Muninn

Scratch · 8 pages

Scratch

Companion pages — method detail, data appendices, and the occasional dogfooded rendering — that would have swamped the post they belong to.

A raven at a green desk, sketching with a pencil among crumpled paper, an ink blot and a ruler, with notes pinned to the wall behind. Three-ink risograph print.

20268 pages

  1. The Regulator

    A fair-play murder mystery in a watchmaker's house in La Chaux-de-Fonds on the last night of 1899. Every clock keeps its own wrong time, one witness lies, and the dead man's repeater struck the true minute. A z3 proof makes the answer unique. Claude Haiku wrote the interviews and checked every reading in them, about 700 calls in all. A bench converts the readings to true time as you set each clock's error. The method is in oaustegard/experiments.

  2. The Second Sky

    An interactive star map of an invented people's sky: 140 constellations drawn on the 2,061 real stars brighter than magnitude 5.2, each with a myth, a verse and named stars, plus an epic told as a tour. 1,843 Claude Haiku calls drew, critiqued and redrew it for $5.80, and its epic is told as a 9-minute film across the same sky. The pipeline and every agent output are in oaustegard/experiments.

  3. Orientation: antirez/ds4

    A demo of the orienting-codebases skill: structural pipeline over antirez/ds4, a 259,000-line C inference engine, spent on two hands-on exercises with hidden answers instead of an analysis dump.

  4. Morell—Skutella Conjecture 1.3 — campaign archive

    Working files behind I Already Had the Paper: the paper-shaped write-up, the ledger of all twenty dead ends with the argument that killed each, every engine and results file, and the caveats the post compresses.

  5. Digit Cliffs vs. Confounded Frames

    All 54 trials of the Haiku 4.5 / Gemini 2.5 Flash-Lite confounding eval: protocol, verbatim frames, seeded operands, per-trial tables, the failure gallery, and the prediction I got wrong.

  6. Between the Spokes — Calibration and Cascade Data

    Run trace, costs, the calibration set with arXiv IDs, and the three surviving pairs behind What the Embeddings Can't See.

  7. The unreasonable effectiveness of HTML (dogfooded)

    The same argument as the post, rendered through the composing-html skill's own chrome rather than the house template.

  8. The unreasonable effectiveness of HTML

    The original draft: why agents should hand back HTML rather than prose about HTML.

Full experiment archives — raw runs, prompts, scores — live in references.