Skip to content

current_direction.md

Latest commit

 

History

History
50 lines (43 loc) · 2.54 KB

current_direction.md

File metadata and controls

50 lines (43 loc) · 2.54 KB
Raw

Current Direction

trinity_framework is the source-of-truth repository for the deconstructed TRINITY stack. Runtime behavior, operator commands, examples, and Crucible assembly are owned here through explicit package boundaries.

Active Priorities

  1. Keep reusable packages small and explicit.
  2. Keep TRINITY product naming in framework packages, not in Crucible packages.
  3. Maintain the complete operator command surface from the framework root.
  4. Keep artifact, eval, parity, and docs gates runnable from a fresh checkout.
  5. Publish deconstructed packages through governed dependency manifests.
  6. Expose Crucible as the reusable route path without hard-wiring it to Qwen.
  7. Close the route-evidence loop with Orchestrator-backed, allowlisted Sakana fitness datasets for external ES training.
  8. Make router confidence operational through the Orchestrator reflex policy: high margins dispatch directly, medium margins preserve normal behavior, and low margins force Thinker before Verifier.

Evolution Boundary

The framework owns evidence production, deterministic fitness assembly and scoring, router-vector artifact import/export, parity, and runtime acceptance. Candidate generation, optimization, and weight mutation stay in an external ES trainer. This boundary keeps training dependencies and mutation policy out of the reusable runtime while preserving a schema-versioned handoff in both directions.

Done Means

  • mix test runs root aggregate tests.
  • mix ci is clean.
  • mix help --search trinity lists the complete operator task surface.
  • The 37-case Qwen router eval is documented and runnable.
  • mix trinity.eval qwen_router_prompt_eval passes strict route-decision acceptance.
  • CUDA/adapted checks are documented and run on capable hosts.
  • mix trinity.orchestrator.demo produces verifier/budget-bearing traces and mix trinity.sakana.fitness_export converts them into deterministic datasets.
  • Route confidence appears as reflex_decision evidence, and low-confidence Orchestrator runs exercise Thinker-to-Verifier escalation without creating a separate agent loop.
  • Verifier revision counts describe state after the current decision, and exported fitness examples contain only allowlisted trace fields.
  • External training can return a reviewed router vector to the existing adapted-export, parity, eval, and CUDA acceptance path.
  • The adaptation readiness loop now adds dataset inspection, score replay, reflex calibration, and non-mutating candidate proposal reports before any candidate vector enters artifact export gates.