PRD 002: Dislocation Research Sleeves

Goal

Expand Pam’s research search space with sleeves that target actual dislocation patterns instead of static heuristic categories.

Problem

The current deployment-grade gate of 100+ bps correctly yields zero trades, but that also means the research loop needs better sleeves to discover where such edges might appear.

Candidate Sleeves

1. Price-jump reversal - eventful price move - near-term fade or continuation decision
2. Spread-shift setup - widened spread followed by normalization
3. Near-resolution dislocation - eventful score plus short time-to-resolution
4. Tradeability shock - sudden improvement or deterioration in tradeability
5. Threshold ladder inconsistency - adjacent threshold markets imply non-monotonic probabilities

Requirements

1. Add bounded sleeve rules to research/strategy.py and mirrored search logic in research/search.py.
2. Keep deployment and research separate: - research sleeves may explore - deployment sleeves must still respect the hard edge gate
3. Log sleeve provenance in replay rationale so we know which family generated each action.

Success Criteria

1. Replay action count comes from multiple sleeve families, not one narrow rule.
2. Search results identify at least one sleeve family with materially better score than the current baseline.
3. Negative sleeves are easy to disable without changing the replay harness.

Non-Goals

1. Fitting a complex ML model
2. Trading on narrative alone

Notes

Bounded, interpretable sleeves are more valuable than a broad opaque model at this stage.