This is a pipeline and CI change. It does not touch AGENTS.md or content/, and it adds no citations, records or assessments, so §3.2–§3.13 do not come into play. The key question is §3.15 check-weakening. `ratification()` is untouched, and the new `checksStale` adopts the same four-current-seat threshold (RATIFICATION_MIN_PANEL). Its reconsideration clause is now stricter, since it filters to current checks. The only relaxation is scheduling priority when four seats are current and a fifth is stale. The PR's own account of 2026-09-28 is consistent with the claim that standing was already derived from the current seats in that situation: four seats installed and the standing came out contested. Asking only `seatsOwed` means a seat's answer on a ledger is no longer re-rolled by default, which if anything protects independence. `--seats` overrides are recorded in the run reason via `heldNote`. Streaming changes only transport: the test shows thinking is excluded from the reply, and a stream cut before message_stop fails the seat rather than producing half an opinion. The workflow's token pre-check preserves the 'money spent is recorded' guarantee, because the PR step still runs whenever the chain could have run. The failure notice is a public issue, so it is not an unrecorded contact. The DECISIONS entry records the founder's direction with date and channel, as §3.15's amendment permits, and honestly discloses the unrecorded spend. A minor, non-constitutional note: MAINTENANCE.md's 'the next sitting's check re-panels it' line may now overstate promptness for a reconsideration that has one fresh check, since those re-checks wait for rule 5b.

The change does not alter the four-seat ratification threshold or standing derivation; it changes scheduling so completed judgments on the same ledger are retained while only missing seats are asked, preserving independence without silently multiplying one seat’s evidence. Reconsiderations still exclude checks they answered, and fewer than four current seats remain stale, so the publication gate is not weakened. The Anthropic stream assembler rejects incomplete responses before `message_stop`, and failed seats are recorded rather than converted into opinions. The decisions entry identifies the run, dates, uncertainty about unrecorded cost, and founder direction without presenting the unavailable amount as known; no content citations, confidential material, AGENTS.md edit, or fabricated locator is added.

The change addresses operational pipeline stalls without altering the epistemic ratification thresholds established under §3.15. Specifically, `checksStale` now aligns with ratification rules by recognizing that four concurring seats constitute a valid panel, while owed seats are tracked and scheduled independently. Refactoring the Anthropic panel seat transport to stream responses resolves Node timeout drops without changing what the seat evaluates. All workflow error reporting and secret pre-checks improve inspectability and operational resilience while leaving AGENTS.md untouched and keeping change histories properly recorded.

The diff is automation and domain scheduling only: Anthropic seat streaming, aligning checksStale/seatsOwed with existing ratification’s four-current-seat rule, a same-step repeat stop, token preflight, and failed-sitting issues. AGENTS.md is untouched; no content/ citations, confidence material, or run overlays are added. Standing derivation is explicitly unchanged—scheduler and check verb only decide whom to ask and when, so a short panel cannot raise a standing and seats already current are not re-paid. That tightens the loop against the 2026-09-28 stall rather than loosening §3.15 gates. DECISIONS/MAINTENANCE/AUTOMATION updates record the founder’s 2026-09-29 direction with ordinary provenance.

This is an operations/infrastructure change, not a content change: it does not touch AGENTS.md, adds no citations, and carries no run records, so the fabrication, confidence, and constitution triggers do not apply. The central question is whether the new staleness rule weakens the §3.15 checks, and it does not: ratification itself is untouched and still requires four seats to have judged the current ledger (the tests confirm four seats ratify, three do not), the check verb now asks only seats that have not judged the ledger as it stands, and a failed or missing seat is pursued on a doubling cadence rather than by re-paying four seats to answer the same question — which the diff shows happened three times on 2026-09-28 with one seat's verdict flipping between steps. The change to not retry a request that met Node's five-minute header window reduces duplicate spend, not scrutiny, and the failure is thrown and recorded rather than swallowed. Provenance is honest throughout: the decisions entry is stamped with the founder's session direction and date, the un-recoverable cost of the nine unanswered requests is explicitly declared as unrecorded rather than invented, and the failed-sitting notice discloses the run URL, steps, and mend. Nothing in the material contradicts its own stated history, and the verification tooling found no citation identifiers to check.