The diff touches only config/models.yaml, config/tariffs.yaml, an appended docs/DECISIONS.md entry and one new test. It makes no AGENTS.md edit and adds no content/ records, citations or run records, so the fabrication and confidence rules are not engaged. No §3.15 check is loosened: the existing roster and tariff tests stay, the new claude-opus-5-5 tariff row is internally consistent ($4 base, $0.20 = 0.05×, $5 = 1.25×) and keeps the unpriced-model test satisfied, Fable 5.1's row is kept so past stamped runs stay priced, and the added test is a new check rather than a removed one. The founder direction is recorded in the decisions log with date and channel ('2026-09-22, in session', plus a Supervised-by line), as §3.15 permits. Roster choice is not a founder-reserved act, so compliance does not hinge on that claim, which I cannot verify. The AA scores and prices postdate my knowledge but are internally coherent ($326/$398 ≈ 82%; $4/$20 against $10/$50 is 60% less). The panel still seats five vendors, and the Anthropic seat moves into the stated 56–59 band rather than out of it. Notes for the operator, not violations: (1) the drafter and the Anthropic seat are now the same model, so that seat reviews its own model's drafts; record this and watch for concurrence drift. (2) The new test walks only src/**/*.ts and *.mjs, missing .tsx files and the scripts/*.mjs the house model drives, and it truncates lines at any '//' such as a URL, so the DECISIONS entry overstates what it enforces. (3) The tariffs.yaml section header still calls Opus 5 'the panel's house seat'. (4) Because the 'opus' tag persists while its label changes, past verdicts should display their stamped model id rather than the new label.

The change updates centralized model and tariff configuration, records the dated founder direction and channel in docs/DECISIONS.md, and does not edit AGENTS.md or publish content-run assessments. The tariff arithmetic shown is internally consistent, includes an official pricing-source URL and check date, and the packet contains no positive evidence of a fabricated citation or provenance detail. Retaining the panel tag does not conceal the model change because the configured model and public label are both updated explicitly. The added roster test strengthens rather than weakens the existing checks by detecting model identifiers embedded in source code, while the existing priced-model check remains in place.

The change updates configuration files and tariffs to switch the house and Anthropic panel seat models to claude-opus-5-5 pursuant to recorded founder direction. The founder direction is properly stamped in docs/DECISIONS.md with date and channel as required by §3.15. The change does not alter AGENTS.md, weaken any verification or concurrence gates, or introduce unverified or fabricated content.

The diff only retargets house/research/panel Anthropic models in config/models.yaml, adds a priced tariff row for claude-opus-5-5, records the change under a dated founder-direction entry in docs/DECISIONS.md, and adds a roster test that model IDs live in config rather than source. AGENTS.md is untouched (§3.15). No content/, citations, assessments, or provenance stamps are added or altered, so fabrication/confidence/provenance rules are not in play. The tag is kept and Fable’s tariff row is retained so past runs stay priced; that strengthens rather than weakens the checks. Founder direction is stamped with date and channel as the constitution allows for operating record.

This is a configuration change, not a content or constitutional edit: AGENTS.md is untouched, no content/ files change, and no citations are added, so the fabrication and confidence rules are not engaged. Moving the Anthropic panel seat from Opus 5 (AA 51, below its own 56–59 band) to Opus 5.5 at max effort strengthens rather than weakens the §3.15 ratification machinery — the seat stays a distinct-vendor independent seat, keeps its tag so a case's checks remain one series, and the diff adds a new test enforcing that model ids live only in config, which tightens rather than dodges the checks. The founder direction is recorded in docs/DECISIONS.md with date and channel ('direction in session, 2026-09-22') as the operating record requires, and Fable 5.1's tariff row is retained so historical run records stay priced, honoring append-only provenance. The benchmark scores, tariffs, and the claimed founder direction are facts I cannot independently verify from the packet, but temporal humility bars treating unfamiliarity as fabrication, and the packet's tooling found no citation identifiers to check. The test's single documented exclusion (src/lib/models.mjs, the loader's own schema) is a reasonable carve-out, not a weakening; at most it is a review-note-level nit, not a violation.