Skip to content

Epic: calibration expansion — self-correcting configuration and measured judgment #8211

Description

@JSONbored

Problem

Epic #8082 closed the loop for TWO knobs and GLOBAL settings: the gate now measures its own judgment against labeled ground truth, replays rule changes against history before they land, and loosens conservatism only where evidence clears a Pareto-floored backtest. That machinery is generic — the corpus carries repo identity in every target key, the knob registry is declarative, the raw-context capture stores the exact inputs AI judgments evaluated — but almost none of that generality is exploited yet. Meanwhile the operator's single biggest wrongness source is now known empirically (retro-label analysis, #8170): decisions that were faithful executions of STALE configuration — rules whose real-world expectations had moved while the config had not. Nothing detects that today; it is discovered through frustration.

Proposal (decompose-first, mirroring #8082's own structure)

Seven tracks, each independently valuable, sequenced by their GitHub relationships (sub-issues + blocked-by — the dependency graph IS the schedule):

  • A. Config-drift sentinel — replay the CURRENT live config against the trailing corpus nightly; alert with evidence when an alternative setting Pareto-dominates what is live. Turns stale-config wrongness into a push notification.
  • B. Per-repo calibration — repo-scoped corpus slices, per-repo knob overrides under per-repo evidence minimums. Repos earn individually different autonomy.
  • C. AI counterfactual backtesting (own sub-epic) — replay prompt/model variants against the raw-context corpus offline, so judgment-layer changes get the same evidence gate as rule changes.
  • D. Knob-registry expansion + tightening unification — more knobs under the discipline; fold the legacy tightening loop under the same backtest standard.
  • E. Calibration curves — reliability curves (claimed confidence vs empirical precision) that DERIVE thresholds instead of stepping hand-picked ladders.
  • F. Reviewer-provider track records — measured per-provider precision, then evidence-weighted reviewer routing.
  • G. Public trust surface — publish per-rule precision + a 'verify this review' reproducibility walkthrough; measured accuracy as the product's public asset.

AMS-side expansion continues separately under #8172 (its own epic, already decomposed).

Deliverables

Deliberately none at the epic level — every deliverable lives in a sub-issue. The epic closes when every sub-issue does.

Links & Resources

#8082 (the closed foundation epic), #8172 (AMS sibling), src/services/loosening-knobs.ts + knob-loosening-run.ts (the registry this extends), packages/loopover-engine/src/calibration/** (the shared primitives)

Boundaries

Same discipline as #8082 throughout: nothing autonomous without explicit opt-in flags, gate-authority changes are maintainer-only PRs, every new capability ships default-off or advisory, and no corpus content ever reaches a public surface — aggregates and checksums only.

maintainer-only — epic; work the sub-issues, not this shell.

Metadata

Metadata

Assignees

Labels

maintainer-onlyOwner-only work — yields no Gittensor points.roadmapOn the Wave-2 agent-layer roadmap board (project 9)

Projects

No projects

Relationships

None yet

Development

No branches or pull requests

Issue actions