DynosAI 0.17.0 — Eval Intelligence

0.17.0 sits on the 0.16.0 governed-team baseline and turns local failures into bounded eval cases without handing quality or routing to the model.

What this release is

Eval intelligence is a learning-loop problem, not a live leaderboard. DynosAI attributes failures, mines local traces, and can open an inbox improvement task. It still does not run live-provider evals and still does not let predictive routing spend model budget.

Shipped

Failure attribution across model / harness / provider / project / infrastructure / governance.
Mining of validations, audit failures, eval records and model-control outcomes into bounded cases.
Host-owned propose: inbox work only, no provider spawn.
Offline regression evidence that does not auto-complete or auto-approve work.
Studio Overview eval cases and dynosai_stats summary.
Offline team_overlap and mined_regression fixtures.

Deliberately not shipped

Live-provider eval quality claims or leaderboards.
Autonomous predictive-routing enablement.
Silent creation of improvement work from MCP agent sessions.
Schema 7 / a new eval authority table.

Authority unchanged

Schema remains v6. Git is source truth. .dynosai/knowledge.db is workflow truth. Eval cases live under .dynosai/runtime/. Cursor ACP and Codex app-server transports are unchanged. 0.16 leases and the 0.15 harness are preserved.

Validation

Release still requires repository checks, Studio asset sync, the stable pytest suite, and website check/typecheck/lint/build when web files change.