e23182d873
CLAUDE.md claims to be the file to trust on what is currently true, which makes staleness in it worse than staleness anywhere else. Several sections were describing a router that no longer exists. The proficiency section was the worst of it. It documented all 27 category/tier decisions returning a single Pareto-dominant model, and spent several paragraphs explaining why that was correct rather than a bug. It IS still a legitimate outcome and the explanation was worth keeping, but it is no longer what happens: sweeping 9 categories x 3 tiers now returns 5 distinct winners at both 50k and 120k of context. Rewritten around what a reader actually needs -- what each outcome means as a diagnostic, and which two corrections widened it (cost priced per request, tier no longer inferred from price), so the section survives the next time the numbers move. "Coding does not discriminate these models" was true of the benchmark and is now false of the data: coding_general spans 0.86-1.00, because feedback.py folded 29 samples of real traffic into glm-5.2-fast and crossed self_eval_min_samples doing it. That is the intended shape of the system -- benchmark sets a floor, real traffic refines it -- so it is now stated that way, with the per-category spread tabulated. coding_refactor and debugging are still flat at 1.00 and that is now an open question rather than a boast. Also stale: 207 tests across 11 files (264 across 14); "the lever is `weights` in config.yaml", a block that no longer exists; an open question about tier-1 composites being compressed by min-max normalization, which cannot be asked any more because there is no composite; an open question asking for a minimum viable eval set, which has existed for some time (23 tasks, four scoring kinds, verified by counting them rather than trusting the number). De-personalised the same way as the README: the author's GPU is out of the opening line, private paths are gone from the outcome-attribution finding (the evidence survives -- a dependency's source directory inside the project venv outranked the project being edited), and a note up front says every measurement here came from one deployment against one account and should be re-run before being trusted. The systemd units were the concrete version of the same problem. All three hardcoded %h/Sources/6krrt, so anyone else copying them per the documented install would get status=200/CHDIR and no useful message. They now say %h/llm-router and the install step substitutes the real path with sed. Verified the substitution reproduces the currently-installed unit byte for byte, so the running service is unaffected; only the templates changed. Recorded while checking: deploy/ has five units and one Ollama drop-in, not the three the docs claimed. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WSkcSD2Jzkxo1Kw27ecfXJ
656 B
656 B