Skip to main content

Published Conformance Results

Maintainer self-test

The repository includes a machine-readable v1.0.0 pre-release report at benchmarks/results/conformance-v1.0.0-pre-release-windows.json. It covers all seven required domains. Timing and allocation values are environment-scoped and are intentionally not promoted as total proxy latency or a universal memory ceiling.

Run: 2026-08-30 on Windows 11, CPython 3.14.7, AMD64; 10,000 timing samples per operation.
Source label: eedb5b2044bfc806ae5cfc33f4627f79cf2c70cc+working-tree
SHA-256: 0921910bdc0438c140f1731315e5683d53197770ae68beae5b6b80179764502c

Because the implementation changes are not yet committed, this is a transparent maintainer pre-release self-test, not a release-grade independently reproducible result. CI should regenerate the report from an exact commit SHA before a formal release.

DomainLatest status
Fragmentation safetyPass — 11 partitions, no reported failures
Raw protected-data egressPass — email, SSN, and credit-card fixtures absent at tested boundary
SSE validityPass — valid JSON events, one [DONE], valid termination, split UTF-8 preserved
Rehydration fidelityPass — exact equality; no placeholder in client-visible output
Audit integrityPass — two signatures verified; tamper negative control detected
Latency measurement completenessPass — distributions recorded; no threshold enforced
Memory bound/measurement completenessPass — retained state within bound; allocation recorded; no RSS threshold

In-process timing observations

Operationp50p95p99
Empty-vault buffer13.9 µs19.6 µs30.6 µs
Protected-token buffer21.2 µs30.8 µs44.0 µs

These are Windows in-process Python operation timings. They exclude ASGI, HTTP, TLS, network, upstream/model, concurrency, and durable audit I/O. The measured Python allocation peak was 4,656 bytes for the declared allocation scope; it is not process RSS.

Independent reproductions

No unaffiliated reproduction has been published yet. This is an explicit evidence gap and the highest-priority community contribution.

Reproduce the report and submit the raw artifact. Do not send representative enterprise traffic or confidential prompts.