Wave-6 item 6. Per the wave-6 brief: "CAG" as a named research
artifact is a 4-page short paper; the deployed reality is
**prefix KV-cache reuse exposed as a billing line item by every
major LLM provider**. We ingest the deployed shape — the
0.1× input-token cache-read multiplier convention that converged
across Anthropic, Google, OpenAI, and DeepSeek between Oct 2024
and mid-2025.
Surface:
- `CachePolicy { provider, write_multiplier, read_multiplier,
ttl_kind: { Explicit | Implicit }, ttl, min_tokens }` —
every shipping cache provider fills this struct. Multipliers
clamped to [0, 100] so transposed values can't blow up cost
arithmetic. `matches_converged_read_multiplier()` checks the
≤0.15 (≥85% off) ceiling.
- `vendor::*` presets verified against the brief's table:
anthropic_5m (write=1.25, read=0.1, 5min explicit),
anthropic_1h (write=2.0, read=0.1, 1h explicit),
google_gemini_flash (write=1.0, read=0.1, 1h explicit,
storage billed separately), openai_automatic (write=1.0,
read=0.1, implicit, 5min typical), deepseek_disk_cache
(write=1.0, read=0.1, implicit, 1h typical).
- `CacheEvidence { cache_read_input_tokens,
cache_creation_input_tokens, fresh_input_tokens, output_tokens,
provider }` — the converged usage-struct shape Anthropic +
DeepSeek already return per call.
- `hit_rate()` — `cache_read / (cache_read + fresh)`,
deliberately excludes creation tokens (those are one-off write
cost, not part of the hit denominator).
- `CostEstimate::from_policy_and_evidence(...)` — realised cost
breakdown: `effective_input_cost`, `output_cost`, `total_cost`,
`uncached_total_cost`, `savings_ratio` (clamped to [0, 1]).
Rejects provider mismatch between policy and evidence (apples
and oranges guard).
Honest scope (v1): types + computation only. We do NOT call any
provider's cache API. Cost units are arbitrary; the consumer
scales to $ / joules / quota. TTL for implicit providers is a
typical observed value, NOT a guaranteed lower bound — the
TtlKind::Implicit tag makes that explicit.
13 tests cover: all five vendor presets satisfy the converged
read-multiplier test; explicit vs implicit TTL classified
correctly; multipliers clamp into [0, 100]; total_input_tokens
sums three categories; hit_rate excludes creation tokens +
zero-input case; cost arithmetic applies read_multiplier
correctly + provider-mismatch rejected + write_multiplier
applied to creation tokens; zero-savings edge case; CachePolicy
+ CacheEvidence JSON round-trip; non-converged read multiplier
fails the predicate.
Co-Authored-By: Claude Opus 4.7 <[email protected]>
d2a9c4ee1f