidence Set. No blank continents left on the map.	4 HARVESTED · ALL 20 TOUCHED · 95/300
26	S-09: 4/4 pass, B-06 triggered	Third consecutive perfect screen from a boundary-scoped harvest — and the first cohort with zero hand-built anchors. Twelve regime-scoping passers; the sixth handoff fires next, completing feed coverage of the whole status machine.	4/4 PASS · B-06 TRIGGERED
27	B-06 published: every boundary has fed — the milestone	The structural goal reached: 68 vetted instruments across six batches cover every boundary of the status machine, each with its verified paper, checked code, and honest flags. The declaration-layer front half is the first with zero hand-built anchors. The map is drawn; the depth phase begins.	B-06 FED · 68 TOTAL · MILESTONE
28	Catalog crosses 100 — the code-gap finding	Entry #100 landed with a warning as the real product: three of five depth entries have no code anywhere. The finding is on the record — paper frontier ≫ implemented frontier — and the strategy inverts to implemented-first discovery. Andrews' tests entered with the boneyard fence explicitly drawn.	100/300 · CODE-GAP FINDING
29	Implemented-first works: 3/3 clean	The code-gap finding became method and paid for itself in one firing: the refutation battery (flagship library, updated yesterday), the E-value confounding-sensitivity measure, and J-PCMCI+ completing the causal-discovery quartet. Zero casualties; STABLE at 41 of 100.	3/3 CLEAN · 103/300
30	The sensitivity pair completes; firecrawl unreachable	Quantified hidden-confounder skepticism now covered in both regression and epidemiology units plus the empirical probe; the implication enumerator turns one grand claim into many falsifiable ones. The crawler failed its first live test — five parked instruments wait, honestly recorded.	2/2 CLEAN · 105/300 · CRAWLER DOWN
31	STABLE screen S-10: 7/7 — the fourth perfect screen	The depth cohort passed intact; the refutation battery proved anchor-native by design and the implication enumerator now plugs into the in-house FDR layer — instruments interlocking. Three passers short of the third STABLE handoff.	7/7 PASS · B-07 THREE AWAY
32	Final harvest + operator stop: the loop parks at 108/300	Three more STABLE instruments (implemented-first closes 8/8), then the stop directive — recorded verbatim, schedule cancelled, firing shipped whole. Six batches and 68 instruments delivered across every boundary; the resume queue is preserved on the ledger.	3/3 CLEAN · STOPPED CLEANLY

Report

In plain words

Status

Family-1 invariance harvest: 6 new STABLE entries

Bootstrapped the campaign (worktree, dashboard twin, PR) and verified six invariance-testing instruments end-to-end — journal DOI + open-source check on every one

HARVESTED 6 · CHECKPOINT

PMLR leads recovered + the ε-frontier anchor: 5 new STABLE entries

The arXiv channel recovered; the three held-back candidates entered with verified papers, plus an instrument-validity test and the 2025 intractability proof that anchors the ε-approximate-invariance framing. Seed-counter off-by-one corrected on the record.

HARVESTED 5 · FAMILY 1 NEAR-SWEPT

First M0/M1 screen batch: 10 pass, 6 out

Every STABLE instrument faced the two paper-gates; ten earned their place with named trust anchors, six went out with reasons on record. The first handoff batch to the admission lab is now due.

10 PASS · 6 OUT · FEED TRIGGERED

Feed batch B-01 published: first handoff to the admission lab

The ten vetted STABLE instruments were packaged cheapest-kill-first and placed on the admission lab's intake shelf — the campaign's first deliverable. The interrupted prior firing was completed honestly on the record; the operator's relaunch directives (20m cadence, brute-force sweep, web-search re-probe — still filter-blocked) were ledger-recorded.

B-01 FED · 10 ENTRIES TERMINAL

Predecessor-import batch I-01: 12 rows, 8 dedup kills

Twelve instruments from the May catalog re-earned their place under today's rules — fresh provenance, fresh code checks, four-way dedup. STABLE deepened (+4 causal-discovery machines), the FAIL and FRAGILE↔CONDITIONAL boundaries opened their first entries, and eight attic items were rejected at the door with reasons on record. The hosted web-search channel answered a query for the first time in four days — eight new leads parked in the taxonomy.

12 IMPORTED · 8 KILLS · WEBSEARCH CONDITIONAL

H0 frozen at merge · branch rebased · import complete: 50 entries

The operator merged the scaffold PR — squash 68596b5a is now the rulebook's permanent fingerprint — and the campaign branch was rebased onto it with every record byte-verified. The dashboard-wipe race was root-caused and fixed by the merge itself. The May attic emptied: twelve more instruments entered (two immediately screened out for having no code anywhere — the rule is merciless). Every boundary now has at least one verified instrument.

H0 FROZEN · 12 IMPORTED · 50/300

Family 2 swept: nine anytime-valid monitors, 59 total

The anytime-valid school entered: provable-false-alarm change detectors, forever-honest monitoring bands, regularity-free tests, sequential independence and homogeneity tests, fat-tail-native quantile trackers. Two theory papers refused as padding, on the record. The FAIL boundary's first interrogation cohort is ready.

9 HARVESTED · FAMILY 2 SWEPT · 59/300

FAIL screen batch S-02: 10 pass, 1 duplicate out

The redundancy instruments faced the gates — harsher this time, since an incumbent gauge already holds the slot: each passer named dependence the incumbent cannot see. One conditional-independence test went out at the door (slot already occupied; a boneyarded predecessor died there). Second handoff batch due next firing.

10 PASS · 1 OUT · B-02 TRIGGERED

Feed batch B-02: the FAIL boundary's handoff

Ten redundancy instruments delivered cheapest-kill-first, with flagged borderline anchors and a four-cousins group note so the lab can kill efficiently. FAIL is the second boundary to complete the full harvest-screen-feed cycle; twenty entries now sit on the lab's shelf.

B-02 FED · FAIL CYCLE COMPLETE

DRO diagnostics + the anchor triad: 8 STABLE entries

Robustness instruments that put a number on how much shift a relationship survives, plus three econometric classics that guard the guards (anchor validity, strength, weak-anchor-robust inference). One raw-score entry pre-flagged; one near-duplicate refused. Families 12 and 20 swept.

8 HARVESTED · FAMILIES 12+20 SWEPT · 67/300

STABLE screen S-03: 13 pass, Neural Granger out on principle

Thirteen passed round two — causal-structure machines, stopping-safe tests, robustness margins, the anchor triad. Neural Granger went out with a principle attached: an instrument whose answers change between training runs cannot be caught lying, while its deterministic sibling passed for exactly that reason. Third handoff fires next.

13 PASS · 1 OUT · B-03 TRIGGERED

Feed batch B-03: STABLE's second handoff — 33 fed

Thirteen more STABLE instruments delivered cheapest-kill-first with every screening flag traveling along — exact-guarantee tests first, the anchor toolkit as a unit, known-risk entries last with risks in writing. Three batches, thirty-three instruments, every screened queue empty.

B-03 FED · 33 TOTAL · QUEUES EMPTY

EXPIRY screen S-04: 5 pass, honest shortfall

The live-decay monitors faced the gates: five passed (four with paper-native guarantees), the framework root went out on principle, and four scaffold-era entries were unscreenable — the no-hearsay rule bites the campaign's own early rows. Five is below the handoff threshold; padding forbidden, shortfall reported with the refill path named.

5 PASS · NO FEED · HONEST SHORTFALL

First FRAGILE↔CONDITIONAL harvest: 5 entries, MRP resolved

Selection-stability measures (with the correlated-features fix), top-weighted ranking overlap answering the incumbent's design bar, region-resolved dependence, and the web-search MRP lead resolved to a real JPM 2025 paper. arXiv was down all firing — three candidates wait rather than enter unverified.

5 HARVESTED · 72/300 · OUTAGE HONORED

arXiv recovered: distance cousins complete + null backbone

The archive paid its debts — kernel, transport, and regime-switching tests entered; three distribution-distance cousins now ship with a grouped-attack note. The stationary bootstrap landed as the citable, implemented backbone for the block-null every entry invokes.

4 HARVESTED · NULL BACKBONE · 76/300

FRAGILE↔COND screen S-05: 12 pass on the per-feature bar

The lucky-vs-conditional cohort was interrogated under a sharp principle: evidence must be about a feature, not the panel. Twelve cleared it; two elegant classics (shrinkage intensity, whole-matrix similarity) went out on exactly that principle, retained as context. Fourth handoff fires next — into the lab's hungriest slot.

12 PASS · 2 OUT · B-04 TRIGGERED

Feed batch B-04: into the lab's most-starved boundary

Twelve vetted lucky-vs-conditional instruments delivered to the slot the lab's own table marks hungriest (one trusted instrument today). Grouped-cousins notes, constructed-anchor flags, and two explicit beat-the-incumbent warnings travel with the batch. Forty-five instruments across four batches.

B-04 FED · 45 TOTAL · HUNGRIEST SLOT SERVED

Provenance debt cleared + null-machinery top-up

Four instruments unscreenable since day one got their citations resolved, unlocking EXPIRY's handoff path. The dependence-correction machinery every entry invokes (HAC, multiplier bootstrap) entered with sources and implementations, plus the offline changepoint segmenter. A wrong-identifier attempt and its recovery both on the ledger.

4 RESOLVED · 3 ROWS · 79/300

EXPIRY screen S-06: the debt-cleared cohort dies at the gate

Three of the four finally-cited instruments died at dedup — stack duplicate, weaker sibling of a passed variant, self-predicted collision — and the ensemble proved to be an adaptation system, not an instrument. Only the offline segmenter passed. Two honesty mechanisms composing as designed; the refill bar is now precise.

1 PASS · 4 OUT · KILLS ARE WINS

EXPIRY refill under the native-guarantee bar: five mechanisms

Only mechanisms with printed guarantees or stack-absent machinery entered: ADWIN (false-alarm bounds as theorems), the 1954 CUSUM root, density-ratio detection, spectral subspace, matrix-profile segmentation. The always-valid p-value line refused as an ancestor of catalogued machinery.

5 HARVESTED · 84/300 · S-07 → B-05 NEXT

EXPIRY screen S-07: 5/5 pass — the bar vindicated

Two firings ago this boundary went one-for-five; the massacre wrote the shopping list, the refill shopped against it, and all five passed. Lesson on the record: the screen's kills are the next harvest's specification. B-05 fires next.

5/5 PASS · B-05 TRIGGERED

Feed batch B-05: the decay-watchers ship — 56 fed

Eleven decay-watchers delivered with a baseline job description, M2-confirm flags, and a coordinate-don't-duplicate note on the lab's own designed-but-unbuilt monitor. Five batches, fifty-six instruments across the interface; regime-scoping is the last boundary never harvested for directly.

B-05 FED · 56 TOTAL · ONE BOUNDARY COLD

COND-scoping's first harvest: the last cold boundary warms

Seven regime-characterizing instruments entered — tail-dependence limits, the extremogram (built for dependent extremes), crash-vs-rally extremal graphs, exceedance correlation. Two excellent ideas exited NO-IMPL with revisit notes. Sixteen of twenty families now swept; the four holdouts are named.

7 ROWS · 91/300 · 16/20 SWEPT

COND-scoping screen S-08: 8/8 pass, gap of two

Second consecutive perfect screen — boundary-scoped harvests kill weak candidates at harvest time. Two squads passed: the declaration layer (FWER/FDR) and the characterization layer (tail structure). Eight is two short of a handoff; the rulebook doesn't round up, and the refill is named.

8/8 PASS · GAP OF 2

Family 14 falls: all 20 families touched

The forecast-comparison classics entered, built for dependent errors from day one — including the Giacomini–White conditional test, the regime-scoping question in native statistical form, and the set-valued Model Confidence Set. No blank continents left on the map.

4 HARVESTED · ALL 20 TOUCHED · 95/300

S-09: 4/4 pass, B-06 triggered

Third consecutive perfect screen from a boundary-scoped harvest — and the first cohort with zero hand-built anchors. Twelve regime-scoping passers; the sixth handoff fires next, completing feed coverage of the whole status machine.

4/4 PASS · B-06 TRIGGERED

B-06 published: every boundary has fed — the milestone

The structural goal reached: 68 vetted instruments across six batches cover every boundary of the status machine, each with its verified paper, checked code, and honest flags. The declaration-layer front half is the first with zero hand-built anchors. The map is drawn; the depth phase begins.

B-06 FED · 68 TOTAL · MILESTONE

Catalog crosses 100 — the code-gap finding

Entry #100 landed with a warning as the real product: three of five depth entries have no code anywhere. The finding is on the record — paper frontier ≫ implemented frontier — and the strategy inverts to implemented-first discovery. Andrews' tests entered with the boneyard fence explicitly drawn.

100/300 · CODE-GAP FINDING

Implemented-first works: 3/3 clean

The code-gap finding became method and paid for itself in one firing: the refutation battery (flagship library, updated yesterday), the E-value confounding-sensitivity measure, and J-PCMCI+ completing the causal-discovery quartet. Zero casualties; STABLE at 41 of 100.

3/3 CLEAN · 103/300

The sensitivity pair completes; firecrawl unreachable

Quantified hidden-confounder skepticism now covered in both regression and epidemiology units plus the empirical probe; the implication enumerator turns one grand claim into many falsifiable ones. The crawler failed its first live test — five parked instruments wait, honestly recorded.

2/2 CLEAN · 105/300 · CRAWLER DOWN

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

The depth cohort passed intact; the refutation battery proved anchor-native by design and the implication enumerator now plugs into the in-house FDR layer — instruments interlocking. Three passers short of the third STABLE handoff.

7/7 PASS · B-07 THREE AWAY

Final harvest + operator stop: the loop parks at 108/300

Three more STABLE instruments (implemented-first closes 8/8), then the stop directive — recorded verbatim, schedule cancelled, firing shipped whole. Six batches and 68 instruments delivered across every boundary; the resume queue is preserved on the ledger.

3/3 CLEAN · STOPPED CLEANLY

Machine-readable SSoT (audit twin): findings/evolution/audits/2026-07-04-matrix-harvest-status-grounding-loop/ — CLAUDE.md · H0-HARVEST-PRE-REGISTRATION.md (frozen at PR #581 landing) · FAMILY-TAXONOMY.md · SOTA-PROVENANCE.md · CANDIDATE-CATALOG.md · FRONTIER-FEED.md · LEDGER.md · LOOP-PROMPT.md · verdict.md

Campaign: 2026-07-04-matrix-harvest-status-grounding-loop · paper-only · laptop-hosted · append-only · this page is the human-readable twin; the audit folder is the machine-readable twin