Toxicity-MM gap-aware replay
supporting-measurementstrategy
Question
Replay frozen strategy policies without bridging corpus gaps or hiding queue ambiguity.
Verdict
Gap-aware aggregate replay exists; a temporal gap is a no-trade coverage outcome and is never interpolated.
- Population
- Frozen private full-book development corpus.
- Resolution
- Five-minute candidates over raw full-book observations.
- Cutoff
- Development span only; aggregate output only.
- Baseline
- Nine frozen policy/ambiguity axes including pessimistic queue treatment.
Limitations
Snapshot queue ambiguity remains an explicit axis; raw observations remain private.
Notes
No additional notes published for this entry.
Source
- Citation
- po-research:toxicity-mm-replay:909e3d35817d9a2a
- Write-up
- docs/research/toxicity-mm-replay.md
- Data
- docs/research/toxicity-mm-replay.json
- Artifact
research-index.jsonv1 — sha256:cb3d976779a5cbcb4188039a8d26d5e558a2d844030a638c2f16fe5a8eb0a2fe- Build data cutoff
- 2026-08-12
- Issues
- #174