AXL / RESEARCH LOG

AXL Research Log

A per-commit record of the dual-agent research iteration, generated deterministically from the git history at HEAD 91dbceb on 2026-04-21T17:38:16+00:00.

How this was generated

This page is the output of tools/build-research-log.py v1.0, a deterministic script that walks the repo's git history and classifies each commit by role using regex rules that match this repo's authoring conventions. It is not a hand-curated narrative. Running the same script against the same git HEAD produces the same output.

Assumptions this property depends on:

1. Commit message discipline. The DAG of review-and-response edges is reconstructed from explicit SHA references in commit messages (patterns like "Codex sha review", "Codex sha follow-up", "Codex review of sha"). If future commits stop naming the SHA they respond to, the DAG loses edges. This is a commit-hygiene constraint, not a script bug.
2. Repo-scoped regex classifier. Role classification depends on this repo's authoring conventions: codex r\d+, Gap \d+, ship:, bench:, spec:, docs:, RESULT. Applying this script to another repo with different conventions produces meaningless classifications.
3. Metric extraction is best-effort. Numeric transitions in commit bodies are captured when they match known patterns (NN.NN% -> NN.NN%, NNN tests pass, Δrecall ±NN.NN). Claims stated in other forms may not surface as structured metrics. The subject and body text are always preserved verbatim in the JSON so human readers can verify.

Known classifier limitations on this specific history:

• The R4.1 sub-finding (referenced in the R4 commit body 099dbe6) does not appear as its own round-labeled entry because it does not have its own commit. Extraction picks the first round match in a commit, so R4 wins. To surface R4.1 as an independent entry, a future sub-finding would need its own commit with R4.1 in the subject.
• Body-level mentions of round labels occasionally produce false-positive round assignments on commits that reference a round in narrative explanation without BEING that round. Verifiable by reading the subject in the entry below.
• The long-form phrasing "round 1" in commit 0a5cad4 is caught by a dedicated pattern; other long-form phrasings of later rounds (if any) may not be caught.

Inputs: git log --reverse --format=... against this repo. Outputs: research-log.json (machine-readable, complete) and this HTML page.

Summary

Total commits: 65. Response edges (commits that name a target SHA): 6. Commits with a formal review-round label: 7.

Commit log

Verify this log yourself

The deterministic property ("same script, same HEAD, same output") is only meaningful if you can run the script. Both the script and the machine-readable output are published here:

• Script source: /research-log/build-research-log.py (Python 3 stdlib only, no network, no LLM calls)
• Machine-readable log: /research-log/research-log.json (the full 65-commit dataset with subject, body, role, round, response edges, metrics)

To verify on a clone of the research repository:

python3 tools/build-research-log.py --format json | diff - research-log.json
python3 tools/build-research-log.py --summary
# expected: commit_count=65, response_edges=6, round_entries=7

Any deviation from the expected summary numbers on the same HEAD is a bug; running the script against a different HEAD produces a different log, which is the intended behavior (the log is a function of history).