Design: define policy simulator features, estimators, and exploration safety

[devkade]

Design: define policy simulator features, estimators, and exploration safety

Parent: #167
Related: #171, #172

Summary

Define how the PolicySimulator estimates the impact of candidate execution policies before dispatch and how it safely chooses exploratory policies for learning.

Scope

Define simulator inputs:

• objective metrics and weights;
• task complexity signals;
• expected touched files/modules;
• dependency depth;
• worker count;
• adapter mix;
• isolation mode;
• verification depth;
• historical strategy stats;
• recent RewardRecord calibration.

Define estimators:

• conflict risk;
• regression risk;
• repair likelihood;
• elapsed/tool cost;
• review burden;
• learning value;
• confidence.

Non-goals

• No neural model requirement.
• No agent launch during simulation.
• No hard-blocking score authority from simulation alone.

Acceptance criteria

• Simulator input feature set is documented.
• Estimator outputs and ranges are defined.
• Exploration policy and safety caps are defined.
• Human override representation is defined.
• Prediction ids are linked to RewardRecord calibration.

Verification

• Fixture examples comparing conservative, balanced, aggressive, high-assurance, and exploration policies.
• Design review against Roadmap: objective-driven learning parallel runtime harness #167 advisory-boundary constraints.

[devkade]

Design decision

needs-design is resolved by #167's Development Order and the shallow-first rule.

Decisions:

• Design: define policy simulator features, estimators, and exploration safety #187 is shallow-first and should not block the runtime MVP.
• The first simulator is deterministic/rule-based, not neural/RL-heavy.
• It may consume objective metrics, rough task complexity, dependency depth, adapter mix, verification depth, historical stats, and reward-ledger calibration when available.
• It produces estimates for conflict risk, regression risk, repair likelihood, elapsed/tool cost, review burden, learning value, and confidence.
• Exploration is capped, explicit, and advisory.
• Prediction ids must link to later reward-ledger records.

Implementation can proceed as Step 7 of #167 after #196 creates real execution evidence.