Skip to content

Latest commit

 

History

History
44 lines (34 loc) · 2.16 KB

File metadata and controls

44 lines (34 loc) · 2.16 KB

Determinism Report

Determinism controls include fixed corpus ordering, fixed timestamps, stable event IDs, deterministic JSON serialization, seeded replay passes, source/compressed/token hash comparisons, and tokenizer drift sentinels.

Release blockers

  • Replay output changes for any golden dataset.
  • Token hash changes for a fixed compressed payload without documented tokenizer version change.
  • Mutation of any existing golden corpus file.
  • Any critical or high forensic loss.

Evidence-aware deterministic compression

evidence_survival_rate is an explicit replay metric for annotated high-value facts:

evidence_survival_rate = evidence_survived / evidence_total

Paper replay specs annotate evidence with stable IDs and simple locators such as method:0, metrics:0, or limitations:0, where the prefix is the fixture section and the suffix is the deterministic sentence index within that section. Agent trace replay specs annotate evidence with event locators such as event:active_task, event:constraints, or event:tool_sequence, which map to extracted operational fields. The validator resolves those annotations against the original fixture and the reconstructed replay state, then emits has_evidence, evidence_survival_rate, evidence_survived, and evidence_total in the in-memory replay rows and JSON artifacts.

The check remains deterministic and model-free: paper evidence uses bounded keyword overlap against the replayed section field, while agent evidence uses normalized exact matching against replayed operational events. No embeddings, LLM judging, external APIs, vector stores, or graph stores are involved.

Fixture Evidence annotations evidence_survived evidence_total evidence_survival_rate
PrefixGuard paper excerpt 2 1 2 0.5
coding_agent_trace 2 2 2 1.0

These fields extend the existing replay artifacts without renaming or removing legacy JSON fields, so downstream CI comparisons can continue to diff the older metrics while also tracking evidence preservation explicitly. Adaptive policy thresholds only interpret evidence_survival_rate when has_evidence is true.