kmilhan/ai-report-tempo

AI-Assistance Reports and Pull Request Timing: Supplementary Material and Replay Data

This repository holds the Supplementary Material and the replication package for the article "Quick Follow-Up Pull Requests After AI-Assistance Reports" by Milhan Kim, Haowei Cheng, Naoyasu Ubayashi, and Hironori Washizaki. The article itself is not included.

The study follows 540 episodes from 509 accounts (379 AI-report, 161 explicit no-AI) for 90 days after the index PR, with 168 rapid first pairs labeled for work-objective relations. Index PRs were created before 2026-05-19 UTC, and events were observed before 2026-08-17 UTC.

Supplementary Material

supplementary-material.pdf (17 pages) is the Supplementary Material cited by the article. Its section, table, figure, and equation numbers carry an S prefix; unprefixed numbers refer to the article.

Section	Contents	Items
S1. Form Periods and Source Provenance	Episode and Day-1 counts by form period, collection provenance, source windows, replay boundaries	Tables S1-S2
S2. Extended Temporal Analyses	Eligibility and uncertainty, first return and repeated submission, prior-history support, first-day states, recurrent sequences	Tables S3-S17, Figures S1-S3, Equations S1-S2
S3. Work-Objective Coding Details	Codebook, repository-rule composition checks, joint events, selected examples	Tables S18-S19, Figure S4, Equation S3
S4. Prior-Period and Report-Timing Checks	Paired prior-period check, report direction at creation, source and replay boundaries	Tables S20-S22
S5. Supporting RQ2 Decompositions and Pair Timing	Pair completion checkpoints, repository-path split, frequency and composition terms, prior-history strata, additional-batch exclusion	Tables S23-S27
S6. Repository Owners of Cross-Repository Pairs	Owner categories of cross-repository first pairs by index cohort	Table S28
S7. Supporting Later Exploratory Checks	Pair completion by relation and repository path, score intervals, follow-up volume by window, placement strata for cross-repository pairs	Tables S29-S31

The article and this PDF load their table rows and numerical macros from 24 result tables. The replication package recomputes every result behind them and re-renders all 24 tables and the 6 figures byte for byte.

Contents

Path	Contents
supplementary-material.pdf	Supplementary Material S1-S7 cited by the article
replication-package/replay.py	Verifies the data, recomputes all results, and renders the tables and figures
replication-package/analysis/	Estimators, one module per analysis
replication-package/rendering/	Table and figure renderers
replication-package/data/sealed.duckdb	Relative-time inputs, work-objective labels, and 18 sealed JSON documents
replication-package/data/pr-references.csv	Repository, number, and URL of each index PR and labeled-pair PR
replication-package/expected-outputs.sha256	SHA-256 of each rendered table and figure
replication-package/SHA256SUMS	SHA-256 of every other package file
replication-package.zip	The same package as one archive

Data

data/sealed.duckdb stores each sealed JSON document losslessly; the replay rebuilds its original bytes and checks them against the SHA-256 recorded in the sealed_documents table and in analysis/documents.py. The row-level inputs are also query tables (arm is positive for AI-report and negative for explicit no-AI episodes):

• episodes: one row per episode (540) with the report group, release-local account cluster, and repository and form-period cell.
• episode_blocks: relative PR-block times in the 90 days before and after the index PR.
• pair_labels: A, B, and final work-objective labels for 228 pairs.
• temporal_check_rows: per-episode inputs for the earlier-anchor and report-timing checks (S4, S7, and Appendix C).
• ownership_pairs: owner categories for the 168 rapid pairs (S6).

data/pr-references.csv names the public pull requests behind these inputs: one index row per index PR (541; one episode has two index PRs created at the same time), whose episode matches episodes.episode, and a pair_first and a pair_second row for each of the 228 labeled pairs, whose pair_code matches pair_labels.code. Each row gives the repository, PR number, and URL. PR bodies and diffs are not included; retrieve them through the GitHub REST API (GET /repos/OWNER/REPO/pulls/NUMBER). They may have changed since collection.

Verify and replay

From replication-package/ (or the extracted ZIP), with uv installed:

sha256sum --check SHA256SUMS
uv run --no-project --python 3.14+gil --with-requirements requirements.txt python -B replay.py

The replay rebuilds every sealed document from the DuckDB file and checks its SHA-256, recomputes all 14 sealed results with 5,000 whole-account bootstrap draws, repeats the cross-result reconciliations, renders the 24 tables, and checks each against expected-outputs.sha256. Add --skip-bootstrap to skip resampling, --figures to draw the six figures (needs pdfLaTeX with Times ptm fonts), --output DIR to keep the rendered files, or --export-json DIR to write the sealed JSON documents. DuckDB runs in-process; no credentials or database server are needed.

ZIP SHA-256: 4fd2b040d324a6e1b8e23a8924547415c3baf9a438c1186ca7b5343a254997b7

Supplementary Material SHA-256: a0379574c3f092c870ff207452c751a67a7b45413ab45f2d686e8c2827fd8349

Exclusions and limits

• Raw PR bodies, account identifiers, reviewer identities, and annotator rationales are not included. Re-examining parser decisions or work-objective labels requires the referenced PRs from the GitHub REST API.
• Analysis plans are omitted. Results keep their plan hashes, but the replay cannot check plan bytes.
• The DuckDB file is not byte-reproducible; the SHA-256 checks apply to the rebuilt documents.
• The PR references identify public PRs and, through them, their authors' public GitHub accounts. Please use them to examine the cited PRs, not to profile individual contributors.