In plain words: the campaign shipped its first deliverable. The ten STABLE-boundary instruments that survived the paper interrogation are now packaged, ordered, and placed on the admission lab's intake shelf (FRONTIER-FEED.md, batch B-01). The ordering is deliberate โ cheapest kill first: instruments whose own papers supply a ready statistical guarantee and working code go to the front (the lab can falsify or accept them fastest); instruments whose trust anchors we had to construct ourselves go mid-pack (the lab's real-data gate is their true test); the two e-value entries ride last because the lab is already examining that family under its own candidate-A track. From here the ten are out of this campaign's hands โ the lab alone decides admit or exclude. Housekeeping on the record: the previous firing was interrupted mid-publication (the batch text existed but was never committed or logged) โ this firing completed it honestly, with the publication date corrected to today.
The handoff (one-way, one interface)
10 SCREENED-M0M1-PASS (STABLE) THE ADMISSION LAB (bigblack)
โ reads FRONTIER-FEED.md as a
โผ order: cheapest-kill-first Tier-2 refill, at its own pace
B-01 โโ 1 H-016 seqICP (CRAN, paper-native type-I control)
2 H-023 IAS (Python, paper-native FWER)
3 H-019 IPP (R, paper-native scoring-rule test)
4 H-025 HSIC-X (needs block-permutation null โ M2 decides)
5 H-017 StabilizedRegr (constructed anchor โ M2 decides)
6 H-021 EILLS (asymptotic bound; ~2K/regime unknown)
7 H-018 Causal Dantzig (no ready code; trivial closed-form + HAC)
8 H-004 PCMCI+ (heaviest compute of the batch)
9 H-001 GROW e-stats (flag: already admission candidate A)
10 H-002 Koning e-values (flag: candidate A family)
โ
โผ
states โ FED(B-01) โ terminal for this loop; verdicts belong to the lab
Operator directives recorded this firing (ledger row 7)
Relaunch at 20m cadence (SSoT says 45m) with a new preamble: work "brute force mode", be exhaustive, and conduct a web search for SOTA concepts/matrices + FOSS implementations. Applied: hosted WebSearch was re-probed rather than assumed blocked โ the probe was filter-rejected again (same class as the 2026-07-03/04 recordings), so the block is confirmed to persist as of 2026-07-06. The provenance ladder (arXiv API โ GitHub search โ Crossref โ PyPI โ firecrawl-on-littleblack) remains the operative channel set; the loop will re-probe periodically. "Brute force / exhaustive" is read as: sweep families exhaustively within the โค25-entries-per-firing bound and the frozen quantity-honesty rule โ never padding.
Coverage counters after this firing
Boundary
Minimum
Verified
Screened-pass
Fed
STABLE
100
19
10
10 (B-01)
FRAGILE↔CONDITIONAL
60
0
0
0
CONDITIONAL-scoping
40
1
0
0
FAIL
40
0
0
0
EXPIRY
40
5
0
0
cross-cutting
20
1
0
0
Total
300
26
10
10
โถ Next iteration
Iteration 5 = resume harvesting โ STABLE remains the most-starved boundary (deficit 81 vs FRAGILEโCOND 60, FAIL 40, COND-scoping 39, EXPIRY 35, cross-cutting 19). Family 1 is near-swept (only GROW/e-value pipeline variants remain), so the next slice is either the predecessor-import batch (re-verify + boundary-tag the still-valid entries of the 2026-05-16 ~95-instrument catalog โ imports count toward targets and several serve STABLE) or the next unswept STABLE-serving family from FAMILY-TAXONOMY.md. The import batch is the cheaper, higher-yield slice and is the default pick.
Merge point (operator): B-01 is published โ per the LOOP-PROMPT operator notes this is a merge-worthy milestone. Scaffold PR #581 and campaign PR #582 are both OPEN; merging them lets the admission loop consume B-01 from main. The loop never merges.
Steering point: no directive = predecessor-import batch next firing. Directives (pick a specific family, reorder priorities, hold harvesting) are ledger-recorded verbatim before taking effect.