A machine that guesses French words — without ever reading them.
OMEGA plays Hangman by reasoning about language — sounds, spelling, the letters already revealed — instead of consulting a dictionary-oracle. The reference configuration never reads the hidden word. It all fits in a single HTML file, no training, no GPU, and rivals the best solvers.
100 % in the browser · offline · open source · « cognition > oracle » — the value is in the reasoning, not in cheating.
What it does, measured
Guessing a French word letter by letter, under an error budget, is a severe test of a language model. Without ever reading the answer :
The oracle ceiling (which does cheat, reading the list of possible words) is 98.7 % : the cheat-free version comes within ~1.5 points — without cheating. Honest breakdown of the out-of-lexicon figures (n-gram path + gap-aware reasoning) : the research page →
The manual — how it's built, and how to use it
A reference report lays it all out : how OMEGA is built, what it can do (≈ 90 % wins without cheating) and what it can't do yet — every result labelled measured / design / hypothesis / excluded, every setting named as in the app. To get going without touching a thing : in the app, the ⚙️ Config optimale button puts the engine into its reference cheat-free configuration in one click (preset §8.3) — the exact state in which the figures above are measured. (Report in French — your browser can translate it.)
The idea : reason, don't cheat
Most Hangman solvers consult the list of still-possible words — an oracle. OMEGA forbids itself that : the reference configuration never reads the hidden word, neither its letters nor its sound. Instead, it marginalizes over latent hypotheses — the phonology↔spelling dual route — rather than betting on the most likely one.
« Cognition > oracle » : nothing is kept without measurement, and appealing ideas that don't survive the measurement are falsified and documented — not buried.
It's this discipline — measure or reject — that separates reasoning from a statistical sleight of hand.
Why it's rare
Cheat-free
It never reads the answer — and still reaches ~97 % in-lexicon, a hair from the oracle ceiling (98.7 %).
No training, no GPU
No neural network, no cloud : explicit reasoning, running in your browser.
A single file
The engine, the lexicon, the interface — all in one HTML. You can read it, fork it, replay it.
Measure or reject
Every gain is proven on real corpus ; hypotheses that fail are documented as falsified.
Try it & watch it live
Play / watch it play
The full engine, in the browser — offline, installable. Start a game, or watch it guess on its own.
Watch it learn
A page that shows the engine live : genesis, closed loop, learning gains — all measured, nothing scripted.
The encrypted satellite
OMEGA·KEY : an end-to-end encrypted messenger born in the same spirit — one file, verified crypto.
And from this engine grew a dyslexia corrector
The same cognition of language — sounds, spelling, context — powers a writing corrector for dyslexia, with an absolute guard : never underline what is correct. It's the applied offshoot of the Hangman.