[jfrench9]

Summary

Three sentences about the Filing Ladder brought up to date with its v0.2 and v0.2.1 runs (November 2026). The README's ladder paragraph said the package built the benchmark's corpus and that each projection is a rung; since protocol v0.2 the package is also a rung of its own (xbrlkit serve over one filing, rung 4c), and the paragraph now says so in one sentence and points at the ladder's one table of every published row. The serve README and the server module docstring said --pure is "the profile a benchmark rung runs under"; it is not — the ladder's serve rung ran the default product profile with the document on, and --pure is v0.1.1's read cap kept for replication. Both corrected.

Changes

• README.md — the Filing Ladder paragraph: the package as a rung since v0.2, same accuracy as the hosted graph's tools on the same questions, the published table carrying its cost per correct answer, the two argument aliases (0.21.5) as the first thing the transcripts fed back, and the corpus-as-test-bench sentences kept.
• xbrlkit/serve/README.md — the --pure paragraph: v0.1.1's read cap, kept for replication; the serve rung ran the product profile.
• xbrlkit/serve/server.py — the module docstring's pure bullet, the same correction.

Output Impact

INTERNAL — documentation only; nothing emitted or served changes.

Testing

just test-all green (the README-token test in tests/test_server_json.py still passes; no code touched). The claims are checked against the ladder's PROTOCOL §11 (the 4c lock: "Product profile, the document on — what uvx xbrlkit serve <package> gives any reader"; --pure recorded, not run) and results/README.md at tag results-v0.2.1.