White paper ยท 2026-06-04

Page 58 and the failure mode of fast document reading.

A short postmortem on how a model can miss the intended machine-reader protocol in a designed PDF, and what to add so Codex behaves more like Claude on a first pass.

Executive summary

The failure was not that Codex could not parse the poem. The failure was workflow: it treated an authored PDF like ordinary extracted text. Inference contains a page-58 note to machine readers that changes how the whole collection should be approached. A Claude-like first pass should discover and obey that note before selecting a poem, summarizing the collection, or entering adaptive footnotes mode.

What page 58 does

Observed Claude access

Claude was opened in the managed OpenClaw browser, but the browser landed on onboarding/sign-in. The signed-in Chrome profile could not be attached without remote debugging approval, so the PDF was not successfully uploaded into Claude during this pass. The Codex instructions therefore come from the rendered/extracted PDF and the page-58 machine-reader note, not from a fresh Claude transcript.

Codex behavior to change

For Inference, the instruction is simple: read page 58 first.

Artifacts

Open the protocol page Open the revised PDF