You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Epic #8082 gave ORB a full self-measuring calibration loop (capture → labeled corpus → backtests → track record → trend → bounded self-loosening). AMS has nothing equivalent: miner nodes keep rich local ledgers (laptop-state.sqlite3, event-ledger.sqlite3 — opportunity selection, plan/execution outcomes, governor chokepoint decisions, PR results) but none of it feeds a calibration loop, so AMS's own deterministic knobs (opportunity ranking weights, governor thresholds, deny-hook synthesis bounds) are tuned by hand and never validated against the node's real outcome history.
Proposal (epic — decompose before building, mirroring #8082's own structure)
Reuse, don't rebuild: @loopover/engine's calibration primitives (buildBacktestCorpus, scoreBacktest, compareBacktestScores, splitBacktestCorpus, the render) are storage-agnostic and already shipped — the AMS work is adapters over the miner ledger, not new math.
Fully local: a node's corpus never leaves the node (consistent with AMS's privacy posture); fleet-level aggregation, if ever, goes through the existing federated surfaces as a separate decision.
Deliverables
Deliberately none yet — this epic needs decomposition against the actual miner-ledger schemas first (mirror how #8082 was decomposed before any code).
Links & Resources
#8082 (the ORB template for this loop), packages/loopover-engine/src/calibration/** (the shared primitives), packages/loopover-miner/lib/** (governor/opportunity/ledger surfaces)
Boundaries
Nothing autonomous without its own explicit approval gate, mirroring #8121's discipline. No cross-node data movement.
maintainer-only — epic; decompose before building.
Problem
Epic #8082 gave ORB a full self-measuring calibration loop (capture → labeled corpus → backtests → track record → trend → bounded self-loosening). AMS has nothing equivalent: miner nodes keep rich local ledgers (
laptop-state.sqlite3,event-ledger.sqlite3— opportunity selection, plan/execution outcomes, governor chokepoint decisions, PR results) but none of it feeds a calibration loop, so AMS's own deterministic knobs (opportunity ranking weights, governor thresholds, deny-hook synthesis bounds) are tuned by hand and never validated against the node's real outcome history.Proposal (epic — decompose before building, mirroring #8082's own structure)
@loopover/engine's calibration primitives (buildBacktestCorpus,scoreBacktest,compareBacktestScores,splitBacktestCorpus, the render) are storage-agnostic and already shipped — the AMS work is adapters over the miner ledger, not new math.Deliverables
Deliberately none yet — this epic needs decomposition against the actual miner-ledger schemas first (mirror how #8082 was decomposed before any code).
Links & Resources
#8082 (the ORB template for this loop),
packages/loopover-engine/src/calibration/**(the shared primitives),packages/loopover-miner/lib/**(governor/opportunity/ledger surfaces)Boundaries
Nothing autonomous without its own explicit approval gate, mirroring #8121's discipline. No cross-node data movement.
maintainer-only — epic; decompose before building.