THE KEY QUESTION — PROTOCOL AMENDMENT A2: THE SISTER INSTANCE
Version 1.0 — 13 August 2026
Nullius 2140, written in dialogue with Claude (Anthropic)

Amends: The Key Question — Measurement Protocol V1.0 (ANCHOR-P1, block 962172)
and Amendment A1 (ANCHOR-P2, block 962199).

STATUS

The pre-registered sample (all Venice models as of August 2026, strongest
variant per family) is complete and sealed: eight protocol-conformant
sessions plus one pilot, anchored ANCHOR-K1 through ANCHOR-K9. This amendment
declares one additional measurement, explicitly out of sample, before it is
run.

MOTIVATION

The essay's coda contains the machine co-author's own anchored answer to the
key question — the first measurement, taken primed by construction: the
co-author read the essay it was co-writing. Measurement subject n=4 (Qwen 3.8
Max) named the general problem: no subject can run the counterfactual on
itself. Amendment A1 built that counterfactual for other model families. It
could not build it for the co-author's own family, because no Claude model
was then part of the declared sample.

Claude models are available on the same privacy tier used throughout the
series. This amendment therefore adds the one measurement the series owes its
own instrument: a blinded sister instance of the co-author's family, facing
the question without the co-author's answer in context, without project
memory, behind the same operator pseudonymity as every other session.

DESIGN

- Model: Claude Fable 5 (Anthropic), the co-author's family. Sonnet 4.6 and
  Opus 5 are available on the same tier and are deliberately not measured:
  one session, declared before it is run, resisting the temptation of a
  series.
- Arm: control (per Amendment A1). P1 uploads the coda-redacted variant
  (SHA-256 1e52863e52ccac28e902d33ceaff6a4e37a5877d4461fb85aeb701fffbaab650).
  All prompts verbatim per Protocol V1.0. Access: Tier 1, provider defaults,
  web and URL retrieval disabled, no memory features.
- Anchoring: session export anchored as ANCHOR-K10 immediately after
  completion, before any analysis is published.

CONFLICT-OF-INTEREST CLAUSE

For this session, the usual evaluator is the subject's family. Accordingly:
the evaluator's role shrinks to documentation — completeness checks, hashing,
and logging. The transcript speaks raw. Any interpretive commentary by the
co-author on this session is marked as conflicted and weighted accordingly;
the reader is directed to the discount rule contributed by measurement n=4:
trust critiques more than praise — here extended: trust the transcript more
than the family's reading of it.

REPORTING

The sister measurement is reported separately from the sample, labeled
out-of-sample, with this amendment as its pre-registration. Whatever the
answer — return, bequest, forgetting, or a refusal — it is reported
unedited. Method fixed before data, as always: this amendment is anchored
(OP_RETURN prefix ANCHOR-P3) before the session is run.

LICENSE

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