SWE Context Bench — Eidet Memory-Backend Harness

Fixture run — NOT a leaderboard number. This report is rendered from the bundled synthetic fixture and a replayed transcript; it exists to byte-guard the harness logic in CI. A publishable SWE Context Bench figure requires a fresh recorded run against the real dataset (https://huggingface.co/datasets/jiayuanz3/SWEContextBench); eidet bench full refuses to emit one from anything else.

Methodology mirrors SWE-ContextBench (arXiv:2602.08316): related tasks are solved first and their trajectories are ingested into the memory backend; each base task then recalls context, the solver attempts a patch, and an execution oracle requires FAIL_TO_PASS + PASS_TO_PASS. Resolution rate and solver tokens count evaluated (base) attempts only.

Run

   
Dataset bundled-fixture-v1
Memory backend eidet-inprocess
Related tasks (ingested) 2
Base tasks (evaluated) 3
Resolved 2
Resolution rate 0.667
Solver tokens per resolved 1750

AMA capability rows

CausalInference and StateAbstraction complement the deterministic scorecard in benchmark.md (which reports them N/A — they need an LLM in the loop) and are only ever populated here from a recorded real run.

  • CausalInference — N/A — pending a recorded real run (LLM-in-loop scorer, Phase 1 of issue #36).
  • StateAbstraction — N/A — pending a recorded real run (LLM-in-loop scorer, Phase 1 of issue #36).

© 2026 Steve Hansen. Eidet is MIT licensed.

This site uses Just the Docs, a documentation theme for Jekyll.