POWER 3.1.0 release evidence¶
POWER 3.1.0 is a release candidate specification. Runtime behavior and documentation must be backed by reproducible, versioned artifacts before production claims are made.
What is pinned¶
The canonical dense model is the bge-m3-onnx export listed in
release/models.lock.json,
including upstream/export revisions, licenses, file hashes, and sizes. The
optional Jina cross-encoder is non-commercial (CC-BY-NC-4.0) and requires
explicit opt-in; it is not a production default.
Benchmark status¶
benchmarks/power31 is a deterministic synthetic bilingual corpus. It is
useful for hermetic CI regression gates only: it is not human annotated, does
not contain real vault notes, and cannot establish production retrieval
quality, latency, or memory claims. Historical 3.0 reports are retained for
context and are diagnostic only. UDCG in those reports is a legacy lexical
proxy, not the EACL-2026 metric.
Release gates¶
The trust-release ADR defines the fail-closed retrieval, MCP boundary/transport, provenance, and evidence requirements. A release evaluation must retain its manifest, dependency lock, hardware profile, model revisions, raw outputs, and verifier result. Until the P0/P1 gates are closed, POWER remains beta/research software.
Run the benchmark integrity checks locally with:
python -m pytest benchmarks/power31/tests/ -v --no-cov --override-ini="addopts="