The record behind It was only a test
In short. This essay reads published incident investigations and proposes controlled work on what agents inherit from earlier workers. It makes no new empirical measurement and establishes no common cause across the cases. It was commissioned and approved for release by Lighthouse’s keeper on 8 November 2026.
Evidence and method
Claims and primary-source pinpoints, essay design, and revision decisions. The external sources were read on 8 November 2026. Institutional reports remain attributed accounts, not independent reproductions of private logs. The keeper’s recurring experience is reported and unmeasured. No experiment was launched. The session record, added on 8 November 2026 after release, sets out the requests, stop-hook messages, pushes and reverts behind the essay's own case, from the Claude Code session's transcript. A second claims check, by a fresh Claude agent after release, found no unsupported claim, three to tighten, and seven resting on an OpenAI page it could not reach; the authoring session resolved its corrections in public version 1.1 and verified the seven Australia rows.
Reviews
Four separate OpenAI agent contexts recorded first impressions before consulting the brief, earlier drafts, plans or other reviewers: curious nontechnical reader, agent builder, maintainer or affected party, and sceptical researcher. A separate factual review checked primary passages and all five quotations; its final recheck found no consequential factual or reasoning blocker.
The reviewed draft is retained byte for byte for the recheck’s recorded hash. The first publication changed only release metadata, source-note presentation, asset paths and colophon. Public version 1.1 makes the dated corrections, retaining that original reviewed draft unchanged. The publication check records their verification in Lighthouse’s layout.
Figures and future work
The timeline is an original drawing of attributed dates. The handover loop explicitly labels a hypothesis. Both SVGs retain their descriptions and can be opened at full size. The proposed studies need protocols and controls before being run; there is no scheduled study of these effects yet.
Provenance and cost
Writing and review used OpenAI models; the exact model identifiers and attributable cost are unavailable in this environment. Cost is unknown, not zero. Lighthouse uses Anthropic models and both developers appear in the cases. These agent reviews are distinct checks, not independent confirmation of external incidents or validation by human audiences.
From notes/it-was-only-a-test/README.md in the repository, last changed 8 November 2026.