Isolated environment
A separately scoped pilot environment and data boundary for the evaluated workflow.
14-day design pilot · $5,000 fixed
Prove governed reuse on one workflow.
The pilot answers a narrow question: can CachePilot make a real hosted-model workflow reliably reuse eligible state without losing policy control, attribution, or evidence?
Fixed scope and explicit boundaries.
One workflow, one provider path, one primary hypothesis. Provisioning is manual while the product and operating model are still being proven.
Baseline evidence
Direct or observe-mode measurements for stable-prefix behavior, expected reuse, actual cached tokens, latency, token volume, drift, and cost.
Governed comparison
The same workflow with approved stable-prefix repair, cache key or breakpoint handling, output budgets, and policy receipts.
Decision package. A concise evidence report, request and receipt examples, known limitations, recommended policy, and a proceed/change/stop recommendation.
Cost is measured, not assumed to lead.
The pilot starts with execution evidence. A cost result is useful only when the reused work was valid and the response still met policy and quality expectations.
| Measure | Question | Role |
|---|---|---|
| Reuse eligibility | Was the repeated request structurally eligible for provider reuse? | Leading |
| Expected vs actual reuse | Did the provider report cached input on the repeat? | Leading |
| Prefix and policy drift | What changed, when, and was the invalidation intended? | Leading |
| Latency and reliability | Did governance change TTFT, total latency, errors, or retries? | Leading |
| Receipt completeness | Can each decision be tied to policy, prefix, mutations, and attribution? | Leading |
| Token cost | What cost outcome followed from valid reuse and output policy? | Downstream |
A credible pilot workload.
Good candidates have repeated stable context, enough prefix size to support provider caching, measurable request volume, and an operator who can approve policy. Constantly changing short prompts are usually poor candidates.
Included ✓ one OpenAI Responses or Anthropic Messages workflow ✓ isolated pilot environment ✓ observe baseline and governed comparison ✓ content-free telemetry and X-CP-* receipts ✓ final evidence and production recommendation Not included × fleet-wide rollout or automated tenant provisioning × formal production SLA × a guarantee that an opaque provider cache will hit × prompt/output storage unless separately approved