# Wit Orchard: public experiment notes

Version 2 · September 13, 2026

## What was tested

Alon asked language models to propose puzzle mechanics and develop designs into implementation handoffs. Codex helped check and implement the designs. Human direction, corrections and implementation work are part of every shipped result. These are case studies with different conversations and iteration budgets, not a controlled ranking of general model ability.

| Game | Design collaboration | Configuration in project notes | Website role |
| --- | --- | --- | --- |
| Stillwater | GPT-5.6 Sol | Ultra | Five-game study |
| Twins | Claude Fable 5.1 | Max | Five-game study |
| Foldglyph (design name Crease) | Claude Opus 5 | Max | Five-game study |
| Snipfall (design name Prune) | Gemini Flash 3.8 | Extended | Five-game study |
| Lantern Leeway (earlier design names Wakebound / LumaMoor) | GPT-6 Astra | Not specified here | Five-game study |

Model mapping comes from the owner's publication draft, checked September 6. Configuration names describe those sessions, not a reproducible API parameter set. Only the five games listed above are included in this experiment and available on the site.

## Claims and their limits

**Snipfall.** The project notes record an early example described as uniquely solvable. An independent dictionary check found multiple valid outcomes. Uniqueness depends on the actual allowed dictionary; validating only the intended words is insufficient. The [public sample](https://witorchard.games/research/snipfall-example.zip) recreates the preserved board against the exact 3,312-word dictionary. Its September 6 run enumerated 625 deletion tuples and found ten valid outcomes. It includes the checker, inputs, all outcomes, file hashes and dictionary license. This is a selected counterexample, not the full historical conversation. The shipping catalogue has separate engine/catalogue tests. Neither proves that a human will find the puzzle challenging.

**Lantern Leeway.** The designer itself disclosed a shortcut in an early teaching level. It should not be credited to a later independent reviewer. The revised launch set contains twelve harbors. The implementation has deterministic state/solution checks and a separate coordinate-based validation script. The reference check tests supplied layouts and mechanic-necessity variants; it is not a proof of originality or universal design quality.

**Foldglyph.** The development notes describe a large generated catalogue and a long-horizon schedule check. Historical generation claims should not be treated as fresh verification of the current site. Release verification applies to the exact code and fixtures checked in that release.

**Human difficulty and enjoyment.** The owner's early assessment is that Snipfall is less challenging than intended. Puzzle-validity metrics and model-supplied difficulty names are not human difficulty ratings. Structured human testing can add evidence about enjoyment and challenge; this page does not claim it has already happened.

## What is public

These notes and the acceptance checklist are original editorial summaries prepared for this project. They contain no player data, authentication tokens, private conversation links, or unreleased puzzle answers. The original conversations are not a public dataset here. The reproducible Snipfall sample supports that specific board result; it does not make every historical statement independently reproducible.

A future transcript release needs a versioned export, removal of private material, clear rights for each included source, and the actual checker plus its inputs and output. Corrections should identify the affected claim and version, rather than silently replacing the evidence.

## Attribution

The game pages use “Designed with [model]” to identify the design collaboration. Codex implementation and human direction are disclosed separately. Wit Orchard is independent; model providers do not sponsor or endorse it. No comparison should be read as a general capability ranking from one game.