validation/code_repair/PREREGISTRATION.md

Code-repair experiment: preregistration (written before any model output was produced)

Date: 2026-10-21. Task list frozen in tasks.csv (sha256 below) before prompts were built or models run.

Question

Does API history from PyPI API Drift help a model migrate code to a newer library release more than ordinary documentation of that release does?

Tasks

150 VersionProbe pairs (v1.2) never used in any audit, the verified split or the classification baseline; change released 2023 or later; at most one task per library and change type. Strata: 60 removals that the dataset records as "still available as X", 45 renamed parameters, 45 removed parameters. Each task gives a snippet written for the old release and asks for a version that works with the new release. Tasks whose original snippet turns out to be valid in the new release at runtime are excluded from scoring (reported).

Conditions (same model, same instructions, items shuffled, one task at a time)

A. Closed book. B. Documentation: an API reference page for the imported module(s) of the new release, generated with griffe from the release wheel (independent of this dataset). C. PyPI API Drift: the dataset's API history for the snippet's names and keywords, the recorded change between the two releases, and new signatures.

Outcomes

1. Runtime validity (primary, automatic): in an isolated environment with only the new release installed, every library name resolves and every keyword binds (inspect.signature(...).bind_partial); placeholder values are not executed. The candidate must reference the library at least once.
2. Intent preserved: a separate model judge, blind to the condition, sees the original snippet, the target version and the candidate, and answers whether the candidate is a faithful migration (same purpose, uses the replacement API) rather than deleting or neutering the library use.
3. Success = valid AND intent preserved. Report per condition with Wilson 95% intervals, by stratum, and paired McNemar tests C vs A and C vs B.

Hypotheses

H1: C > A on success. H2: C > B on success. No result will be dropped; all prompts, answers, verdicts and scripts are published. 9b295971852886b8acc2cc61b183cb2356ea0a684115ccc9584dcd3ac92a069e repair/tasks.csv