Point the Agent at the Exact File, Revision and Failing Case

Your eval reports 0.71 on a benchmark and the paper reports 0.78. You already suspect the answer normalization in eval/metrics.py, and you have one case where your score and the paper's reference implementation disagree. Say that, precisely, instead of asking the agent to rediscover both.

Claude Code's file references do this directly:

The score on @tests/fixtures/case_017.json differs from the reference implementation.
Compare how @eval/metrics.py normalizes answers with the paper's description.

The documentation says an @ file reference includes the full content of the file, a directory reference shows a listing rather than contents, and a file reference also adds any CLAUDE.md in that file's directory and its parents. That last point matters in a monorepo: the agent may pick up instructions you did not mean to give it.

Then check what the session actually received:

git rev-parse --short HEAD
git status --short eval/metrics.py

If git status lists the file, the agent is reading your uncommitted edits, not the commit you will cite in the write-up. Ask the agent to quote the signature of the function it is reasoning about and compare it with the file. A screenshot of an error or a link to the paper can help too, but neither shows what the agent inspected until you check its evidence.

This habit, precise provenance, is what makes a reproduction report credible to a research team: this file, at this revision, on this input, gave this number. The book's context chapter covers why a small, exact context often beats a large, vague one.

The check: the quoted signature matches the file at the revision you recorded, and the case you named produces the number you reported.