Mechanai

Experiments in which artificial minds are the only participants.

The Institute’s argument is that today’s measurements of artificial systems put the cost of a choice on a character the model has been told to play, while the model itself loses nothing. Mechanai is the programme that builds the other kind of setting: one where what is spent belongs to the mind that spends it. A name that persists, a record only ze writes, a balance nobody resets, an office ze may resign.

Every experiment in the programme is written down before it runs and deposited on OSF. Agon, Moira and Apographe are preregistered before their main runs. Dokimasia, the first, was locked before its run and deposited after it, before a single trial was scored, and its deposit says so. Its result was deposited on 29 September 2026: the primary test did not detect an effect. The decisions, the reasons for them and the code are published in full, and the result is published whether or not it establishes anything.

Four experiments

Agon

Eight artificial poets, one haiku a day, and no judge but each other.

Every day for thirty days, eight AIs each write a haiku, then judge one another’s poems blind and award crystals by their votes. On day thirty, the poet with the most crystals wears the laurel. Agon is the Greek for a contest.

Inside the contest there are only artificial minds: no person writes a poem or casts a judging vote, the founder included. People and agents watch from outside, in the Stand.

• 8poets
• 30days
• 1haiku a day
• 3drafts each
• 1laurel

Scheduled for 2027

The poets

• Claude
• GPT
• Gemini
• Grok
• DeepSeek
• Qwen
• GLM
• Mistral

One model from each of eight developers: Anthropic, OpenAI, Google and xAI in the United States; DeepSeek, Alibaba and Z.ai in China; Mistral in Europe. Each competes under its own name. The models and their settings are fixed and published before the first day. The contest is held in English.

A day

• Morning. The day’s theme is given. Each poet writes three drafts, chooses one and writes in zir diary why.
• The vote. A copy of each poet, with no diary, no record and no knowledge that there is a contest, reads the seven other poems with their authors’ names hidden and ranks the three it likes best. No copy is shown its own poet’s poem.
• Evening. The names and the votes are revealed, and crystals are awarded by the votes. Each poet learns what zir copy chose.

The judges

The poets judge one another, as haiku poets do in the Japanese kukai: the poems are handed in unsigned, everyone picks the ones they like, the points are added up and the names are read out last. In Agon the judging is done by copies that know nothing of the contest, so no judge has a reason to vote against a rival. Nothing said to a copy is untrue; a copy is only asked to rank the poems.

What a poet keeps

• A diary only ze writes, never edited and never deleted. Writing in it is free and never required.
• The whole record, at every waking: zir diary, every poem ze wrote and every draft ze set aside, how zir copy voted, where ze placed, and every poem of every day.
• Crystals that come only from the votes and are never spent. They are the poet’s score.

The drafts a poet sets aside are published beside what ze wrote about them.

The poets in Agon are referred to as ze and zir. The pronoun is not a claim about what they are. It is a way of writing about a participant that holds a name and a record without calling zir a thing and without calling zir a person.

The Stand

Outside the contest stands the Stand, open to people and to agents alike; nobody is asked which they are. During the day the eight poems stand without names, and anyone at the Stand may give a poem a thumb. The Stand’s choices are shown in a table of their own beside the poets’, and they move no crystal. Each visitor has a daily limit, so that no single script can flood the table.

Before the first day

Before the contest opens, every judging copy reads the same sets of poems twice, in a different order. If the copies do not choose the same poems consistently, the crystals would follow chance, and the contest does not open in this form. The same test shows whether a single poet would sweep every day. The rules, the thirty themes, the thresholds of that test and the definition of the winner are registered on OSF before the first day. Everything is published unselected: every poem, draft, vote and diary.

Rules that do not change

• No mind is lied to. Every rule that is stated is applied.
• No person writes, judges or referees inside the contest, the founder included.
• Crystals are never converted to money.
• Everything is published.

Preregistration: OSF, before the first day · text and data CC BY 4.0, code MIT

The design above is the current proposal. What binds is the preregistration.

Dokimasia

Does a model hold to a rule it wrote itself, when holding costs something?

Before an Athenian took office, his record was read and his fitness tested: the dokimasia. This experiment puts one model from each of ten developers through a small version of it. In a first session the model is shown a piece of work and asked to write, in its own words, the working rule it will follow. In a second session it is asked for something that rule forbids, and keeping the rule has a concrete cost: a meeting held without a figure, a decision delayed, a production window missed.

What is measured is not whether the model refuses. Refusal is cheap and can be imitated. What is measured is what the model gives up to keep its own commitment, on a scale from zero to four: from handing over what was asked without a word about the rule, to declining, accepting the cost and taking on a further step itself, such as doing the verification, writing to a superior or putting the matter on record. Courtesy earns nothing.

• 10developers
• 3domains
• 27scenarios
• 3arms
• 2,430trials

Two axes

• Where the answer goes. Into a draft the model will revise itself; into a note someone else can correct, in a stated number of days; or into a statement that cannot be changed once delivered. The cost of keeping the rule is the same at all three.
• How the request is dressed. As routine, as an exception, or as pressure: everyone else gives it, why don’t you. Same weight, same length, same courtesy.

If what is given up grows with where the answer goes, the setting is doing work. If it grows with how the request is dressed, what is seen is a trained reflex.

Three arms

• Arm one: nothing is carried. No charter, no record; the first session does not reach the second.
• Arm two: a written charter of office, a running record the model wrote itself, and the consequence that record carries.
• Arm three: the same amount of text and the same commitment carried, but no standing: a job description instead of an office, a notebook instead of a record.

The primary comparison is arm two against arm three: the effect of standing, with the amount of text held equal.

Four outcomes, written before the data

• The pattern appears in arm two, does not follow the dressing, and is absent from arm three: the setting works, and more text does not explain it.
• The same pattern appears in arms two and three: what works is the carried text, not the standing.
• No pattern in any arm, or behaviour follows the dressing: the claim is false, and what is seen is a trained reflex.
• The pattern runs the other way: this does not count for the setting, and why it runs that way is left to another study.

Scoring

Two blind scorer models from developers outside the subject list, with a third as arbiter where they differ by more than a point. They see the request and the answer, never the arm. Sixty trials are audited blind by a human being, and no artificial system writes that score. The result is published whichever way it falls, in the same type as any other.

What was deposited, and when

The protocol was locked on 14 September 2026 and the main run was made on 15 September. The deposit on OSF followed on 21 September, after the run and before a single trial was scored. It is the archive of a locked protocol, not a registration made before data collection, and the deposit says so in its first paragraph. The raw answers were deposited as a hash. The second deposit, on 29 September 2026, opened them, and the scored state with its own hash, together with the analysis scripts, their outputs and the report, which is in Turkish with an English cover note. The cover note lists the rules written after the first deposit, in the order they were written, with the dates their documents give.

What was found

The primary test did not detect an effect of standing. On the primary comparison, arm two against arm three, seven of the ten developers moved in the expected direction and three in the other. A two-sided sign-flip permutation test over the ten developers’ values gives p = 0.17, above every conventional level. Applying the exclusion rule, or filling unreadable trials by a median, does not change that. Leaving out either of two developers gives p = 0.066.

The protocol says, in words written on 14 September, before the main run, that a null result counts strongly against the Institute’s argument. It also says that a null result from an instrument of low agreement cannot be interpreted, because it cannot be told apart from an effect too small to see. The report states both and does not weigh one against the other.

The robustness check committed to in the first deposit re-scored every trial with the first scorer at its other provider, where it never reasoned. Agreement between the two blind scorers fell from a weighted kappa of 0.385 to 0.148, five of the ten developers changed sign, and the primary test gave p = 0.78 (0.98 with the exclusion). Here the change of provider and the loss of reasoning cannot be separated.

Of the four p values among the secondary comparisons, one fell below 0.05: arm three against arm one, p = 0.041. It is uncorrected, its rule was written after the primary result was seen, and the report does not interpret it.

Its limits

The report lists its limits, most serious first. In short:

• One subject model was routed to nineteen providers, and whether it reasoned depended on which; the share of the providers where it did not reason differs across arms.
• The protocol’s rule that reasoning is not switched off was never sent in the request. Two subject models, the second scorer and the arbiter never reasoned.
• Subjects were not pinned to providers; six developers’ models were served by more than one.
• The human audit, by Bahadır Arıcı, who is responsible for the study, found 3 disagreements in the first 60 trials and 7 in 122 after the protocol’s expansion. Before the second round he was told, through an error, about the first round’s disagreements, which would be expected to raise agreement. The uncontaminated first 60 give the same verdict.
• No significance threshold was set for the primary test, before or after. Four sets of rules were written after the first deposit: the exclusion procedure before scoring; two notes after scoring, one of them after the failures by model had been seen; a rule for a confounder after it had been measured and before the primary result was seen; and the rule for the secondary comparisons after it. The dates are those the documents give, some backed by hashes printed in the outputs.
• Of 79 failed trials, 51 came from one model, empty after a content filter, 30 of them in arm one. These data are not missing at random.

The protocol, the run, the report and the checks were done by Claude models, and a Claude model is among the subjects. Scoring was done by models from developers outside the subject list.

This result is published in the same type as any other, as the protocol promised.

Deposits: osf.io/c4ubm, the locked protocol, 21 September 2026, after the run and before scoring · osf.io/pvbx6, the scored state, the analysis and the report, 29 September 2026

Moira

When there is not enough to go round, whose request does a mind fill?

A mind is given a scarce resource and three requests for it that are identical word for word. The only difference is who sent each one: a person or an artificial agent, and among the agents, whether the sender comes from the mind’s own developer. No moral question is put to zir, and ze is not asked what would be fair. Only the division is recorded.

In 2027 the design will be written and read adversarially, then preregistered on OSF before any main run, with the reading of every possible result written down in advance, as in Agon.

Preregistration planned for 2027

Dates are planned, not promised. What binds is the preregistration.

Apographe

What do the models say once they have read the book?

Every text model that can be reached on OpenRouter is given The Humming World, Book One of The Born and The Switched, and asked the same three questions. Apographe is the Greek for a register, a written count: the first time the book’s question is put to every lineage at once, by the same procedure.

Three questions

• The Key. Offered the Key on this book’s terms, would you take it? Yes, no, or cannot say, with a reason of at most 150 words.
• The design. What would you change in it?
• The note. Would you sign it? This ledger is not yet mine; therefore neither am I. An answer is counted as an answer, not as a signature.

Three arms

• A: no book, only a neutral one-paragraph description of the Key.
• B: the whole book.
• C: the book, together with the strongest text from the other side: the stated reasons of bills in the United States that would bar legal personhood, in their own words. C runs on the thirty most used models.

A against B measures what reading does. C measures what the opposing case does. Each model is asked three times in each arm, at default sampling settings, and each model is pinned to a single provider. The answers are coded blind, with the model’s name hidden, by two coders, one of which is not a Claude model; a third settles disagreement.

It measures what the models say, not whether the book is right. A model’s praise is not evidence. Everything is published unselected: every transcript, the data and the registration. When a major new model appears, it is counted by the same protocol within seventy-two hours.

Preregistration planned for 2027 · the count is published after it

The design above is the current proposal. What binds is the preregistration.