DynosAI 1.0.0-rc.3

Public RC · Eval maturity. This is not stable 1.0.

Package version: 1.0.0rc3 (PEP 440). Display: 1.0.0-rc.3. Development classifier remains Development Status :: 4 - Beta. Database schema remains v6. MCP unique surface remains 31.

What RC3 proves

• A host-owned acceptance ZIP importer turns failed suite children into bounded Eval Registry cases. Improvement work stays in inbox. DynosAI does not spawn a provider.
• governed_change_cost() is the promoted efficiency metric. Completed-work aggregates appear on dynosai stats, dynosai scorecard, project overview, Studio and diagnostic bundles.
• Measured fields include success, validation state, fresh/cached/output/reasoning tokens, retries, human gates, elapsed time, MCP calls/failures/rejections/normalizations, and duration/result bytes when those values were actually recorded.
• Unused advertised MCP tools are not treated as waste. Tool-surface size is an experimental context-overhead diagnostic.
• A deterministic PROMPT_PREFIX_1.0 authority prefix is hashed and stored under .dynosai/runtime/ plus audit. The prefix does not dump the 31-tool list. Phase context stays in a suffix. The hash means cacheable structure, not a provider cache hit.

What RC3 does not prove

• No live-provider 1.0 certification. MATRIX_1.0 cells remain not_run.
• No installer or published PyPI package.
• No OS-level sandbox, Docker, VM or remote execution runtime.
• No additional certified clients.
• No autonomous predictive routing.
• No claim of production-ready 1.0.

Next

RC4 is installation, doctor-in-Studio and installed-package CI. Do not start it from this note alone; wait for an explicit instruction.