This is operator-tooling work: `scripts/arbiter.mjs` re-asks a seat whose reply could not be parsed as JSON, with the same packet, same seat identity, and no fallback model. It does not loosen any §3.15 check — quorum, the graded concurrence scale, the lone-seat parking kinds, and the rule that a refusal fails the seat are all untouched; only a transport-level parse failure is retried, and a second unreadable reply still fails the seat with both replies' cost attached. Provenance is strengthened rather than hidden: the new optional `askedAgain` field on `ArbiterSeatSchema` records that the verdict is the second asking and the panel report displays that note beside the seat's vote, so no reader is misled about how the verdict was obtained. AGENTS.md is not edited; the DECISIONS.md entry is a dated, channel-stamped record of a founder direction in session, which §3.15 expressly permits as the operating record. No content/ files, citations, or assessments are involved, and the citation tooling found no identifiers to check. Nothing in the diff fabricates a locator or publishes confidential material.

The change preserves strict vote parsing and permits only one retry before the seat fails, so it does not accept malformed output or reduce the required multi-model concurrence. The retry keeps the same model identity and packet, while the resulting record discloses that it was a second asking and aggregates both successful replies’ usage and cost. The decision log and schema make the procedural change inspectable, and the diff neither edits AGENTS.md nor alters citation, confidence-material, or content-verification safeguards.

The change adds a single retry when a panel seat's response fails JSON parsing due to syntax or transport formatting errors, preserving the exact model, packet, and failure semantics if the second attempt also fails. Full accounting and provenance are maintained by combining token and dollar costs, recording an `askedAgain` notice in the schema, and surfacing the re-query in the panel markdown report. The amendment is properly recorded in `docs/DECISIONS.md` as a founder direction given in session per §3.15, without touching `AGENTS.md` or weakening constitutional checks.

The diff only touches the arbiter retry path, an optional askedAgain stamp on the seat schema, and a dated DECISIONS.md entry. It does not edit AGENTS.md, add content citations, or publish confidence material. A single re-ask on unparseable JSON keeps the same seat, packet, and refusal-fails rule, bills both calls, and surfaces the second asking in the report—so it does not weaken §3.15 checks or dodge concurrence; it only recovers a transport accident. Founder supervision is recorded in the decisions log as required for operating-record directions.

This is a transport-robustness fix to the arbiter, not a loosening of any §3.15 check: the seat is re-asked with the identical packet and system prompt, its identity and model are unchanged (no fallback model), a refusal still fails the seat, and a second unreadable reply still fails it — so no invalid or unratified judgment gains a path to publication that it lacked before. The change preserves the constitution's transparency requirements: the retry is billed (both replies summed on the seat's cost), recorded via the optional `askedAgain` schema field, and displayed in the panel report next to the seat's own word, so ratification records stay honest about how a verdict was produced (§3.14/§3.15 accountability). It does not edit AGENTS.md, fabricate any citation, or touch confidence material; the founder direction is recorded in docs/DECISIONS.md with date and channel as §3.15 requires. The retelling in DECISIONS.md matches what the diff implements, and no content or run records are affected. One residual risk — a seat deliberately emitting malformed JSON to buy a retry — is mitigated because the retry carries no new information beyond a neutral parse-error notice, and the second asking is permanently marked on the record for readers to weigh.