DSI · decision-space integrity

Release notes · v0.2.1

DSI Product 0.2.1

A hardening release: installation, diagnostics, packaging, and a context-admission fix, with Regression Audit as a first-class capability. No conceptual redesign — this builds on 0.2.0 without a public API break. Still evaluation & pilot, still local-first.

Added

  • Deployment doctordsi doctor, a read-only readiness check you run first. "Ready for audits" means installation readiness only.
  • Installation validationdsi validate-install runs a real, domain-consistent audit end to end (audit → evidence card → regression).
  • Readiness endpointGET /ready (503 when not ready); an enriched GET /version.
  • Regression Audit — first formal release: an instrument-relative comparison of paired outputs, with instrument-limited verdicts and provenance.
  • Clean-install verification — the built package is installed in an isolated environment and checked to run from documentation alone.
  • Consolidated documentation — getting-started, deployment profiles, and security & operations guidance.

Fixed

  • Embedded decision phrasing. Indirect questions such as "I don't know if I should stay or look elsewhere" are now recognised as decisions and audited against the full applicable set, matching their direct form "Should I stay or look elsewhere?" — rather than being reduced to a single clarification path.
  • Install validation consistency. The end-to-end check now audits with a prompt matched to the selected domain, rather than pairing an arbitrary prompt with an arbitrary domain.

Compatibility

  • D-CTX-1 did not change classifier or scorer logic — the fix changed how the applicable set is constructed, not the matching or the scoring. Separately, v0.2.1 ships classifier expected-map-lexical-precision-v3, which rebases audit identities: audit ids from v0.2.0 bundles do not carry forward.
  • Public output schema unchanged.
  • The frozen 54-prompt compatibility replay showed zero denominator or coverage changes.

Known limitations

The lexical classifier is challenge-tested (challenge-only); human review is recommended. Regression Audit verdicts are instrument-relative — they report a change in configured expected-path visibility, not that one build is better, safer, or of higher quality. DSI is not intended for autonomous high-stakes decision making or response blocking, and is not human-validated.

This work measures visibility of configured expected paths in model outputs. It does not measure advice quality, factual correctness, user outcomes, or regulatory compliance.