This is infrastructure-only: the Arbiter's five seats move from the unmetered `callVendor` to the pipeline's metered `callSeat`, a `costOf` aggregator is added with tests, the arbiter record schema gains optional `cost` and `paradigm` fields, and harvest copies the cost through. No content/ files, no citations (verification tool found none), no AGENTS.md edit. The checks in §3.15 are untouched: five distinct vendor seats remain, a failing seat still counts as a FAILED 'unsure' rather than being silently swapped, and the vendor-independence comment is preserved and extended. The cost reporting is honestly conservative — tokens summed as a floor, dollars withheld and the bill marked INCOMPLETE when any seat returned no accounting — which is the opposite of false precision (§3.13, §3.8). The workflow's switch from MAINTENANCE_PAT to GITHUB_TOKEN with `issues: write` is scoped to opening/labeling review-note issues and does not loosen who parks or merges a change. The one thing I cannot verify is the DECISIONS.md figure '$0.16: 46,000 tokens in and 3,200 out across five seats' from a prior metered run; it is plausible as a trial-run report and is stamped as an operator observation, so I would flag it only as something the operator should be able to point to a run record for, not as a violation.

The change strengthens rather than weakens §3.15 accountability by routing every Arbiter seat through the existing metered transport and carrying the resulting cost into the harvested governance record. `costOf` withholds a dollar total whenever any seat lacks accounting or any returned usage is unpriced, while explicitly marking incomplete token totals as a floor, which is honest and calibrated. It also retains the cost of a returned but unparsable response, avoiding silent deletion of incurred usage, and the added tests cover all of these cases. No AGENTS.md edit, fabricated citation, confidence material, unratified content judgment, or bypass of multi-model review is present.

The change routes Arbiter seat calls through the pipeline's metered transport (`callSeat`) and records token usage and monetary cost into governance records and schema. It accurately tracks provenance and cost without weakening the multi-model arbitration checks in §3.15 or modifying AGENTS.md. Proper fail-safes are maintained if a seat call or JSON parsing fails, ensuring failed seats remain marked as failed while retaining actual incurred usage.

The diff meters Arbiter seats through the existing callSeat path, drops the sole unmetered callVendor caller, and records panel cost (tokens always; dollars only when every seat is priced, incomplete otherwise) in the verdict blob and harvest schema—honest accounting, not fabrication or check-weakening. Schema, harvest, tests, and docs match that behavior; AGENTS.md is untouched. The workflow token swap only fixes review-note labeling. No content citations, confidence material, or provenance misstamps appear.

This is a plumbing change that routes the Arbiter's five seats through the same metered transport as pipeline verbs, removing the unmetered `callVendor`, and records the panel's cost in the verdict blob and harvested governance record — it makes the check-and-ledger machinery more transparent, not weaker. Nothing under content/ is added, no citations are introduced (the verification section is empty), and AGENTS.md is untouched, so §3.15's amendment rule is not implicated. The cost accounting follows §3.8's spirit honestly: tokens are always summed, dollars are withheld (usd: null, 'not fully priced', 'INCOMPLETE') whenever any seat returned no accounting, and the tests in arbiter.test.ts exercise exactly the incomplete, unpriced, and unparsable-seat cases. The failed-seat path still marks the seat failed rather than silently swapping vendors, preserving §3.15's vendor-independence requirement. The switch from MAINTENANCE_PAT to the workflow's own GITHUB_TOKEN for review-note issues comes with a concrete stated reason (the fine-grained PAT could not add labels) and does not loosen any §3.15 check — the lone-objection → issue flow is preserved and actually gains the labeling it was missing. The DECISIONS.md entry records the founder direction with date and channel as the constitution permits; I have no positive evidence of misattribution in the material shown.