ScaffoldScope v0.3.1

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.