docs(model-notes): fold the kc-model-audition bakeoff under model headings

[usjoh]

Saves two weeks of unsaved work and makes it reach the scoreboard.

What this was

A four-arm bakeoff from 2026-08-02 — sonnet-5, opus-5, opus-4.8 and the free nemotron-3-ultra-550b extracting knowledge from dense session transcripts, scored against a pre-registered key, then re-judged blind by a different-lab grader. It had been sitting uncommitted in the working tree since it was written.

Verified against the eval log before touching it: run kc-model-audition-20260803T001232Z-p41601, 18 attempts, outcomes matching the write-up exactly.

gpt-5.6-sol                 PASS x7          ← "7/7 judge tasks passed first try"
claude-opus-4.8             PASS x2          ← corpus 1 + corpus 2
claude-sonnet-5             PASS x2          ← corpus 1 + corpus 2
nemotron-3-ultra-550b:free  PASS x2          ← corpus 1 + corpus 2
claude-opus-5               PASS x1 + FAIL x4 ← corpus 1 passed; corpus 2 OpenRouter outage

The problem

The heading was a glob:

## Anthropic models via opencode (`openrouter/anthropic/claude-*`) — KC-extraction bakeoff

The notes matcher resolves model ids from headings, so it matched nothing. Three findings reached no scoreboard row:

• sonnet-5, opus-5 and opus-4.8 show zero measurable separation on this workload
• verbosity buys nothing — opus-5's 5131 words scored the same 11/11 as opus-4.8's 1660, at 3x length and ~2.4x wall-clock
• the free 550B nemotron ties both frontier arms once you stop substring-scoring

That is precisely the evidence you want in front of you when routing this kind of work.

The fold

Per-model results split under headings carrying the eval log's own slugs, following the convention the working sections already use. Judge evidence moves under gpt-5.6-sol, the model that earned it. Every method and scoring lesson carried across verbatim under a dated process heading — the fold script asserts none were dropped.

Also records the routing trap the outage leaves behind: those 4 FAIL rows drag opus-5 to 33% on the default scoreboard, while models --attributable excludes them and shows 100% over the one task it completed.

Measured effect

	before	after
dated notes reaching no scoreboard row	46 under 18 headings	44 under 18
Notes column for the 4 arms	empty	populated

Content is re-filed, not rewritten; per-model bullets use only numbers already in the original text. Full suite: 344 pass.

Precedent: 0fc0042 (fold cttc run section into the gpt-5.6-sol heading).

🤖 Generated with Claude Code