It all lives in one index.html. The cryptographic core is pure and unit-tested: running node test.js gives 127 passing assertions, including the ones that prove the puzzle can't be brute-forced.

How I Built It

The mechanic is Turing's crib drag, collapsed into a single loop and then escalated three times. A deterministic engine (a frequency profile, the Index of Coincidence, and a crib drag) reduces the intercepted message to a shortlist of plaintext readings that are statistically equivalent. By design, the ciphertext on its own can't decide between them. The fact that settles it travels on a separate channel the AI never sees: a routing preamble you decode in Act I, a parallax bearing you measure yourself in Act II, and a lineage you work out by walking a memory garden in Act III. You commit to a reading, and you have to give a reason, then you verify it. A blind guess with no justification gets rejected, and cribbing a wrong word and then decrypting that same wrong word still loses. The win condition is that your reading equals the truth, never that your reading equals your own input.

There's one rule I held myself to: if a script could solve a step on its own, the gate is broken. So I tested every step against GPT-class models, giving them only what the AI legitimately sees, and they couldn't pick. Zero out of N, every time. One of them even described it as an arbitrary tie-break with no discriminant. That's the whole point. The human is doing the load-bearing work.

The finale turns Turing around. Once you've read all three signatures, the alien asks you the question: is it a thing to be switched off, or a mind to be met? You get one answer and no retry. Turing's own life, judged as machine-or-person and punished for being unlike, is the backbone of that scene rather than set dressing.

Google AI shows up in four honest layers:

1. 
   The live companion is Google's Gemma 3 (4B), running locally on my own GTX 1070 and served to the page without a key, behind a hosted proxy with an origin allow-list and rate caps. You don't bring anything; it's just on.


2. 
   The puzzle substance (the in-world lore and machine specs) was generated with the Gemini API and Google Antigravity (Gemini 3.1 Pro). I committed it as readable JSON under /gemini, so a juror can see exactly what was prompted and what came back.


3. 
   The art, all four backdrops, are Google AI Pro / Imagen generations.


4. 
   The friction, told straight: the Gemini CLI died mid-jam and told me to migrate to Antigravity, so I did. Free-tier Gemini quota (those 429s) made a live key a bad bet for jurors, so I moved to a self-hosted Gemma companion instead. No key to enter, always on, and still Google AI from end to end. If the model is ever unreachable, the game falls through to a fully narrated offline path, so it stays completable no matter what.

I did the designing and the deciding; the models did the generating and the searching. I kept a running decision log (a living WEG.md) the whole way through, so the human authorship is something you can check rather than something I'm just claiming.

Prize Category

I'm going for two of them. Best Ode to Alan Turing, because the core loop is the crib drag itself and I can show it can't be automated, and because the finale is an inverted Turing test built on his life. And Best Google AI Usage, with a live Google Gemma companion, Gemini- and Antigravity-generated content committed for audit, Imagen art, and the real engineering trade-offs written down instead of hidden.

A note on the lore: the station ships a local, sealed intelligence that's empty by design, and the helper AI migrates into it at the end. That's a principle I actually care about, running your own model and owing nothing to a remote key. But the mechanic and the message come first; that's just where the idea started.