Skip to content

ScaffoldScope v0.3.1

Latest

Choose a tag to compare

@github-actions github-actions released this 16 Aug 02:41
· 4 commits to main since this release
v0.3.1
badecb2

ScaffoldScope 0.3.1 is an immutable research-instrument snapshot.

Initial public release from the current repository. Version 0.3.1 is used
because GitHub's immutable-release protection prevents reuse of a release tag
from the deleted predecessor repository.

Added

  • A readable, zero-runtime-dependency Python harness for paired coding-agent
    ablations across context policies, tool surfaces, and treatment instructions.
  • Four built-in context treatments: no compaction, reactive summarization,
    periodic summarization, and budget-aware selective retention.
  • Deterministic planning, hard experiment budgets, resumable trials, provider
    usage provenance, and intention-to-treat reporting.
  • Local and network-disabled Docker execution backends with recorded runtime,
    task-source, implementation, plugin, and image provenance.
  • SWE-bench import, full-matrix export, and immutable official-evaluation
    overlays without rewriting raw generation results.
  • Integrity-checked traces, patches, reports, and deterministic evidence bundles
    with redaction-aware context commitments.
  • A starter project, a zero-cost scripted demonstration, a versioned plugin API,
    operator documentation, and a statically rendered Next.js project website.

Analysis contract

  • Summary schema version 2 separates infrastructure-valid generation accounting
    from evaluator-valid outcome analysis, so pending external evaluations retain
    their token, cost, latency, and context-exposure records.
  • Inferential labels require a preregistered primary contrast, at least 20
    independent tasks, and at least 98% pair coverage. Scripted runs and panels
    below 10 tasks do not receive intervals.

Known limits

  • ScaffoldScope 0.3 is an alpha research instrument, not a security boundary or
    a guarantee that a benchmark is uncontaminated.
  • The built-in scripted demonstration checks the workflow; it is not evidence of
    model capability.
  • Imported SWE-bench generations remain outcome-pending until official evaluator
    results are ingested.

Read the documentation, verify downloads with SHA256SUMS, and report vulnerabilities through private vulnerability reporting.