Iteration 19 — Sweep Un-Parked via Firecrawl: Two Refill Candidates for the STABLE Gap

Iteration 19 — Sweep Un-Parked: Two Refill Candidates for the STABLE Gap SWEEP EXECUTED

In plain words: the literature sweep finally ran — not through the blocked search provider, but through the project's own self-hosted research scraper (reached by its Tailscale address). Three searches with verifiable sources produced two serious candidates for the campaign's one open instrument gap ("can anything identify true regime-invariance?"): (A) a family of e-value invariance tests — mathematically immune to the exact perfect-fit breakdown that killed the old instrument, with always-valid error guarantees; (B) invariance-guided regularization — the constructive workaround from the same research line that proved exact invariant-set pursuit computationally hard. Both enter the queue with a mandatory entrance exam already waiting: the perfect-duplicate ground-truth test from iteration 12, which any candidate must pass before its verdicts mean anything.

Preflight (A0): 2026-07-03 22:18 UTC — load1 4.21/32c · CH + sidecar active → ALL PASS. No ClickHouse; sweep via firecrawl (littleblack, Tailscale IP; MagicDNS non-resolving — noted); zero synthetic data.

Refill candidates (paper-screen survivors, default-EXCLUDE)

Candidate	Sources	Why it fits the gap	Entrance condition
A. E-value group-invariance testing	AOS 24-AOS2394 · arXiv 2310.01153 · arXiv 2606.00878	M0: STABLE + Expiry (both starved) · M1: Ville/anytime-valid bound · e-values stay well-defined under perfect fits — the exact degeneracy that killed causalicp (iter 12)	Must pass the iter-12 ground-truth duplicate calibration before ladder entry
B. Invariance-guided regularization	computational-limits-of-ICP paper	Tractable surrogate for invariant-set pursuit from the NP-hardness lineage already on file	M1 guarantees to extract from full text; same calibration gate
Conformal-under-shift findings (arXiv 2403.15025 · ICLR-25 Wasserstein-conformal · JMLR online-conformal) fold into the Frontier #4 reference pool.

Candidate

Sources

Why it fits the gap

Entrance condition

A. E-value group-invariance testing

AOS 24-AOS2394 · arXiv 2310.01153 · arXiv 2606.00878

M0: STABLE + Expiry (both starved) · M1: Ville/anytime-valid bound · e-values stay well-defined under perfect fits — the exact degeneracy that killed causalicp (iter 12)

Must pass the iter-12 ground-truth duplicate calibration before ladder entry

B. Invariance-guided regularization

computational-limits-of-ICP paper

Tractable surrogate for invariant-set pursuit from the NP-hardness lineage already on file

M1 guarantees to extract from full text; same calibration gate

Conformal-under-shift findings (arXiv 2403.15025 · ICLR-25 Wasserstein-conformal · JMLR online-conformal) fold into the Frontier #4 reference pool.

▶ Next iteration

Standing for the operator: the 73 grid-corroborated draft statuses and the Frontier #4 precondition review. Next firing: candidate A's full-text extraction (firecrawl scrape of the arXiv HTML) + the ground-truth duplicate calibration harness for e-value invariance testing — the entrance exam before any ladder walk. Expiry re-grounding triggers on the next completed regime slice.

Iteration 19 · 2026-07-03 · sweep executed via firecrawl (LEDGER row 25, FRONTIER Tier-1 refill) · append-only