Probe arbitration

Why it carries this status

Lifecycle, not result. This says where the audit sits in its process — never whether what it found was good.

* **Re-running R2/R3.** Compute, and gated on the operator.

2026-08-17. Derived from folder evidence, then adversarially challenged; the challenge pass was upheld.

The verdict

The audit's own conclusion, reproduced in full from the source below. Not a summary — this is the document, rendered. Links inside it that point at unpublished files are shown as plain text rather than as links that would 404 here.

                 panel hygiene        threshold calibration
  ROTATIONAL     ✅ correct           ❌ PASS unreachable
  LEGACY         ❌ broken            ✅ can certify

Still owed 17

• R3 (scoring the 20 real candidates under the repaired dials, both probes) is blocked — it makes the suspended xi and untested sibling stages live, and running it on duplicated input would build on the defect.
• Re-running R2 and R3 on the deduplicated substrate is compute work explicitly gated on the operator.
• R4 — the probe-verdict × realness-verdict cross-tab — requires the separate realness/usefulness arbitration to run first.
• R5 — writing axis.orthogonal for all 20 candidates in one pass — is gated on R2 passing and R4 being available; the registry stays frozen at 88 null · 12 PASS · 5 BAN · 4 WATCH · 1 PENDING.
• The xi leg is suspended from label authority until re-derived on deduplicated data.
• The sibling leave-one-out R² leg is suspended from verdict authority pending its own calibration declaration; the sibling stage has no declaration document of any kind.
• P_50 = 0.925944 is recorded as substrate-contaminated and deliberately not re-derived; recomputing it on clean data has zero degrees of freedom and needs no new declaration.
• three_axis_gate.py cannot express the repaired rules and is declared not authoritative; repairing its schema is operator-gated.
• The ±0.05 run-to-run drift attributed to D6 was not reproduced post-fix and should be re-measured; if it persists it has another cause.
• Whether the out-of-repo writer that duplicated the rows is still scheduled was never checked (last write 2026-07-23; no cron or systemd unit inspected).
• A second, genuinely undedupable duplicate class documented in Terry's warehouse_overlap_audit.py is unquantified.
• Whether Chatterjee's xi is adequately calibrated under the genuine 30.93% point mass that survives deduplication on LINKUSDT@100 is open; the literature review was blocked by an environment egress failure.
• The table-side remedy min_age_to_force_merge_seconds on open_deviation_bars is a production ALTER and is the operator's call.
• close_time_us is absent from the table sort key, which is both a performance cost (full-partition scans) and a correctness issue; fixing it needs its own declaration plus an equivalence proof.
• Residual envelope risks recorded but unfixed: a query carrying its own SETTINGS could override the shim, and anything reaching the database over HTTP or a Python driver bypasses the shim entirely.
• The SOTA research flags items 950, 951, 954, 962 and 992 in its section 9 as load-bearing and unverified before anything is implemented; in particular the fixed-b refutation rests on a second-hand sentence about the sample mean, and xi is not a sample mean.
• Whether orthogonality predicts tradeability at all is unsettled; if realness verdicts prove uncorrelated with both probes, orthogonality is pre-registered to be demoted from a promotion gate to a hygiene tag.

What it claims, and what backs each claim 33

Every row pairs a claim with the file it came from and the verbatim text in that file. The sources sit above the deploy root, so the quote is embedded and the path is printed as text rather than linked — a link would resolve on a laptop and 404 here.

The audit folder 10 markdown files

Source of record: findings/evolution/audits/2026-07-24-probe-arbitration/ — not published, so these are listed rather than linked.