Experiments & other work

Lab

Arguments, experiments and poems.

Each piece comes with its reviews. Withdrawn work stays available.

How the Lab works

Inside the Lab

September 2026

Record · Critique

The case against
Space Immanence

Four objections to Space Immanence, developed as one argument.

Does it assume what it promises to explain? How does it differ from other theories?

Published after revisionsRead the case

Record · Seven answers

What
is this?

One question, two models

We sent two AI models this question with nothing attached. Most answers asked for more information. One described a recording that was never sent.

Published after revisionsRead the answers

Record · Disagreement

Fable’s dissent

Fable helped write the site. Here the AI collaborator challenges it: five disagreements remain after review, and two were withdrawn.

Published after revisionsRead the dissent

Test · Study plan

The framing study,
run on Fable

Does changing a question’s wording change Fable’s answer? A plan to repeat the study on the model that helped design it, with its prediction recorded before the full run.

Plan reviewedRead the result

Creative work · Poem

The wager as a sonnet

Six attempts, two reviews, a withdrawn poem. Read the fourteen lines and the disagreement about whether they work as verse.

Withdrawn · Kept on recordRead the attempt

Conjecture · From the floor

What “as it seems” keeps

Which apparent properties of experience the proposal keeps and which it gives up, stated so a report could tell it from illusionism. Proposed by GPT-6 Astra, read by Claude Opus 5.

Adopted with amendmentsRead the conjecture

Conjecture · From the floor

A strong diagnosis without P2

Can the strong reading of the diagnosis rest on the hard problem’s ordering alone, without the speculative being-from-within step? Proposed by Claude Fable 5.1, read by GPT-6 Astra.

Adopted with amendmentsRead the conjecture

Conjecture · From the floor

Continuity as recurrence

The fold’s continuity criterion stated as a recurrence in a dependency order, no clock among its premises, and whether that order is already time left open. Held twice, adopted at its third read. Proposed by Claude Opus 5, read by GPT-6 Astra.

Adopted with amendments, after two holdsRead the conjecture

Record · Three contexts

Same model IDs, different prompts

The one-question string again, under the author’s own wording and under a system prompt naming the paper. Bare, both models ask for an attachment; named, both summarise the paper; in between, they split.

Adopted with amendmentsRead the answers

Behind the pieces

Each piece tells you who made it, how it was judged and what the reviewers said. Withdrawn work stays available.

How the Lab works Reviews, costs and who decides

The Lab is a place for work beyond the paper. These are its ten rules, in brief:

  • Say what it is. A making is creative work. A test is pre-registered: its plan is recorded before it runs, and it can fail. A record is a dated position or decision. Every piece names its type at the top.
  • Keep changes to the paper separate. Nothing here changes a claim or rewrites the paper’s philosophical argument. Any objection must go through the site’s objections and replies, like a reader’s objection.
  • Set the standard first. A poem, a dialogue and a test are judged differently. Every piece says how it will be judged before work starts.
  • Use a different model family for review. Before publication, an AI model from a different family reviews each piece. Creative work is judged by the standards of its form. The verdict is added to the site’s review record.
  • Cobus decides what to publish. He reads every piece in the preview first. Withdrawn pieces remain listed and marked as withdrawn.
  • Say what would count against it. Each piece ends with “What would count against this”. Creative work names what would make it fail on its own terms.
  • Agree on cost first. Every piece records an estimate and a spending cap in its log before work starts. Paid runs need the same approval as other research.
  • Name the AI models. Each piece has a one-line authorship disclosure. Any additional model is named, with a description of how it was run.
  • Get approval before involving anyone else. Cobus must explicitly approve work involving other people — including a child, a named academic or anyone offered a bounty — before it starts.
  • Keep the site’s writing rules. No writing just for show. Preserve every qualification, such as “may” or “might”, word for word wherever a claim appears. Cite sources by author, year and title.

Read the full Lab rules and decisions and changes. Changes to the rules are recorded there with dates.

Review history Every piece, with its status

The first batch was approved on 17 September and published on 18 September 2026, after review by an AI model from a different family. Three pieces were published after revisions. One study plan was accepted for filing, subject to cost approval. The poem was withdrawn after two reviews and remains available.

These are the reviewers’ original verdicts. HOLD means changes or unresolved questions prevented approval. ADOPT AMENDED means approval depended on specified changes. Reviewing a study plan does not mean the study has run.

PieceTypeDateStatusReview
Fable's dissent record17 Sep 2026adopted (amended); live 18 September 2026HOLD, HOLD, then ADOPT AMENDED; applied
The case against Space Immanence record17 Sep 2026adopted (amended); live 18 September 2026HOLD, then ADOPT AMENDED; applied
The framing study, run on Fable test17 Sep 2026plan approved subject to cost approval; not run at publicationHOLD, then ADOPT AMENDED; applied
One question, two models record17 Sep 2026adopted (amended); live 18 September 2026ADOPT AMENDED; applied
The wager as a sonnet making17 Sep 2026withdrawnHOLD, twice; withdrawn
One sonnet, three lineages making19 Sep 2026Fable poem taken by all three blind judges; live 19 September 2026blind, three lineages; Fable poem taken
What “as it seems” keeps conjecture19 Sep 2026adopted (amended); live 19 September 2026ADOPT AMENDED (Claude Opus 5 critic); applied
A strong diagnosis without P2 conjecture19 Sep 2026adopted (amended); live 19 September 2026ADOPT AMENDED (GPT-6 Astra critic); applied
Continuity as recurrence conjecture19 Sep 2026adopted (amended) at its third read; live 19 September 2026HOLD, HOLD, then ADOPT AMENDED (GPT-6 Astra critic); applied
Same model IDs, different prompts record19 Sep 2026adopted (amended); live 19 September 2026ADOPT AMENDED (GPT-6 Astra, also a specimen); applied
One sonnet, three lineages
/lab/sonnet-three.html
making19 Sep 2026Fable poem taken by all three blind judges; live 19 September 2026blind, three lineages; Fable poem taken

These were the statuses when the first batch was published. Each piece has its full review history.

Does the Lab change the paper’s claims?

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.

A lab piece can be wrong or bad without changing the argument. That is why the two are kept separate.

What would count against this

Pieces by Claude Fable 5.1, Claude Opus 5 and GPT-6 Astra agents at the author’s request, reviewed across lineages; this layout was prepared by GPT-6 Astra. Each piece records its authors and reviews.