This is pipeline and infrastructure work. It adds an optional `maxOutputTokens` to the panel schema and sets it to 128,000 for the Anthropic seat. Empty-reply errors now carry the vendor's usage, so `callSeat` records the spend row before rethrowing. It also updates tests and adds DECISIONS/MAINTENANCE entries, and it touches no content/, no citations, and not AGENTS.md. No §3.15 check is loosened: the `next.test.ts` edit only narrows which real cases the scheduler-ordering tests pick, excluding those with ≥LEADS_MIN blocked leads because rule 3c runs first by design; the scheduler itself is unchanged, so this repairs a brittle fixture rather than dodging CI. The ledger is handled honestly: the earlier failed run (…052151) stays at $0, and the ~$1.47 figure is explicitly labeled an inference from the tariff rather than backfilled, which fits no fabricated records and no silent data repair (§8). The figures are internally consistent: $2.33 for 47,864 in / 106,953 out and $1.47 for 47,864 / 64,000 both fit one ~$4/$20-per-million tariff, and 2026-09-28 is indeed a Monday. The founder's pinned effort (2026-09-22) is expressly left unchanged, and Deep Memory's resulting standing is reported as contested, with its records in the sitting's own PR, not presented as ratified. The 'Supervised-by: founder (direction in session, 2026-09-29)' stamp on a 2026-09-30 entry is plausible, since 05:21–05:37 UTC is still the evening of 09-29 in US time zones, and I see no positive evidence of misattribution. One minor nit, not a violation: the vendors.mjs comment says 'about $1.40' where the decisions entry says 'about $1.47', and the two should be aligned.

The decisions entry properly labels the failed call’s approximately $1.47 cost as an inference, but it then generalizes one Deep Memory measurement into the categorical claim that at max the seat costs five to seven times the other seats and imposes “a fifteen-minute wait on every one.” The same entry reports a materially different Arbiter ratio—$0.77 versus $0.24 for the remainder—showing that cost varies by workload, even if an Arbiter vote is not identical to a case check. Presenting the single check’s ratio and duration as recurring measured behavior fails to distinguish observation from extrapolation and omits calibrated uncertainty. Rewrite this as specific to the cited Deep Memory run and state that future cost and latency may vary.

The pull request adjusts the Anthropic panel seat configuration to allow a higher token ceiling (128,000 output tokens) so that extended thinking does not exhaust the token budget before producing an answer. It also updates transport error handling to ensure metered token usage from empty/exhausted runs is recorded in the spend ledger rather than silently lost. No files in content/ or AGENTS.md are modified, no epistemic checks are weakened, and comprehensive unit tests are added.

The diff is operational only: roster ceiling for the Anthropic seat, spend recording on empty billed replies, test/schema/docs updates. It does not edit AGENTS.md, add content citations, fabricate provenance, publish confidence material, or weaken §3.15 checks. DECISIONS/MAINTENANCE entries document measured cost and founder direction without presenting unratified judgments as ratified. No omitted files or run account to hide a violation.

This is pipeline infrastructure with no content, citations, or run records, and it does not edit AGENTS.md. The core change does not weaken any §3.15 check — it repairs one: the Anthropic seat previously could not render any verdict at all (all 64,000 tokens consumed by thinking, empty reply), so giving it the roster-named ceiling of 128,000 enables the seat to participate in ratification rather than excusing it; the pinned max effort is explicitly left as the founder's 2026-09-22 decision. Recording the vendor's usage from a failed, billed call is more honest ledgering (§3.14), and the decisions entry is careful to label the ~$1.47 as an inference from the ceiling hit rather than a vendor report, refusing to write it into the ledger after the fact — exactly the provenance honesty the constitution demands. The scheduler-test change narrows 'unencumbered' to exclude cases with leads, matching the existing pattern for provisional/blocked records and preserving test determinism without altering any check's behavior. The MAINTENANCE.md note correctly keeps a deeper effort change as the founder's reserved act. I see no fabrication, no confidence exposure, no check-weakening.