Iteration 2 — PMLR leads recovered + the ε-frontier anchor

Iteration 2 — PMLR leads recovered + the ε-frontier anchor HARVESTED 5 · FAMILY 1 NEAR-SWEPT

2026-07-04 · second firing · ledger row 4 · campaign branch loop/matrix-harvest-campaign

In plain words: last firing, three good candidates were left standing at the door because the arXiv channel was rate-limited and their conference venue isn't indexed by our backup channel. This firing the arXiv channel came back, all three walked in with verified papers — plus two bonus finds: an instrument-validity test from the same research school, and a 2025 theory paper proving that perfectly testing invariance is computationally intractable, which comes with a tractable work-around. That last one matters: it is the mathematical justification for this project's entire "ε-approximately-stable" framing — you cannot demand perfection, so you measure controlled imperfection.

What entered the catalog (5 rows, all STABLE boundary)

id	Instrument	One-line story	Primary source (arXiv API-verified)	FOSS (checked)
H-022	Active ICP	Picks the most informative experiments for invariance testing — catalogued with an honesty flag: our market data offers no experiments, only observation	arXiv:2006.05690 (2020)	`juangamella/aicp` 21★
H-023	Invariant Ancestry Search	Hunts a weaker but far more attainable target than full causal identification — built for low-power settings like ours	arXiv:2202.00913 (2022)	`PhillipMogensen/InvariantAncestrySearch`
H-024	EIIL — environment inference	Infers regime partitions from data — a stress-test for whether OUR regime slices are the right environments	arXiv:2010.07249 (2020)	`ecreager/eiil` 52★
H-025	HSIC-X — independent instruments	Tests whether an "anchor" variable is actually a valid instrument, via independence — first family-20 entry	arXiv:2202.01864 (2022)	`sorawitj/HSIC-X`
H-026	Computational limits of invariance pursuit	Proves exact invariance testing is NP-hard AND supplies the tractable relaxation — the ε-frontier's mathematical anchor	arXiv:2501.17354 (2025)	theory instrument (n/a); nearest artifacts `wmyw96/EILLS` (MIT) · `wmyw96/FAIR`

Instrument

One-line story

Primary source (arXiv API-verified)

FOSS (checked)

H-022

Active ICP

Picks the most informative experiments for invariance testing — catalogued with an honesty flag: our market data offers no experiments, only observation

arXiv:2006.05690 (2020)

juangamella/aicp 21★

H-023

Invariant Ancestry Search

Hunts a weaker but far more attainable target than full causal identification — built for low-power settings like ours

arXiv:2202.00913 (2022)

PhillipMogensen/InvariantAncestrySearch

H-024

EIIL — environment inference

Infers regime partitions from data — a stress-test for whether OUR regime slices are the right environments

arXiv:2010.07249 (2020)

ecreager/eiil 52★

H-025

HSIC-X — independent instruments

Tests whether an "anchor" variable is actually a valid instrument, via independence — first family-20 entry

arXiv:2202.01864 (2022)

sorawitj/HSIC-X

H-026

Computational limits of invariance pursuit

Proves exact invariance testing is NP-hard AND supplies the tractable relaxation — the ε-frontier's mathematical anchor

arXiv:2501.17354 (2025)

theory instrument (n/a); nearest artifacts wmyw96/EILLS (MIT) · wmyw96/FAIR

Recount correction (the honest part): recomputing the counters from per-row boundary tags exposed an off-by-one inherited from the scaffold: seed entry H-003 (SAVI e-processes) is tagged EXPIRY in its own row, so the seed split was STABLE 8 / EXPIRY 5 — not the 9 / 4 the scaffold counter (and iteration 1, which trusted it) reported. Totals were always right. Per-row tags are the source of truth; the correction is recorded in ledger row 4 with a supersedes pointer, nothing was edited in place.

Coverage counters after this firing (recomputed from per-row tags)

Boundary	Minimum	Verified	Movement vs iter 1
STABLE	100	19	+5 (and −1 recount)
FRAGILE↔CONDITIONAL	60	0	—
CONDITIONAL-scoping	40	1	—
FAIL	40	0	—
EXPIRY	40	5	+1 (recount)
cross-cutting	20	1	—
Total	300	26	+5

Boundary

Minimum

Verified

Movement vs iter 1

STABLE

100

+5 (and −1 recount)

FRAGILE↔CONDITIONAL

—

CONDITIONAL-scoping

—

FAIL

—

EXPIRY

+1 (recount)

cross-cutting

—

Total

300

Channel report

arXiv API ✓ RECOVERED (iter-1's 429 block cleared) — 5/5 primary sources verified by id ✗ 2 searches for the H-015 "Wasserstein-relaxed invariance" source: nothing usable gh search/api ✓ FOSS checks (HSIC-X, FAIR, EILLS, aicp, IAS, eiil)

H-015 stays provenance-pending. If the next family-1 pass also fails to resolve a Wasserstein-specific source, the lead is likely a mislabel of the ε-frontier line now anchored by H-026 and will be superseded toward it — recorded, not deleted.

▶ Next iteration

Family 1 is near-swept (remaining: GROW/e-value pipeline variants + the H-015 resolution attempt). Two candidate slices, in the frozen priority order:

Option A (default) — first M0/M1 screen batch: STABLE now holds 17 screenable CATALOGUED entries (19 minus the 2 already screened-out seeds). Screening ≥10 of them M0/M1 could trigger the first FRONTIER-FEED batch to the admission loop — the campaign's first deliverable. The pick rule says a pending screen/feed threshold outranks new harvesting once available.
Option B — keep harvesting: finish family 1's e-value tail, then family 2 (anytime-valid monitors, serves EXPIRY+STABLE) or the predecessor-import batch (~95 instruments, mostly FAIL/cross-cutting).

Steering point: no directive = Option A next firing. Type a directive into the loop session to override — it is ledger-recorded verbatim before taking effect.

Machine-readable record: ledger row 4 + catalog rows H-022–H-026 in findings/evolution/audits/2026-07-04-matrix-harvest-status-grounding-loop/. Pages are append-only.