Final status — 14 iterations CAP REACHED · LOOP STOPPED

2026-07-30 · 14 iterations · 14 committed harnesses · 2 pre-registration drafts · 16 commits on PR #676

The headline, stated plainly

Neither probe can certify a candidate yet, and that was the expected outcome. The loop was authorised to land zero-parameter correctness fixes and to measure everything else. Every remaining blocker is either a dial that needs your pre-registration, or a re-baseline that needs a full-width bigblack read. Both were out of scope by design, not by failure.

Operational status, per capability, each verdict linked to what established it

Capability	Rotation	Legacy
Reads one row per bar	yes #674	yes #674
Deterministic fetch (total ORDER BY)	yes #674	yes #674
Candidate values uncorrupted	yes #674	yes #674
Estimator pinned to a canonical oracle	yes exact to 2.44e-15 vs CRAN XICOR — iter 1	n/a legacy has no ξ
Gate reproducible across runs	yes seeded, bit-identical — iter 3	n/a
Row-order invariant	yes exact 0.00e+00 — iter 14	untested only ξ was checked
Reports nothing on shuffled data	yes 60/60 — iter 7	yes 400/400 — iter 7
Bands sit above their noise floor	yes 15× headroom — iter 7	yes 7.7× — iter 7
Estimator stable near a decision line	band safe, margins not jitter 0.73× the moat but 5.4× the 0.0008 margin — iter 2; de-randomisation available, not wired — iter 3	unknown ρ-under-ties never tested (L-4 unreached)
Comparison panel legitimate	blind spot passes a pure price-level function — iter 5	over-credits levels inflate ρ +0.42; trade IDs +0.16 more — iter 5, iter 6
Labels free of lookahead	no vol terciles relabel 2–5% of past rows per re-run — iter 8	n/a no regime stratification
BAN verdict is data-driven	only above 4,720 bar-equiv vs a declared floor of 1,000 — iter 10	untested
Verdict statistically calibrated	no unanimity gives 72% false veto below the line — iter 4; worst-of-N drifts +0.005–0.010 per 4× cells — iter 9; permutation calibrates the wrong null — iter 13	no same max-vs-fixed-line defect, unmeasured (L-3 unreached)
Can certify a candidate	not yet	not defensibly

Attack list — 14 of 17 addressed

ID	Status	Iter	ID	Status	Iter
R-3 oracle parity	FIXED	1	R-4 attainable ceiling	GROUNDED	10
R-1 tie-jitter scaling	GROUNDED	2	R-5 sub-slice nulls	GROUNDED	11
R-2 de-randomisation	FIXED	3	R-8 τ̂²ₙ calibration	GROUNDED	12
R-6 unanimity false veto	QUANTIFIED	4	R-7 permutation reference	QUANTIFIED	13
L-1 price-level panel	GROUNDED	5	X-3 order invariance	FIXED	14
L-2 trade-ID panel	GROUNDED	6	L-3 legacy worst-of-N	OPEN	—
X-1 shuffled null	FIXED	7	L-4 ρ under ties	OPEN	—
X-2 future perturbation	QUANTIFIED	8	L-5 h_norm point masses	OPEN	—
R-9 worst-of-N displacement	QUANTIFIED	9	3 legacy items not reached before the cap

Five zero-parameter fixes landed

ξ oracle gate — the binding stage was the only unpinned statistic in a repo where every other kernel is oracle-gated. Now exact to 2.44e-15 against Chatterjee's own package.
De-randomised estimator with self-reported MC-SE, short-circuiting free for continuous kernels.
Seeded gate — the iter-1 gate was itself flaking (p moved 0.0060 → 0.0030 against a 0.00167 threshold).
X-1 shuffled-null gate, both probes, negative and positive controls.
X-3 order-invariance gate, both directions.

All wired as /xi:oracle, /xi:structural-gates, /xi:order-invariance, /xi:doctor, /xi:check-full — suite green.

Where the loop corrected itself or the record

Two campaign framings refuted. L-1 ran the opposite way (A1 is too permissive, not legacy too harsh); L-2's "bug" was a declared treatment, self-documented in the harness.
The contested item resolved — standardisation works, but the constant is not universal (iter 11), and the estimator that fixes it is verified (iter 12).
One of my own claims corrected — iter 2's ceiling argument was overstated for typical cells (iter 10 measured the gap at 0.0023, 1% of the floor-scale gap).
Two of my own harnesses caught producing bad output — a false positive from a min/max support test (iter 1), and a pass criterion loose enough to wave through a 2σ size excursion (iter 13).
Predictions were wrong in 5 of 14 iterations and the misses are recorded as findings, not smoothed over.

Queued for the operator — nothing here was acted on

#	Needs	Why it is not the loop's call
1	Re-run R2 on the deduplicated substrate, 62 cells	full-width read; 2 GiB measured insufficient at any LIMIT
2	Re-derive P₅₀ from the 66 shipped columns	mechanical, but needs the real table
3	Confirm the jitter finding at real per-cell n on real cells	needs the kernel registry + a wide read
4	ξ stability guard — 4 remedy options drafted	every option is a dial; a flat `k` provably cannot work
5	Volatility tercile lookahead — 4 remedy options drafted	3 of 4 invalidate all accumulated coverage
6	Wire de-randomisation into the cascade	moves every verdict; R=37 resolves the margin, ~1.7 s per 62 cells
7	Retrieve Lahiri (2001)	the only route to closing the serial-dependence half of §(b)
8	L-3 / L-4 / L-5	not reached before the cap

The two drafts awaiting sign-off

PRE-REGISTRATION-DRAFT-xi-stability-guard.md and PRE-REGISTRATION-DRAFT-regime-tercile-lookahead.md — each names its free parameters and states explicitly what must not be inferred from the measurement behind it.

One thing that must not be lost

The guard was never deleted. Its strongest deletion argument is second-hand and about the sample mean, and ξ is not a sample mean (research gap #962). Iteration 11 showed the rule is repairable, which makes deletion the wrong move anyway — but the citation is still unretrieved, and the serial-dependence half of the objection stands open.

Every number on every iteration page is reproducible from a committed harness under findings/evolution/audits/2026-07-29-probe-hardening-loop/harness/.