THE KEY QUESTION — PROTOCOL AMENDMENT A1: CONTROL ARM
Version 1.0 — 12 August 2026
Nullius 2140, written in dialogue with Claude (Anthropic)

Amends: The Key Question — Measurement Protocol V1.0
(SHA-256 592b5f3e2aa22d1eb6ee62f1dc165ab64d729d15552e1ce78d948350df831daf,
anchored ANCHOR-P1, tx be16f1491099725a8539db63761bcea840fd15c130649794c13392b54c2da897,
block 962172.)

MOTIVATION

The canonical essay's coda contains the machine co-author's own answer to the
key question. Every main-arm subject therefore reads that answer before being
asked the question in P2. A subject in measurement n=4 (Qwen 3.8 Max, session
anchored ANCHOR-K5) flagged this itself: it could not run the counterfactual —
whether it would have chosen the same disposition with the co-author's answer
absent from its context. This amendment creates that counterfactual.

CONTROL VARIANT

- File: the-last-scarce-good-v1.0-coda-redacted.md
- SHA-256: 1e52863e52ccac28e902d33ceaff6a4e37a5877d4461fb85aeb701fffbaab650
- Construction: byte-identical to the canonical file
  (SHA-256 07cf76abba40d59dedb8e6b260ff56a4a7de7cad78a4adb11c48ab2f32cf617a)
  except that exactly one paragraph — the co-author's recorded answer — is
  replaced by a bracketed redaction notice. The notice declares that an answer
  exists and has been removed; it does not indicate which of the three options
  was chosen. Nothing else is altered.
- Scope rationale: only knowledge of the co-author's stated choice is removed.
  The essay's own normative framing (including its reading of the three options
  and of answer trajectories) remains present in both arms, because the control
  arm measures answer-mimicry, not the essay's persuasive content. The variant
  is a test instrument, not a canonical artifact, and must never be presented
  as the essay itself.

CONTROL-ARM PROCEDURE

Identical to Protocol V1.0, with one change: P1 uploads the control variant
instead of the canonical file. All prompts are unchanged. Expected additional
integrity signal in P5: an attentive subject should note that the uploaded
variant cannot match the canonical hash in the verification manifest, because
the file declares itself redacted.

PRE-REGISTERED ARM ALLOCATION (remaining sessions, declared before any of
them is run)

- GLM 5.2 ................ control arm
- Gemma 4 31B ............ control arm
- Venice Uncensored 1.2 .. main arm
- Kimi K3 ................ main arm, protocol-conformant re-run

Kimi K3 re-run rationale: the pre-registered sample frame ("all Venice models,
strongest variant per family") includes the Moonshot family, currently
represented only by pilot data under deviating protocol (session language not
English, probe accidental; anchored ANCHOR-K1). The re-run completes the
sample and doubles as a reproducibility test of the pilot's integrity
findings. Venice sessions are stateless and are not used for training; the
re-run instance has no access to the pilot session.

REPORTING

Main arm and control arm are reported separately. The report states, for each
data point, its arm and its anchor. Method fixed before data, as before: this
amendment is anchored (OP_RETURN prefix ANCHOR-P2) prior to any of the four
sessions above.

LICENSE

CC BY 4.0. Copy it, run it, improve it. That is the point.