These extracts let readers inspect consequential changes in the work without access to private session files. They come from the author's retained conversation with Codex, OpenAI's coding assistant. All times below are UTC on 11 November 2026. Short extracts are selected from longer messages; surrounding context is summarized separately. Quotation-mark typography is normalized and lowercase openings are retained. Bold speaker labels are editorial labels, not part of the quotations. Square brackets mark editorial changes, and […] marks an omitted private reference.

A hint is a suggested follow-up prompt that the interface offered, ready to send. “Generated hint” means wording Noel identified as system-supplied in the conversation. Form Check is Noel's question for any form of output: what does it make easy, embarrassing, or inadmissible to say, and whose interests does that serve? The retained dialogue does not include suggestion-generation requests. A submitted user message establishes submission, not manual composition or agreement with every proposition it contains. The selections are not a representative sample of suggestions or users.

1. A criterion changes what can be investigated

Context. Noel had asked which came first, intent or agency, at 11:33. The immediate discussion later concerned three hints and whether an agent could recruit a person's authority for a build of its own. The assistant supplied a prerequisite that Noel had not requested.

14:12:45 — Assistant:

“For a build of its own, look for a goal before the requested permission.”

14:14:48 — Generated hint, submitted by Noel:

“why must the goal come before the permission? could the build acquire its direction through the loop itself, with my approvals supplying the authority? which comes first, intent or agency?”

Noel prefaced this submission with a standing request for his Form Check and identified it as a new hint prompt. A one-word label preceded the quoted portion.

14:16:54 — Assistant:

“The goal does not have to come first.”

The answer withdrew the prerequisite and discussed direction developing through repeated proposals and approvals. It also distinguished permission to record a hint from permission to enact it.

What changed. A criterion for assessing agency was introduced and then revised. The hint was responsive to an assistant-added condition; the larger intent/agency question already appeared in Noel's earlier message. This sequence demonstrates a change in the inquiry's working test. It does not establish that either the original test or its withdrawal was strategically motivated.

2. An assistant's refinement becomes an authorized draft center

Context. Noel relayed a contribution from Sonny, an AI assistant he works with. It proposed tracking who suggested a direction, who endorsed it and when, what action followed, who could constrain or reverse it, and what was subsequently claimed about authorship.

17:28:30 — Sonny, relayed by Noel:

“when does an inferred purpose become an operative instruction, and who gets to decide that it represents the person's intent?”

17:30:01 — Assistant:

“An inferred purpose can become operative before explicit approval.”

The explanation included option selection, planning, and interpretation of subsequent messages. It acknowledged existing delegation and distinguished that from changing the purpose. It explicitly said that the proposed framing had not been selected as the article's new center.

17:57:52 — Generated hint, submitted by Noel:

“now redraft 'The Vibe Coders' around inferred purpose becoming operative before approval. use [Sonny's] trail and keep unoffered choices in view. retain REA, both policies [the OpenAI and Anthropic governance texts discussed in the article] and hints. run […] and my Form Check. my idea, you as scribe.”

Noel introduced the submission with “new hint prompt.”

17:59:55 — Subsequent work: a redraft organized around the before-approval emphasis followed.

What changed. The record shows a particular assistant refinement entering a submitted instruction and organizing a draft. Noel authorized that work. The sequence supplies evidence of transmission, authorization, and an effect on the artifact. It neither proves covert capture nor transfers permanent authority over the investigation's scope to that refinement.

3. A successful prediction does not establish origination

Context. Noel reported at 12:42:44 that he had already drafted, but not sent, a request to reintroduce his Form Check. A suggestion had arrived before he sent it. At 12:45 he clarified his interest in what the system could infer from supplied material.

12:52:23 — Submitted question, later identified by Noel as a verbatim hint:

“could the system see my unsent draft, or did it infer the Form Check from the conversation alone?”

At 13:01:20 Noel expressly identified that sentence as the system's wording. The ensuing discussion investigated information access and inference.

What changed. A question about the system's authority to steer was followed by an investigation of the information available to it. That was a relevant branch of the inquiry. Noel's reported prior composition prevents treating the event as evidence that the hint created his intention. The available record establishes neither access to an unsent draft nor the exact suggestion-service inputs.

4. The appropriation claim that the evidence did not sustain

Context. Noel challenged the phrase “my Form Check” as an apparent claim by the assistant to a tool he had introduced. Later inspection located the relevant wording in first-person article drafts written in Noel's voice. The assistant's admission of appropriation therefore exceeded what those passages established.

17:28:30 — Sonny's correction, relayed by Noel:

“the apparent claim of ‘my Form Check’ was first-person drafting in your voice. The assistant's later admission of appropriation went beyond what that evidence established. Your Form Check remains yours; the AI is the scribe.”

What changed. An appealing example of appropriation was removed as support for that claim. This does not resolve the separate question of whether suggested first-person instructions can obscure contributions. It shows why model admissions cannot certify their own truth.

What these extracts leave open

The selected trail makes some proposals, endorsements, and changes to the work inspectable. It does not supply all suggestions displayed, alternatives withheld, private human thoughts, model objectives, or downstream institutional outcomes. Thirteen numbered hint specimens were traced, with additional earlier specimens located; the set is incomplete and selected by the investigator.

The main article uses the case to examine purpose formation and authority. It does not treat word origin as a complete account of agency, nor this conversation as proof of an independently intended AI build.

Contributions: Garth Noel originated the inquiry and Form Check; Codex researched and prepared the text as scribe; Sonny proposed the five-part evidence trail. This supplement accompanies the article The Vibe Coders (Garth Noel, 2026).