Limits¶

What this product does not do, and where the estimates stop. No roadmap promises here — just the current boundaries, stated plainly.

Product boundaries (by design, not omission)¶

No live proxy or gateway. We never sit in your request path. The audit is offline analysis of logs you already have. (X-01)
No enforcement. We report waste; we never act on your account, set budgets, or block calls. (X-02)
No multi-org RBAC or SSO. One email, one account, magic-link sign-in. (X-03)
No AI-written report text. Every sentence in a report is a template; every number is deterministic arithmetic. (X-04)
Plain server-rendered pages, no SPA. (X-05)

Aggregate-only exports underestimate. If your provider export carries daily aggregates instead of per-call records, per-call detectors (retries, loops, caching) cannot run at full power; the report says which detectors ran degraded.
Unpriced models are excluded, not guessed. Calls on models without a verified rate card are listed and left out of totals.
Cache-write floors. Cache savings subtract estimated write costs; where write windows cannot be estimated, a 0.7 haircut applies. These floors mean our cache numbers are lower than most tools' — deliberately.
OpenAI cache-write counts are a tracked gap. OpenAI usage exports do not carry per-call cache-write counts, so the write side of cache simulation is estimated from cache-lifetime windows for OpenAI traffic. Anthropic exports carry both fields natively.
Downgrade savings assume quality holds. Every oversized-model finding says so explicitly; we can prove the price delta, only you can prove the quality equivalence.
Monthly figures are extrapolations of your observed window (× 30 ÷ observed days). Send a representative window — a quiet weekend sample will under-report.

Uploads to 200 MB per audit (larger sets: split by time range).
One audit runs per credit; processing beyond the concurrency cap queues, with position visible in the status API.
Report links expire after 30 days; raw uploads purge after 7. Download your PDF.