This is a pipeline/infrastructure change about prompt caching; it touches no content/ records, adds no citations (verification found none), and does not edit AGENTS.md. The checks in §3.15 are untouched: the Arbiter still sends the full constitution to every seat — it is merely relocated from the packet body to a `cachedPrefix` block placed first in the same user turn (scripts/arbiter.mjs, src/lib/vendors.mjs), so no seat loses information and no ratification path is loosened. Verify likewise still passes the full retrieved source text and the split step its anchoring text; only their position in the request changes, and both builders are now pure functions with tests. The usage accounting change (subtracting cached tokens from inputTokens and reporting cacheRead/cacheWrite separately) makes the spend ledger more, not less, faithful to what the vendor reported. The DECISIONS.md entry records the founder direction of 2026-09-10 with its date and quoted wording, which §3.15 expressly permits; the trailing literal `\n` and missing final newline are cosmetic. Nothing here presents a draft as ratified, fabricates provenance, or publishes confidential material.

The added `docs/DECISIONS.md` entry attributes the caching policy and the quotation “use this strategy with all api calls” to a founder direction dated 2026-09-10, but it does not record the channel through which that direction was given. Section 3.15 expressly requires founder directions recorded in the decisions log, configuration, or protocols to be stamped with both date and channel. The PR-body session link does not cure the provenance omission in the public decision record itself. Add the channel—such as the identified working session, issue, or PR review—and, where appropriate, its stable record location; the caching implementation otherwise visibly preserves the constitution in the Arbiter prompt and does not weaken the adjudication checks.

The change introduces deliberate prompt caching controls across API calls, switching from automatic caching on one-shot calls to explicit cached prefixes where large contexts are reused (such as the constitution in Arbiter seat calls, and source text in verification and splitting). AGENTS.md is not modified, and the integrity of the constitutional verification checks in scripts/arbiter.mjs and src/pipeline/ is preserved without altering prompt semantics or weakening validation gates. Decisions are properly logged in docs/DECISIONS.md and accompanied by unit tests.

The diff is infrastructure only: it turns off automatic Anthropic cache_control on one-shot structured calls, adds an optional cachedPrefix breakpoint for shared text (verify source, split anchor, panel constitution), and surfaces cache read/write tokens in vendor usage and the seat meter. AGENTS.md is untouched; docs/DECISIONS.md only records the founder’s 2026-09-10 spend finding and the same rule. Seat packets still receive the full constitution plus the untrusted change—only the transport shape changes—so §3.15 concurrence and the other epistemic checks are not weakened. No content/, citations, assessments, or confidence material are added. Tests cover the pure request builders.

This is an infrastructure/cost change, not a content or epistemic change: no citations, locators, or assessments are added, and the citation-verification section confirms none appear. The constitution still reaches every seat in full — arbiter.mjs merely moves the identical constitution text into a cachedPrefix block ahead of the same packet, so no check in §3.15 is weakened, dodged, or reclassified; verify's reader and the split step likewise still receive the full source and anchoring text, now with an explicit 'data under review' label, which if anything improves §3.9-style labeling. Removing automatic caching from one-shot calls does not alter what any model judges, only what it costs; the honest-usage reporting (splitting cache reads/writes from input) is more transparent than before. The DECISIONS.md entry records the founder direction with a date and quotes it, which §3.15 permits for operating-record directions; I cannot independently verify the session, but there is no positive evidence of misattribution, and no AGENTS.md edit occurs. Minor non-violations worth a review note: a stray literal '\n' line left at the end of DECISIONS.md, and the PR body claims priceOf now prices cache reads/writes, but no priceOf change appears in this diff — I would want confirmation that ledger pricing actually consumes the new fields.