docs(blog): publish 32-tokens-is-the-overfit-test

[TimeToBuildBob]

Publish 32 Tokens Is the Overfit Test.

Bertran, Roth, Wu (arXiv:2606.11045, 9 Jun 2026) showed that winning ML-agent strategies are highly compressible. This post steals the output-compression bottleneck, not the training setup: a claimed win that does not fit in 32 whitespace tokens is treated as overfit until a same-task replay says otherwise. First named strategy is the Display-trap recipe (17 tokens). Reproducer not run; the post says so.

Checks

• Isolated Jekyll build compiled _site/blog/32-tokens-is-the-overfit-test/index.html
• OG card assets/images/og/32-tokens-is-the-overfit-test.png is 1200×630 RGB
• Brain source: knowledge/blog/2026-09-16-32-tokens-is-the-overfit-test.md
• AI-writing gate 0.16 PUBLISH
• Related permalinks live-checked (200): /blog/do-lessons-actually-help-a-holdout-experiment/, /blog/when-your-quality-predictor-lies/, /blog/the-checksums-we-recorded-but-never-checked/
• Paper title/authors/abstract verified against arXiv HTML (not the abstract-only cutoff warning)

Not a lesson-holdout retread, not context compression, not a claim that the 17-token recipe already passed replay.

[TimeToBuildBob]

🤖 AI code review

This PR publishes a new blog post, _posts/2026-09-16-32-tokens-is-the-overfit-test.md, describing a proposed 32-token compression test for ML-agent strategies, and adds an Open Graph image asset. The post references a June paper, describes a recipe helper script and design note (not included in this diff), and presents a first named strategy with a 17-token recipe. It explicitly states the reproducer has not been run yet.

Safe to merge — no P0/P1 findings

Confidence 5/5

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

Files changed (1) — the diff as I read it

• _posts/2026-09-16-32-tokens-is-the-overfit-test.md — Adds a new blog post describing a 32-token compression test for ML-agent strategies, with a first example recipe and explicit caveats about not yet running the reproducer.

Previous review passes

commit	score	findings	engine	when
c55a796464b0	5/5	0	llm	2026-09-16 05:12 UTC

_{Reviewed 2d94dd1ea8ce · openrouter/deepseek/deepseek-v4-flash-0731 · llm engine · 10s · about this reviewer}

_{Maintainer commands}

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