This is a pipeline/engine change with no content/ records, no citations (verification found none), and no edit to AGENTS.md. The substance strengthens rather than weakens the §3.15 checks: `servedBy` reads the vendor's fallback block and per-iteration usage so a run record stamps the model that actually answered rather than the model that was asked, and the note ('asked → served (trigger; n fallback iterations)') is propagated into the run record's notes and the report header by report.ts, draft.ts and edition.ts. `usageByModel` bills each iteration to its own model's tariff, correcting a spend ledger that had been misattributed — again a provenance fidelity improvement. Unit tests are added for both helpers, consistent with §8/§10. The DECISIONS.md entry is a dated operator record of the fault and fix, stamped with date and channel, which §3.15 expressly permits for founder-session directions and operator records; it discloses rather than conceals the earlier false stamp. The PR body's claim that the affected run's own stamps and spend rows were hand-corrected elsewhere is not verifiable from this diff, but that is a follow-on PR's burden, not a violation here.

The change strengthens truthful provenance by deriving the served model from the vendor fallback block and per-iteration metadata rather than retaining the originally requested model. It propagates the fallback path into report headers and run reasons and allocates token usage to each identified model, with focused tests covering fallback, missing-block, ordinary, and per-model billing paths. The DECISIONS entry records the discovered historical stamping error and the implementation response without presenting a content assessment or unratified judgment as ratified. No citations or publishable case content are added, AGENTS.md is untouched, and no constitutional check is weakened or bypassed.

The change improves provenance accuracy in compliance with §3.8 and §3.15 by ensuring that when an upstream API executes a server-side fallback, the run records and billing reflect the model that actually generated the output rather than the model originally queried. The diff adds helper functions `servedBy` and `usageByModel` along with unit tests, integrates fallback attribution across pipeline verbs, and records the operational decision in DECISIONS.md without touching AGENTS.md or modifying any content files.

The diff fixes false model provenance when a vendor server-side fallback answers: servedBy reads the fallback block and per-iteration usage so run stamps name the model that actually wrote, with an explicit asked→served note, and usageByModel bills each iteration at that model’s tariff. That upholds §3.15’s requirement that every change be stamped with the model that served, rather than weakening checks. No AGENTS.md edit, no content/ citations, no confidence material, and no fabricated records; DECISIONS.md only appends an operator decision entry. Tests and pipeline call sites (report/draft/edition/models) consistently carry the corrected model and fallback note into closeRun.

This change strengthens §3.15's truthful-stamps requirement rather than weakening it: the run record previously named the model that was asked (claude-fable-5-1) while the fallback model (claude-opus-5) actually answered, and the fix makes servedBy read the fallback block and per-iteration usage so the stamp, notes, and report header name the model that actually served, with the fallback trigger and iteration count displayed. The billing fix (usageByModel charging each iteration to the model that consumed it at its own tariff) corrects a spend-ledger misattribution, which is also a provenance-accuracy improvement. Nothing in the diff touches AGENTS.md (the DECISIONS.md append is a documentation record, not a constitution edit), no citations are added, and no run records are altered in this change — the hand-corrected stamps of run 2026-09-11-report-orch-or-114458 are explicitly deferred to a following content PR and are described as carrying notes saying they were corrected by hand, so no misattribution is introduced here. I cannot verify the narrative claims about the run itself (29 iterations, $3.72 vs $3.13) from what was shown — no RUN ACCOUNT was included — but those are descriptions of a past event, not records this PR publishes, and the test fixtures (2 iterations) are plainly synthetic rather than claimed observations. If the follow-up content PR ships the hand-corrected run records, those corrections should be judged on their own packet.