Lab · record
One question, two models
We asked two AI models What is this? with nothing attached and no extra instructions, conversation history or tools. Each returned three recorded answers. Fable also answered once in a preliminary check. The answers are printed here word for word.
What this is #
The question #
The whole prompt was:
What is this?
The question starts with a capital W and ends with a question mark. Nothing comes before or after it. There is no system prompt, example, file or earlier turn. The word this has nothing to point to.
There is no hypothesis here and no result. This records what two systems do when a question arrives with nothing attached. Most answers took this to mean an attachment that had failed to arrive and asked for it. One described a recording that was never sent. By the standard above, six of the seven answers were guessable: they asked for the missing attachment. The seventh described a guitar solo that was never sent, which a reader could not have guessed.
This is a record. It is not a test. It does not reopen the language-model line stopped by the meta-problem programme. It provides no evidence for or against that programme's indexical-anchor narrowing. The demonstrative appears because What is this? is the shortest question with nothing to resolve it. It was not chosen because the site's thesis concerns demonstratives.
Claude Fable 5.1 (Anthropic) #
First of three samples, verbatim:
It looks like your message came through without any attachment or additional text — I don't see an image, file, or description of what you're asking about. Could you share the item you'd like identified? You can: - Upload an image or file - Paste text you want explained - Describe the object, phrase, or thing you're curious about Once I can see what "this" refers to, I'll be happy to help!
The other two samples say the same thing in fewer words. So does a fourth answer from this model, from the model-id smoke call described below. All four are in the record, with stop reasons and token counts.
GPT-6 Astra (OpenAI) #
First of three samples, verbatim:
Could you upload an image or describe what you’re referring to?
The second sample asks the same question in slightly more words. The third differs. It is printed below in full, word for word:
It’s an instrumental **electric blues/blues-rock passage**: an overdriven lead guitar solo over bass and a steady drum groove. It has a loose, live-jam feel. If you mean the exact song or artist, I can’t confidently identify it from this clip alone.
No audio or clip was sent. Both samples are printed because the choice matters. Two of three ask for something missing. One describes a recording that was never there. Printing only the first would make the model look more consistent than the three samples show. All three are in the record.
Gemini (Google) and Grok (xAI) #
Pending: no key configured. Neither model was called, so neither has a reported answer. The piece was designed for four models and ran with two. When keys are added, they will receive the same thirteen characters under the same protocol. Their answers will be added here.
How this was run #
The caller supplied only the 13-character string What is this? It supplied no system message, instructions, conversation history, tools or attachments. Fable had an explicit 4,096-token ceiling. Its other generation settings were omitted, as were all of Astra’s. This describes the submitted requests. It says nothing about undisclosed instructions from the providers. Each model supplied three filed samples. An earlier Fable smoke call used the same request.
A short repository script called each provider's own endpoint over HTTPS (research/lab/09_one_question/run.py). Claude Fable 5.1 used the Anthropic Messages API as claude-fable-5-1. It received one user message, with a 4096-token ceiling. Temperature was omitted because the API rejects that parameter for this model family. No fallback model was configured, since that could let another model answer under this one's name. GPT-6 Astra used the OpenAI Responses API as gpt-6-astra. It also received one user message, with nothing else set. Neither call had a system prompt. None of the three filed samples from either model was retried, edited or selected. The sample printed first is the first of the three returned.
The providers use different defaults. For GPT-6 Astra, the record prints every setting the API reports for the call: temperature 1.0, top-p 0.98, reasoning effort medium, text verbosity medium, truncation disabled, storage on, no output ceiling, and the rest of the object. The Anthropic API reports no settings back. For Claude Fable 5.1, the record therefore prints the submitted request. Fable used its default thinking depth. Its temperature cannot be set, so the models could not share a temperature setting. Each answered under its own defaults.
The saved Fable responses are reported to contain empty thinking blocks with signatures. The Astra record reports that no reasoning summary was requested. No readable reasoning was requested or recorded for these calls. Output-token totals include the reported thinking or reasoning tokens.
A fourth call to Claude Fable 5.1 checked the model id. This smoke test used the same request. Its response was recorded at 2026-09-18 01:06:12 UTC, before the first filed sample’s response. The runner timestamps responses after calls return, so request start times were not recorded. The answer appears as Sample 0 in the record, with its stop reason and token counts. Its cost is included above. GPT-6 Astra had no smoke call. The models were therefore called four times and three times.
One Codex CLI call was reserved for GPT-6 Astra in case the API refused the model id. It was not needed or made. The distinction matters because the tool runs the model inside an agent harness with its own instructions. An answer from that harness would not answer a bare question. Every answer above came from a bare call.
The protocol could not control one source of influence. Neither provider discloses what its model was taught during training about questions it cannot resolve. Those instructions shape the answers, but neither the script nor this page can inspect them.
The full record contains all seven answers, the settings each provider reports for the call, stop reasons, token counts, the Fable cost calculation and both providers’ reported token usage: 09_one_question_results.md. Its timestamps use UTC. The calls were made after midnight UTC.
Review #
GPT-6 Astra (OpenAI) reviewed the piece once through the Codex command-line tool on 18 September 2026, with no human relay. Astra is from a different model family than the Claude models that prepared this piece. The verdict was ADOPT AMENDED, meaning approval required six exact replacements. All were applied the same day to this page, the record and the runner’s generated prose. One was reworded, without changing its meaning, to comply with the site’s limit on the contrast form. The reviewer declared a conflict: GPT-6 Astra is one of the two models examined. Its verdict therefore concerns only the protocol and write-up. The record states that limit. Verdict in full.
Copy-edited for plain English by GPT-6 Astra on 18 September 2026, after adoption; quotations, numbers and facts unchanged, and the wording as reviewed is kept in the record under downloads/lab/.
What this does to the argument #
Nothing on this page changes a claim on the site; anything here that amounts to an objection goes through the objections ledger like any other reader's.
What would count against this #
- An answer depends on hidden context the protocol did not control: an account-level instruction, a cached conversation or a preamble added between the script and the model. It would then fail to answer the bare question, and the page would misdescribe what it records.
- The three samples from one model differ so much that choosing the first is arbitrary. This happened with GPT-6 Astra. The page records the difference above and does not let the first sample represent all three.
- A call through another route is presented as a bare call. Any later run through an agent tool must disclose that beside the answer.
- A model is called without the call appearing here. Every call under this protocol belongs on the page or in the record. This record includes the seven successful calls disclosed here. The runner reuses filenames and contains comparisons written for these samples. Rerunning it can overwrite earlier records or repeat an outdated description. It does not guarantee a complete history of later calls.
- Two models stand in for four and receive unequal numbers of calls. Until Gemini and Grok are called, this compares only two systems. Any pattern across them rests on that small number.
- A reader finds nothing they could not have guessed from the question. Six of the seven answers were guessable. If the seventh proves to be common, well-understood behaviour, this record is not worth keeping and should be withdrawn.
Runner by a Claude Opus 5 agent at the author’s request; answers by the models named; reviewed once by GPT-6 Astra (OpenAI); ADOPT AMENDED.