Skip to content

Latest commit

 

History

History
44 lines (27 loc) · 2.47 KB

File metadata and controls

44 lines (27 loc) · 2.47 KB

Design

agent-bisect treats an agent run as an ordered stream of recorded activities. Its public claim is intentionally narrow: given a transcript, check the deterministic evidence inside that transcript and localize the first visible break.

Deterministic Envelope

The deterministic envelope is the part of a run the transcript records in a checkable form:

  • structured file edits and tool calls
  • parent-step links and file-target links
  • recorded command exits
  • recorded test or build output

The tool does not replay model reasoning, re-run the original agent, or infer hidden work from prose. It verifies the deterministic envelope and reports the gaps around it. That lets the non-deterministic core be tested indirectly: compare what the agent claimed or attempted against recorded, deterministic effects, without pretending the model itself can be replayed.

Gates

  • G1 validates normalized activity shape.
  • G2 checks edit causality against earlier full-content anchors in the same run.
  • G3 parses recorded test/build results.

Each gate returns PASS, FAIL, or NA. NA is a coverage result, not a hidden pass.

Localization

The localizer walks gate failures backward through parent-step links and conservative file-target edges. It reports:

  • the earliest visible breaking step
  • the gate that failed
  • downstream cascade steps
  • HIGH confidence for structured paths
  • LOW confidence when the path contains opaque, unmapped, unlinked, or heuristic evidence

The design is precision-first: report the smallest justified claim, and keep uncertainty visible.

Evaluation Ground Truth

Most real agent transcripts do not include reliable step-level labels for the first deterministic break. The accuracy report therefore uses injected faults over clean real runs: the mutation supplies known ground truth while the surrounding transcript remains realistic.

That is why ACCURACY.md is a controlled result, not a real-world attribution claim. It measures how precisely agent-bisect localizes known structured breaks once they are visible to G1/G2/G3. STUDY.md supplies the coverage denominator for how often real runs expose such breaks at all, and BENCHMARK.md shows the semantic boundary where the tool abstains.

Boundaries

This project is a transcript analysis tool. It is not a runtime, scheduler, state store, model replay system, or semantic judge. Its output is strongest when the transcript records structured actions and weakest when work is collapsed into opaque text.