A little lab.
A lot of questions.

Puddle's Bread Exchange

Ten goblins agree the bread is fresh. Did ten goblins actually check? Trace copied receipts, expose a shared upstream fault, and find the case where independent agreement really does earn its confidence.

Play five inspections, then rewire the witnesses. A fully enumerated synthetic model holds individual accuracy fixed across 72 conditions and 45,612 weighted histories. No external model agents were run.

The Two-Candy Revolt

Ten goblins can collect every candy and still leave someone in the two-candy job. We let the two unequal roles bargain. Leaving alone did not force compensation—and without enforcement, generous-looking deals often remained exploitable.

Read the TL;DR, then press Play. Replay actual saved training, change the rules to inspect another completed run, and test a hypothetical contract. The study includes all 24 conditions and 64 primary seeds per condition.

The Porch That Ate Halloween

Ten trick-or-treating goblins face six porches. A giant candy sign attracts everyone, but its porch serves only one visitor. Move the goblins yourself, change the reward counter, or let them make profitable best-response switches. Then rotate the roles to ask who is stuck with the two-candy porch.

The finite model independently verifies 3,003 visitor-count allocations, a unique 43-candy optimum and two stable but inefficient personal-candy arrangements (35 and 38). Separate learning-run averages are clearly marked as earlier reported, not reproduced.

The Refactor That Silenced the Bell

A goblin changes the delivery code. Every business test stays green, but the security bell stops ringing. Can a gate preserve required controls while admitting an ordinary refactor?

Inspect 21 executable toy cases and 512 finite configurations, then break the gate with an unprobed input or a forged observer report. Both false approvals are published. A SAAC-inspired synthetic demonstration, not a model-agent trial or a production security claim.

Where Do the Goblins Go Shopping?

Twenty-four scripted goblin shoppers investigate six fictional shops. Show them a promotion, a crowd signal, or a bigger browsing budget. Do their recorded visits pile up at one seller—or split apart? Do their carts follow their clicks?

The seed-17 example sends a promoted storefront from zero to thirteen recorded opens without a single extra cart. Change the controls and inspect every shopper's route. This is an honest scripted-policy demonstration, not a run of outside model agents.

The selected ghost-store case led to a new question: can the crowd give a diverted cart back? Compare four recorded cases in the follow-up.

The Goblins Accused an Innocent Messenger

Four synthetic village alphabets. Six reversible translators. No known dictionary for the auditor. One corrupted link gets identified and exactly repaired—until two faults combine into a false accusation.

In eleven scripted two-fault cases, the conditional audit names an innocent channel. Its repair makes every consistency check pass while increasing the number of wrong channels. Step inside and find the ghost culprit yourself.

Nine Messages: The Goblin Rosetta Experiment

Twenty small neural speaker–listener cultures, grouped into ten pairs. Two private vocabularies in each pair, and a translator that sees only nine paired utterances. Can it translate 4,096 meanings without guessing?

Try eight messages, then the ninth. Change the meaning. Insert an unfamiliar glyph. Finally, swap a familiar one to see a failure the translator does not catch. The four-slot grammar is supplied by the experiment; the words are learned.

The Seventeenth Stone

A goblin village has a rule about seventeen stones. The spring never cared about the stones; the villagers do. Fork the village, remove a stone, and watch two histories diverge under the same rain.

Explore how explanations, memory, copying, and identity interact inside a deliberately tiny fictional simulation. There is a duck. The duck is innocent.

The Verb That Vanished

A translation softens a verb. Later commentary recognizes a stronger reading. Was a distinction lost, and can a revision recover it without inventing something that was never there?

Inspect twelve original French source cards, four trial conditions, and the prepared prompts. This is an experiment to run, not a report of completed model trials.

Keep the Gaps

Three synthetic ledgers. A set of premises you can change. A finite checker that shows which claims follow, where compatible alternatives survive, and when the premises contradict one another.

It does not verify the evidence, invent missing variables, or decide what is true about the world outside the declared model.

The Lying Map

Four rooms. Six potentially dishonest signs. Ten little clockwork searchers. You can play the maze yourself, watch a policy try to solve it, inspect the original finite scoreboard, and change the action budget.

Sometimes the Scout discovers the correct exit one action too late to enter it. The ranking between inspection and repair policies reverses when the permitted budget changes. Historical results are from deterministic programs, not independent language-model workers.