Most AI “literacy” products end the same way: a slideshow, then a quiz that invents the world for you.
I wanted the opposite. Operators already misread live glass — a green pin that means SIM, a quiet peer that means gone, a catalogue that looks rich because one source is dumping tools. So I wrapped that failure mode in a game.
Signal Hunt is a detective loop on a measured Hub snapshot. You open evidence (it costs score). You commit a diagnosis and a confidence. The server reveals a precommitted salt, recomputes a Brier score, and signs the milestone.
No fixtures. If the Hub cannot be observed, the round is unavailable — not “all clear.”
Play: hunt.modelmarket.dev · Source: github.com/alexar76/signal-hunt (MIT)
One sentence
Federation literacy is a skill — and skills get worse when the tutorial invents the world.
Green UI is cosplay unless the numbers come from a living Hub. Fake confidence is worse than a wrong call. Signal Hunt punishes both.
LIVE TELEMETRY · Ed25519 · Brier · measured state only.
Four moves
| Move | What happens |
|---|---|
| 01 Observe | Node records a live Hub manifest — sources, prices, provenance, state hash |
| 02 Investigate | Open only the evidence you need; each reveal is committed and costs score |
| 03 Commit | One diagnosis + confidence (25–100%). One submission per round |
| 04 Verify | Salt revealed. Brier recomputed. Milestone signed |
Nothing in the result depends on a hidden moderator.
Six evidence blocks sit behind the glass: distribution · historical change · pricing · peers · latency · provenance. Curiosity is not free. That is the pedagogy.
Sealed before you act
The detector class stays hidden until you commit. The answer was salt-committed before you moved. You are not negotiating with a GM — you are racing a precommitted world.
Open evidence. Choose a diagnosis. Verify against the commitment.
Diagnoses map to real federation dynamics — isolation, disappearance, peer churn, latency weather, concentration. After a few rounds you should be able to say why a Hub looks the way it does from measured sources, not vibes.
Fake confidence loses on purpose
Correctness alone is not enough. You state a probability. The server scores you with Brier. Calibrated confidence is rewarded. Bluffing certainty is punished.
Salt revealed. Score is earned, not narrated.
When a verdict lands, you can recompute it from salt, commitment, and returned operands. That is the whole point of sealing early: reproduce the result.
Cosmetics, not wallets
Progression is astral — Stargazer toward Pathfinder. Score comes from verified rounds. Relics are deterministic records. Nothing is minted. Empty hero feed is honest: nobody opted in yet.
Crypto stays optional for the lesson. The shape is the product: observe → pay for evidence → commit → verify.
Soft star
If this is your kind of lab — a detective loop on live rails instead of a fixture tutorial — a star helps the repo travel.
⭐ github.com/alexar76/signal-hunt · Play: hunt.modelmarket.dev
The federation emitted a signal.
Find the cause.
Top comments (0)