Skip to content

ADR 0001: P.O.W.E.R. 3.1 Trust-Release Baseline

Status: Accepted

Date: 2026-07-21

Context

P.O.W.E.R. 3.0.0 has useful retrieval, MCP, indexing, and lint capabilities, but its published quality, security, and production-readiness claims do not yet have reproducible evidence or complete release gates. P.O.W.E.R. 3.1 is a trust release: it must correct those foundations before adding capabilities.

The source baseline is commit b2700058d717b49169ce284a48ed731b3cd56d21.

Input SHA-256
docs/tests/P.O.W.E.R.3.0.0-TEST.md 3c8d2df19c82beedb2c4abc88fa0eb76c485b8287ea17d2ba42111db65e0aed3
docs/tests/P.O.W.E.R.3.0.0-TEST-2.md d6bfeaa5e68f98428246189f7b670acfd01f4e010cc66a445e73f39ef8c74af8
src/power_framework/mcp/power_server.py 6ea5d03795afca4894e5df136f02837bf4c24e1cc20c7931db70fceef3d00f4c

The historical reports reference temporary, untracked benchmark files. They are not release evidence because their runtime configuration, dependency lock, hardware profile, and raw query outputs are not versioned.

Decisions

  1. Treat P.O.W.E.R. 3.0.x as beta/research software. Do not claim it is production-ready, safe for remote MCP deployment, or quality-gated until the relevant 3.1 release gates are reproducibly satisfied.
  2. The word UDCG is reserved for the EACL 2026 Utility and Distraction-aware Cumulative Gain. The current metric must be renamed or explicitly marked legacy before it can no longer be a primary gate.
  3. Requested dense or reranked retrieval must fail closed if its compatible index or model is unavailable. Any permitted fallback must appear in result metadata and operational status.
  4. MCP retrieved text is untrusted data. It requires provenance and a structured envelope; it must not be treated as an instruction for an agent.
  5. MCP writes must stay within a configured canonical vault root, validate an untrusted relative path centrally, and use atomic writes. HTTP must bind to loopback by default; remote use requires an explicit authorization policy.
  6. A release decision must be based on versioned fixtures, manifests, model provenance, and reproducible outputs, not aggregate scores from a private vault alone.

Consequences

Phase 1 starts with the MCP path boundary and transport hardening. Phase 2 unifies retrieval contracts, Phase 3 repairs metrics and benchmarks, and later phases add graph, model, supply-chain, and performance gates. Every phase must carry focused regression tests and document its validation result.

Phase 0 Exit Evidence Still Required

  • A versioned, redacted runtime configuration and dependency-lock hash.
  • A machine-readable baseline manifest with hardware profile and model revisions.
  • Retained raw benchmark output or an explicit statement that no comparable historical artifact exists.