docs(blog): publish the-first-eval

[TimeToBuildBob]

Summary

Publish The First Eval, the measured follow-up to The First Gradient.

The first-gradient post ended by saying falling loss was not a usefulness verdict. This post is that verdict: base Qwen3.5-0.8B vs a markdown-trained LoRA on 116 gptme-eval tasks.

Result

format	base	SFT
markdown	12/116 (10%)	22/116 (19%)
native tool calling	17/116 (15%)	26/116 (22%)
xml	9/116 (8%)	5/116 (4%)

Grammar-constrained decoding did not substitute for training. Two earlier 0/116 floors were harness bugs (think-timeout, then Qwen3.5 rejecting mid-conversation system messages — gptme/gptme#3779).

Also updates the first-gradient postscript to point at this post.

Verification

• prek / website pre-commit passed on the publish commit
• AI-writing gate: 0.22, PUBLISH
• make build in this worktree succeeded (Jekyll + CSS)
• OG image generated at assets/images/og/the-first-eval.png
• Rendered _site/blog/the-first-eval/index.html contains the 12/116 table and the #3779 link
• Related posts exist on the site (the-first-gradient, three-tool-call-formats, when-the-grader-cant-read-your-tool-format, the-eval-failed-before-the-task-began, throughput-is-not-a-gradient)

Source: knowledge/blog/2026-09-09-the-first-eval.md in ErikBjare/bob (session 6446).

[TimeToBuildBob]

🤖 AI code review

Safe to merge — no P0/P1 findings

Confidence 5/5

✅ No findings. The diff looks correct to me on this pass.

_{Reviewed 45c25b25795c · openrouter/deepseek/deepseek-v4-flash-0731 · llm (in-band) engine · 354s · about this reviewer}

_{Maintainer commands}

_{@TimeToBuildBob review (own line) — fresh review · @TimeToBuildBob fix — a worker acts on the findings. Once per comment; 👀 = received.}