Skip to content

Milestones

List view

  • Self-Improvement · Phase 5 — Close

    Automate the cycle. Scheduled collect→reward→train→eval→gate→deploy; CalibrationMap-style observe→persist→reload→apply behind a feature flag; every model/dataset/bench report versioned with instant rollback. Playbook §08.

    No due date
    •0/2 issues closed
    0% complete2 open0 closed

  • Self-Improvement · Phase 4 — Validate

    Prove it got better. Baseline-vs-candidate on the frozen bench suite (SWE-bench, terminal-bench, smoke, loop-bench), promotion gate with no-regression thresholds, redeploy through the Provider trait preserving judge ≠ worker. Playbook §06–§07.

    No due date
    •2/5 issues closed
    40% complete3 open2 closed

  • Self-Improvement · Phase 3 — Refine

    Train on the signal. SFT / DPO / RL datasets from labeled traces (dedupe + decontaminate against bench), then fine-tune: SFT on winners → DPO on preference pairs → online RL with stella's verify ladder as the reward function. Playbook §04–§05.

    No due date
    •2/9 issues closed
    22% complete7 open2 closed

  • Self-Improvement · Phase 2 — Reward

    Turn outcomes into signal. Export the verify ladder's LadderDecision as reward labels (+1.0 deterministic pass … −1.0 deterministic fail), composite reward with cost/steps/revisions, judge reasoning as soft signal only (feedback-airlock rule). Playbook §03.

    No due date
    •1/2 issues closed
    50% complete1 open1 closed

  • Self-Improvement · Phase 1 — Instrument

    Make the loop observable. Capture full execution trajectories (prompt messages, tool calls, stage trajectory, diff fingerprint, cost) via the HookRunner channel (stella-core/src/hooks.rs:249) and land them in ContextStore episodic memory. Playbook: stella-docs/playbooks/self-improving-model.html §02.

    No due date
    •1/1 issues closed
    100% complete0 open1 closed