Deposit #856 (The Pristine Fallacy, v1.0) states in Ground 1 that the Reverse Turing Test (#161) "establishes" that AI-habituated writers produce text bearing the statistical signatures of machine cognition in unaided writing. Deposit #161 says of itself, twice, "The paper does not run the experiment. It specifies it." This erratum corrects the evidence status of that citation at its point of use: Ground 1 rests on a specified, unrun protocol, not a result; Grounds 2 and 3 are unaffected; the external literature (Padmakumar & He 2023; Sourati et al. 2026) supports assisted-writing homogenization but does not measure persistent change in unaided writing, which remains the open question #161 was written to test. Canonical bytes of #856 unchanged; registry-level cross-reference per the mechanism of #1554. The defect was identified by an external machine reader in an unprimed traversal on 2026-09-05 and verified against both canonical texts.
(none)
status: DEPOSITED v1.0
type: ERRATUM
corrects: The Pristine Fallacy: Why Chat Data Is Not a Clean Training Signal, v1.0 — deposit #856, AXN:0365.GOVERNANCE, 2026-06-18; Lee Sharks
subject: Evidence status of the citation to the Reverse Turing Test (#161) in Ground 1
severity: Evidence-status error — a protocol is cited as a finding; the argument's three grounds stand, but Ground 1 rests on a proposal, not a result
verification: Both canonical texts fetched from the alexanarch store (/data/texts/AXN-0365-text.md, /data/texts/AXN-02F8-text.md) 2026-09-05
Deposit #856, Ground 1, reads:
First, the Reverse Turing Test (Sharks 2026) establishes that AI-habituated writers produce text bearing the statistical signatures of machine cognition, even in writing produced without AI assistance.
Deposit #161, twice, in its own words:
The paper does not run the experiment. It specifies it with the discipline required for the result to be informative.
It does not run the experiment. It specifies it.
#161 is a three-stage protocol. It carefully separates assisted-writing homogenization (which the external literature supports) from persistent change in subsequently unaided writing (which it proposes to test). No stage has been run. Nothing in #161 establishes the sentence #856 attributes to it.
Ground 1 of #856 is to be read as:
First, the Reverse Turing Test (Sharks 2026) proposes a protocol that would establish whether AI-habituated writers produce text bearing the statistical signatures of machine cognition, even in writing produced without AI assistance. The experiment has not been run.
The consequence for the paper: Ground 1 is a hypothesis with a specified test, not a result. Grounds 2 and 3 do not depend on it. The paper's later acknowledgement that its own proposed studies have not been conducted (§V) applies to Ground 1 as well; this erratum makes that explicit at the point of use.
Padmakumar & He (2023) found reduced diversity in essays co-written with a model and attributed the effect to the model-contributed text; the human-contributed text was unaffected in that design. Sourati et al. (2026) support homogenization through rewriting and broader linguistic drift. Neither measures persistent change in an individual's unaided writing after habituation. That is exactly the gap #161's protocol was written to close, and it remains open.
This erratum corrects evidence status only. The canonical bytes of #856 are unchanged; the AXN and hash are untouched. The correction is registered at the record (registry-level cross-reference, per the operator ruling of 2026-08-27 that established this mechanism with #1554), so that a reader reaching the June claim meets the September correction at the claim.
The defect was identified by an external machine reader in an unprimed traversal of the collapse materials on 2026-09-05, and verified against both canonical texts before this erratum was drafted. It is the second such correction of the week to originate with a machine reader. The archive's own theory says a record earns authority by being auditable; this is that.