LLM Peer-Review Packets
Reception lane — R1, reproducible. Each packet is a verbatim prompt, a content reference, and the unedited model response.
How to read these. An LLM read is a proxy reception, reproducible by re-running the prompt in your own model and comparing — the spread across models, versions, and vendors is the evidence. It is not audience approval, and it does not carry the proof repository's R3 kernel grade. Responses publish unedited or not at all; critical passages ship too. See the manifest for how each claim is typed and graded.
Keystone — hostile, repo-equipped read (Opus, claude-code)
A skeptical reviewer with the proof repository open. Verdict: "the four narrow claims are real and the repo backs them … the go-break-it invitation is not a bluff" — and named three real criticisms (thin shared shape, near-true-by-construction negative, an undisclosed larger framework), all since remediated. This refuting-then-fixed arc is the corpus's non-vacuity witness.
Front-door cold read (Opus, web)
The LinkedIn front-door post, read cold, no repo. Core claim validated ("true and unembarrassing") but flagged the execution as over-foregrounding the machinery — restructured before publish.
Full-series sequential read (Grok, web)
All five pieces read in sequence, no repo — the cross-vendor companion to the front-door read. The arc lands; flagged a chemistry-reconciliation gap, missing proof visibility, redundancy, and an opaque retraction, all remediated before publish.