Checklist for a claim about machine-mind competency

Section 7 of the Tree G position paper, as a standalone page. Draft 1, 2026-10-30. A claim passes it before it is registered: as a Tree B pre-registration, as an entry under ../../spine/03-adversarial-registry/, or as a result this programme reports. The entry answers each row by number, in the entry itself. A row that cannot be answered is recorded as unanswered, and the claim goes on the registry's open-problems list. Tree A citations are to sections of ../01-a-ledger/writeup.md. Source details are in reading.md.

#	Test	Why	Tree A
1	Name the task family the competency is claimed over, and the convention for counting performance.	An advantage on one family of problems is offset on the rest (Wolpert and Macready 1997, p. 70, https://doi.org/10.1109/4235.585893), so a claim with no family has no content.	0.3 used one family; the structural finding is stated for that family only (4.1, 5).
2	Metric substitution: score the claim on at least one second metric that tracks what the system optimises, chosen before the run.	A metric the system does not optimise can show behaviour that is not there.	Delayed gratification was zero on total inversions in all 490 runs of the authors' code (3.5, 3.7).
3	Vary every implementation choice the claim should not depend on (scheduler, turn order, seed, compressor) and report the effect beside the claimed one.	A nuisance parameter can move the number more than the manipulation does.	Turn order changed bubble sort's score by 78%, more than the whole barrier effect (3.3, P5).
4	Matched-effort baseline: compare against the simplest known mechanism given the same effort (the same rule with one parameter changed, or a gradient-follower), in addition to chance.	A score against chance "is only as good as the null model" (Chis-Ciure and Levin 2025, p. 13, https://doi.org/10.1007/s11229-025-05319-6), and beating chance shows only that the system has some structure.	One rule with a slower clock on half the cells clustered as much as two algorithms (3.6); textbook bubble sort scored 0.079 with no barriers (3.5).
5	Random-substrate null: run the untrained or randomly initialised instance of the same architecture on the same task and metric, and subtract its competency.	Random structure already carries usable competency, and a seed is cheap.	Untrained Q-table and random reservoir (4.2, 4.4); label-only chimera controls (3.2).
6	One currency: state effort and competency in the same units, with every convention for what counts as effort fixed before the run.	Conventions alone can flip the sign of a residual.	Two rules gave the same untrained system residuals of opposite sign; no rule put a program and a trained network on one scale (4.4).
7	Pre-registration: freeze predictions, metrics, nulls, thresholds and each party's reading of each outcome, with a date and hash, before code. Set tolerances against the quantity being compared.	Without fixed readings the same result points both ways (Birch 2025, p. 21, https://doi.org/10.31234/osf.io/af7c9_v1).	A tolerance of 43,219 bits set against effort absorbed every residual, against competencies of at most 8,866 bits; commit order did not prove the sequence (2, 4.3).
8	Report medians and per-run spread with the mean, and say how many runs carry the effect.	A ratio statistic can be set by a few runs.	With barriers the authors' code had flat medians while 4 or 5 runs in 100 set the mean (3.7).
9	Adversarial replication: name a reviewer who expects failure before the run; publish the code; in a replication, also run the original authors' code unmodified.	A reimplementation can be the source of a result.	0.1c ran the authors' code and ruled that out (3.7); nobody outside the group has rerun Tree A (5).
10	No self-report as primary evidence, and no prompt that contains the hypothesis or mind vocabulary.	Models trained on human text are "incentivized and enabled to game our criteria", non-verbal dispositions included (Birch 2025, pp. 16-17).	None.
11	History-dependence: say whether the effect requires prior instances of the pattern, and test it.	It separates a history-free pattern from Sheldrake's habit reading (`../../Review.md` section 5).	None.
12	Manipulable variable: name a variable that switches the effect and show the switch.	To cause a supervenient property one must cause its base (Kim 1998, p. 42), so the handle is expected on the substrate.	The move probability switched clustering from 0.244 to 0.320 (3.6).
13	Say whether agency is attributed to the system or to the pattern; a claim of pattern agency states the measurement that would differ.	Agency in the pattern is the one claim the ledger cannot price (`../../Review.md` section 5).	None.

A positive result that passes every row is read as stated in draft-accounting.md section 4: the accounting left out a term, and finding the term is the next experiment.