Skip to content

calibration: generalize backtest-gated loosening into a bounded knob registry beyond the satisfaction floor #8159

Description

@JSONbored

Problem

The #8121 narrow start hardcodes ONE loosenable knob (the satisfaction floor). The epic's full intent is a registry of loosenable configuration values, each with its own hard safety bounds, all gated on the same backtest discipline.

Requirements

⚠️ Required pattern. Mirror KNOWN_THRESHOLDS (src/services/threshold-backtest.ts) / KNOWN_LOGIC_RULES (scripts/backtest-logic-check-core.ts): a declarative registry mapping each knob to its rule ids, candidate steps, and hard bounds — never per-knob bespoke loops.

  • Every knob keeps the narrow start's invariants: smallest-step-first, visible-improved + held-out-non-regressed, hard minimum, write-path direction re-checks, audit event per apply.
  • Do not add a knob without real fired/override ground truth for its rule ids.

Deliverables

  • The registry + generalized evaluator (refactor of satisfaction-floor-loosening.ts, keeping its tests green)
  • At least one additional knob wired end-to-end with its own bounds and tests

Links & Resources

#8121 (parent), src/services/satisfaction-floor-loosening.ts (the shape to generalize)
maintainer-only — extends autonomous config change; each knob is a deliberate decision.

Metadata

Metadata

Assignees

Labels

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

Projects

Status
Done
Status
Done

Relationships

None yet

Development

No branches or pull requests

Issue actions