OATH v0.11: an accusation is itself a claim, and four accusations had no receipts

[fathomlab]

Scored under PREREG_oath_v11_row_ordinal_retraction_2026_08_25.md, frozen on cbd2864 before styxx/certify.py was touched.

Outcome: V11_ORDINAL_RETRACTION_SHIPS. All nine gates pass. V11_ORDINAL_LABEL ships True.

What it does

Every cycle from v0.1 to v0.10 asked whether a token was GROUNDED. This one asks whether it is a CLAIM.

A markdown table's first column is where this corpus writes row numbers. Extraction takes them like any other token and the obligation predicate binds them, so a row number had to ground in a receipt or be accused. The certified frame's entire standing accusation surface was four of them, in one document. The VERIFIED half of the same column was worse: L26 was sworn to scale_test_result.json:per_item[2].i, an index leaf equal to its own subscript. Exhaustive substitution over the class answers 46 / 50 / 21 across 117 mutants — a 0.4274 false-attestation rate on tokens that assert nothing.

The clause is a status-level ABSTAIN with reason row_ordinal_label at the is_spec tier, gated on three conjuncts: first cell of a table DATA row, an exact header from a nine-entry vocabulary frozen in the prereg, and a cell that is entirely a bare non-negative integer ≤ 100. Value-blind, so doctoring the digit does not switch it off. Never non-extraction, so every silenced token stays countable by coordinate.

The retraction

This performs, once and by design, the event v0.9's G4 forbade: four UNGROUNDED→ABSTAIN conversions and one OATH-FAILED→OATH-HELD flip. The licence is that a status may be withdrawn only when what is shown false is the accusation's presupposition (claimhood), never its verdict.

• Genuine verifications destroyed: 0
• Conversions outside the target document: 0
• The committed certificate was never hand-edited; it reproduces
• KNOWN_VERDICT_DRIFT shrinks in this commit, which is the only way an entry may leave it

Process receipts

A second adjudicator, blind to the prior panel and to the clause's reason codes, re-adjudicated the full 11-token roster plus a fresh draw of 10: 11 LABEL / 10 CLAIM, zero LABELs outside the roster. Its verifier pin was corrected after the fact and the wobble is disclosed rather than smoothed.

The battery was then adversarially audited by four independent lenses, each finding put to two skeptics instructed to refute. 18 findings raised, 2 survived, both naming one defect: G4′b's bars passed vacuously on an empty list, so a truncated panel artifact would have cleared the gate by omission — this cycle's own thesis turned on its instrument. Repaired, along with a voiding G1 that crashed before writing the outcome it existed to produce. Re-derived at a reconstructed frozen tree: every gate verdict byte-identical, only diagnostics added.

Reviewer notes

• The battery VOIDs on G1 if re-run on this commit. The ship adds a certified document and moves the frame it pins — that is G1 working. Reconstruct the frozen tree first (hold the RESULT certificate aside, rebuild the ledger) to re-derive.
• The RESULT dogfoods itself and reports the finding against its own certificate: it is OATH-HELD, and nine of its verified bindings are sworn to positions rather than measurements — the v0.8 coincidence channel, CLOSED_NEGATIVE, in this cycle's own output and not fixed by it.

Suite green. ruff clean over styxx/. git diff over every committed *.certificate.json empty.

🤖 Generated with Claude Code