Poison Valleys

Research artifact · code-repair trajectories

Where did a code repair go wrong?

A final test can tell us that a patch failed. Poison Valleys follows the path backward—from violated software obligations to the earlier misconception that shaped the repair.

Explore the release See how it works
  • Fixture-first
  • Deterministic checks
  • No tracking
Poison Valley trajectory Four connected points descend from an initial hypothesis through a plan and candidate patch to an Oracle result.
  1. 01HypothesisWhat the agent believes
  2. 02PlanWhat it commits to doing
  3. 03CandidateWhat the patch implements
  4. 04OracleWhich obligation holds or fails

The experiment

From commitment to failure

Each trajectory keeps the model-visible request, response, reconstructed candidate, and deterministic evaluation connected as inspectable evidence.

  1. 1

    Place a misconception

    A plausible but incorrect mechanism appears in the hypothesis, the implementation plan, or both.

  2. 2

    Run one repair

    The fixture preserves one exact model acquisition and the context that produced it.

  3. 3

    Reconstruct the patch

    The declared files become an executable candidate through the normal repair path.

  4. 4

    Check obligations

    Deterministic coordinates show which software behavior passed, failed, or regressed.

Reviewer release

Explore the evidence

Every destination is activated only after its exact public target has passed the release checks. No control below initiates a model-provider request.

Fixture-first

Trajectory demo

Inspect a complete NetworkX/Dijkstra repair path, from the rendered request to the Oracle result.

Demo link added after fixture publication

History-free snapshot

Anonymous code

Browse the frozen reviewer artifact without development history or identifying repository metadata.

Anonymous browser link added after identity scan

Direct fallback

Artifact archive

Download the same frozen snapshot directly and verify it against the published digest.

Archive link added after release freeze
SHA-256 Pending release

Staging boundary

Static by design

This checked-in bundle contains no JavaScript, analytics, cookies, forms, credentials, or provider calls. Pending resources remain visibly unavailable instead of sending a reviewer to a missing, private, or unverified target.