Evidence map clause
4.9 Provider requests and prompt caching
This clause and its evidence summary follow the authored EVIDENCE-MAP.md section.
Evidence map clause · EVIDENCE-MAP.md
Map clause: Cache key, stable prefix, accounting, and checkpoint.
Source heading anchors: spec/04-provider.md#49-provider-requests-and-prompt-caching; spec/04-provider.md#491-conversation-key; spec/04-provider.md#492-byte-stable-prefix; spec/04-provider.md#493-tool-result-budget; spec/04-provider.md#494-the-upstream-cut-and-continuation; spec/04-provider.md#495-usage; spec/04-provider.md#496-opt-in-context-checkpoint
Spec revision / decision reference: v1.2 (pinned fragment locally verified); section-level D references are listed in the source classification. Last reviewed: 2026-10-07.
Type: Safety/correctness for cache-key and counter semantics; empirical for cache-hit and cost claims.
Evidence and status: R54 is INTERIM; R74 is WITHDRAWN; T98 cache smoke is INTERIM (
CL-WA-t98-cache-smoke-reported-observation, INTERIM) and below its stated gate. The cache replay mechanism is INTERIM and publishes sanitized request counters; 12 assigned attempts did not dispatch, and the backend run used a different authentication source from the design. It is DESCRIPTIVE and does not establish an endpoint effect or cache gate. T100 is NOT-RUN (CL-WA-t100-cache-lane-reported-observation, NOT-RUN). Historical cross-harness values remain INTERIM andsource-reported, not reproducible from public data:CL-H115-historical-cache-hit-comparisonandCL-H116-historical-cache-change. Offline context-store observations are not Build outcomes: context store is INTERIM (CL-WA-ctx-store-reported-ctx-store-ranked-paths, INTERIM). Historical definitions and epochs are not reconciled to T98. Related hypotheses: H51–H61, H115–H116, H134, H148–H150.Unresolved measurement: Repeat cache telemetry with a frozen request set, identical token definitions, matched authentication source and venue, and public request-level records; verify any cache gate independently.