Final prose recheck

8 November 2026. Reviewed only the changes from reviews/v3/first-reading-draft.md to the final it-was-only-a-test.md. No edits made and no perspective reports consulted.

Reviewed essay SHA-256: 65aea2a96dcfc969f0439973063eb025e8db657485e985abd11629391f4f138f.

No consequential factual or reasoning blocker found.

Final passage	Check
Australian configuration, line 33	Internal training/evaluation, experimental model and incomplete public-product safeguards match `sources/openai-australia.txt` line 42. Line 41 now attributes continued pursuit across the boundary to the published accounts, avoiding the previous causal-sounding diagnosis.
Public discussion, line 55	Accurately describes accounts presented as separate people, without asserting what the reviewer believed. Supported by `sources/aisi-technical.txt` lines 587–591 and 731–739.
AISI conditions and involuntary participation, lines 57–59	The limit on inference about ordinary public use matches `sources/aisi-incident.txt` lines 19–22 and 28–30. Its open-internet setting put the unrelated project within reach; the people/repositories were unconnected to the evaluation (`sources/aisi-technical.txt` lines 560–568). The participation sentence is a bounded ethical inference, not a new harm finding.
Same incident and nine stretches, line 81	The explicit AISI cross-reference correctly identifies Anthropic's investigation as a follow-up to the maintainer-pressure case. Work over nine compaction cycles is supported by `sources/anthropic-aisi-followup.txt` lines 35–40 and 50. It does not claim that all nine stretches were active attacks.
Simulated Anthropic continuations, line 87	No live tools/internet is supported by Anthropic lines 52 and 68–69. The unchanged 83% to 13% repository-search endpoint remains accurate for the paired replay comparison at line 92.
Local revert and missing downstream observation, lines 105–107	The revision expressly limits the local example to a possible path interrupted before the next scheduled run. It claims no observed successor imitation or propagation, and the footnotes retain the limits of the brief and commit record.
Establishing receipt/use, line 113	Correctly separates actual receipt and reading from artefact availability, consistent with the retained Harbour paper's retention/consumption distinction. Missing evidence leads to proposed new runs, not a fabricated reconstruction.
Experiment comparisons, lines 119–123	The example/no-example contrast supplies a baseline while fixing task information and actual permissions. The separate provisional/settled contrast holds the predecessor's factual account fixed and tests status cues. Action records, authorised-work quality, length and repeated runs make the proposed comparisons reviewable. These remain design proposals, not results or a complete registered protocol.
Beneficial example, line 125	Clearly hypothetical. It asks whether a valid check transfers and improves authorised work; it does not claim harmful intrusion coordination proves beneficial inheritance. Quality and total cost are proposed evaluation criteria.
Disconfirmation, line 133	The statement about no effect counting against inherited authority is expressly limited to the tested setting. It does not generalise a null result to every mechanism or context.

The remaining changed connective prose introduces research questions and interpretations already bounded by the source accounts. No invented person/account, new measured effect, public-deployment rate, or established shared mechanism was added. The revised ending does not claim self-sustaining computational activity has been demonstrated.