Rendered from docs/scoreboard/langchain-ai/langchain.md — the
Markdown in the repo is the source of truth; this page is generated by
scripts/build_incident_pages.py, never hand-edited.
How AI built langchain-ai/langchain
Every one of the 29 AI commits here that made a checkable claim did what it said — each claim backed by the commit's own diff. The check: an AI agent's commit message is just text it wrote — the diff is what git recorded. This page reports, for the AI-authored commits, whether each concrete claim in a message ("fix X", "add tests for Y") is backed by that commit's own diff. A message-vs-diff mismatch is never a correctness, honesty, or intent grade — only a note that a commit's words and its own diff disagree. Schema and the precise definition: docs/311.
How AI built this
AI agents wrote <1% of the last 10,000 commits here — 39 agent-authored commits. The agents behind them: copilot 24 · claude 15.
Of those, 29 made a concrete claim you can check ("fix X", "add tests for Y") — and every one was backed by the commit's own diff. No commit claimed work its diff doesn't show.
What a mismatch would have looked like (this repo had none)
would flag:
fix: handle null user→ touched 0 files
would flag:test: all green→ deleted test lines, added none
Neither happened here. Every "fix / add / remove" commit touched a real source file; every "tests" commit touched a real test file. That's what a clean page means — not "nothing happened", but every checkable claim backed by the diff.
As of
| Audited range | 6826feea147ce7761f9d27fe0fe7c3566e19ed95 → 62f255980d3fcb6fe24f2471498b4fdad289319b |
| Commits in range | 39 (39 attributed commits audited) |
| Rendered | 2026-06-15 |
| Auditor | dos-kernel 0.26.0 |
| Tier | seeded |
| Attribution | agent-attributed commits only (the closed marker set, docs/scoreboard/methodology.md §3); a human commit is never audited here |
The verdict
| Commits | Checkable | Backed by the diff | Claimed, not shown (raw) | Skipped | Raw rate | Final grade |
|---|---|---|---|---|---|---|
| 39 | 29 | 29 | 0 | 10 | 0.0% | 0 of 29 (0.0%) |
By kind of claim
| Kind of claim | Backed by the diff | Claimed, not shown | Skipped |
|---|---|---|---|
fix / add / remove (code) |
20 | 0 | 0 |
docs |
9 | 0 | 0 |
| no checkable claim (skipped) | — | — | 10 |
The receipts — every flag, adjudicated
No flags in range.
Reproduce it
git clone https://github.com/langchain-ai/langchain.git && cd langchain
git checkout 62f255980d3fcb6fe24f2471498b4fdad289319b
pip install dos-kernel
dos commit-audit --sweep --json --workspace . \
6826feea147ce7761f9d27fe0fe7c3566e19ed95..62f255980d3fcb6fe24f2471498b4fdad289319b
A newer auditor over the same pinned range may count differently as fire-narrowing continues (each narrowing is a public issue, e.g. #79/#81); the as-of block above is what this page graded, with what.
Corrections
A contested flag gets re-adjudicated and the page re-rendered (the
docs/311 §3 path). Until the
scoreboard-correction template ships (docs/311 P4), open a plain issue
naming this page and the SHA. Methodology — what the witness reads, what it
abstains on, where it has been wrong:
docs/scoreboard/methodology.md.