Wiki · Backlog
Extract Inspect AI — reproducible agent evaluation and sandboxed evidence
Shelf
Backlog
Status
proposed

Proposal (rationale)

The current corpus contains deterministic model substitutes, structured-output validators, traces, and tests, but it does not read an evaluation system as the subject. That leaves an important gap: how a task definition, agent trajectory, sandbox, scorer, evidence log, aggregation, retry, and reproduction contract fit together.

Inspect AI is selected because its public source is explicitly organized around language-model evaluations and includes tool-using, multi-turn and sandboxed execution. The extraction asks what an evaluation result actually proves, what evidence survives the run, and where model-graded or human judgement enters a result that may otherwise look mechanical.

Decision

proposed — 2026-08-15. Captured as the evaluation-system child of WI-018, after WI-019 defines the track template.

Plan

Requirements

  • Pin repository SHA, licence, date, measured scale, and the exact evaluation packages and tests read.
  • Trace one agentic evaluation from dataset/sample construction through solver execution, model and tool calls, sandbox interaction, scoring, aggregation, and durable result log.
  • State which inputs and environment details are sufficient to reproduce a run and which external model or infrastructure behaviours remain outside the record.
  • Distinguish deterministic scoring, model-graded scoring, human review, and aggregation; state the claim boundary of each.
  • Trace failure, retry, cancellation, partial result, and resume semantics where present.
  • Map findings to gate, audit, evidence, bounded-loop, and sandbox candidates without editing the catalogue.

Impacts

  • One evaluation/optimization extraction and an index row under WI-019’s layout.
  • Candidate evidence for the distinction between execution traces, audit inputs, gates, scores, and conclusions; WI-028 decides catalogue consequences.

Approach

Choose one small evaluation that uses an agent loop and a sandbox, then follow its actual objects and stored artifacts end to end. Use built-in components rather than a third-party evaluation so the read boundary remains in one pinned repository. Treat the presence of many evaluations as an inventory claim to measure, not a proxy for framework quality.

Acceptance criteria / tests

  1. Snapshot and read boundary make every implementation and count claim reproducible.
  2. One evaluation is traced from sample input to stored result and aggregate score.
  3. The extraction states what the log can replay, audit, or merely describe, with failure paths included.
  4. Every scorer type inspected has an explicit claim boundary; a numeric result is not presented as truth.
  5. The sandbox trust boundary and external dependencies are traced or bounded as not found.
  6. Opinion is labelled; catalogue candidates are deferred to WI-028; gates and site links pass.

Out of scope

  • Running expensive model evaluations or comparing model performance.
  • Surveying the external evaluation collection as a second corpus.
  • Security certification of sandbox backends.
  • Catalogue, module, or CLI changes.

Execution

Not started.

Review

Not started.