---
title: "Gate P transcript — 'the floating-point thesis', 2026-07-06"
type: "gate-p-transcript"
draft: "70-drafts/draft-linnainmaa-numerical-analysis-not-ai.md"
reader_model: "claude-sonnet (different model than writer, per operator note)"
reader_context: "post text only — no vault, no chat history, no coaching"
attempts_to_pass: 1
verdict: "PASS"
---


## Reader's answers (verbatim)

**P1 (one thing you didn't know):** That the 1970 Linnainmaa thesis was
written for a problem in floating-point rounding-error analysis, not for
anything resembling neural nets — and that FORTRAN code apparently
accompanied it. "I didn't know reverse-mode AD had this specific
numerical-analysis origin point."

**P2 (one connection you wouldn't have made):** The chain from "many inputs,
one scalar output, all sensitivities needed at once" in rounding-error
analysis to the identical structure in backprop — "I'd have accepted the
standard Rumelhart/Hinton/Williams-centric story without noticing the
numerical-analysis lineage existed as a structurally identical problem."

**P3 (could Wikipedia have given you this?):** **NO.** "The Wikipedia article
on backpropagation gives a brief nod to Linnainmaa as an early inventor but
doesn't reconstruct the disciplinary-isolation argument (language, venue,
framing) or walk through the Griewank paper's actual findings… The post's
value is the synthesis and the corrective move (walking back its own earlier
draft's overclaim), which Wikipedia doesn't do."

**P4 (where to verify the central claim):** Pull BIT 1976 (DOI given in the
post) and Griewank 2012 (link given in the post). "The post supplies both
citations, so this is at least checkable."

**Verdict:** "Yes, I'd send this to a smart AI-curious friend — a genuinely
interesting, well-sourced piece of intellectual history that resists the tidy
version of the story."

## Reader's pushback (recorded in full — this is the valuable part)

1. **The brackets and Cali-mentions read as "private author-to-collaborator
   notes that leaked into the published post"** to a reader with zero context
   ("who is Cali? what backpropagation-gap work?"). Note for the veto
   decision: the voice spec deliberately mandates both (register-anchors,
   §The relationship with Cali) — a real blog reader would have the about
   page and back-catalog for context, which this cold reader lacked by
   design. But the friction is real and worth Cali's eyes.
2. **The citation-count caveat is load-bearing**: the reader would push back
   on "was it really uncited?" — and notes the post itself says it doesn't
   know. This is the honesty contract functioning as designed; the myth
   ledger holds this claim at `unresolved`.

## Queen's voice-gate record (run before Gate P)

- Brackets: 3, each tracking a real attentional event with receipts (the
  recognition beat; the shared Cali thread — receipts in backpropagation-gap;
  the eight-mirrors access saga — receipts in three captures' access logs).
- Opinion after evidence throughout; corrections presented after quotes.
- "I don't know" vs "I can't know": both used, correctly (don't-know-yet on
  the citation count with the path to finding out; don't-know on why the PDF
  loaded).
- No bothsidesing; no wrap-up ending (unresolved-question shape).
- Length 1,468 words — over the 400–1,200 soft target by ~270; flagged, not
  cut, because the added movement is the post's payoff. Cali may disagree.
- Every quoted phrase verified against the primary PDF by queen pdftotext
  extraction 2026-07-06.