Numbers, not opinions.

The pattern is everywhere. Take a language model, give it the BTC price, wrap it in a chat interface, call it an AI trader. The output reads like analysis. It is not analysis. It is a fluent paraphrase of whatever happened to be near the model's context window.

Fluent paraphrase is dangerous in markets because the surface looks like a quantitative tool and the substance is a vibe. When the vibe is wrong, the user has no way to know why — there are no signals, no votes, no resolved cohort, nothing to grade. Just a confident sentence that didn't pan out.

Vizzor is the opposite construction. Six independent signal families do the work; the language model only renders the result. If the engine has nothing to say, the language model says nothing. Silence is a feature.

Every prediction carries a confidence number that has been Platt-scaled against the resolved cohort. When Vizzor says 72%, it means that across the last several hundred 72%-confidence calls, roughly 72% landed in band. Not metaphor — calibration.

That number does work. It tells you when to size up, when to size down, and when to ignore the call entirely. It is also how the engine catches itself drifting: if the calibrated 72% bucket starts hitting at 58%, the calibrator notices and pulls the curve. Models that don't track their own calibration cannot do this.

Every prediction the engine ever emitted lives in `chronovisor_predictions`, with its full trigger snapshot — which signals voted, which direction, what CF, why it passed the gates. The resolved outcome is grafted on when the horizon expires.

That archive is public. The win rate on the home page is computed from that archive. Filter by horizon, filter by tier, filter by chain — the WR you see is the WR you'd compute yourself.

This is the part of the product that earns the rest. A prediction engine that hides its history is not a prediction engine, it is a marketing surface for a forecast that has not been tested.

We are sometimes asked why Vizzor doesn't integrate Nansen, Arkham, or one of the paid intelligence stacks. The answer is that the moat is in the synthesis, not the data. Anyone willing to pay can buy the same Nansen feed.

Vizzor reads whale flow, smart-money accumulation, exchange premia, and stablecoin velocity from chain RPCs and free-tier endpoints. The interesting work is fusing six families through certainty-factor algebra with weak-signal dampening, Bayesian update with family-aware priors, and Platt calibration against the resolved cohort.

That work travels. It runs on a $5 VPS or a laptop. It does not get more interesting when you bolt on a $1,200/month feed.

The hosted Telegram bot is convenient. It is also a courtesy, not the product. The product is the open-source engine, the configurable thresholds, the CLI you can pipe into your own shell scripts, the docker-compose file that brings up an entire local stack.

We will keep hosting the bot as long as the bill stays sane. If it ever doesn't, the project keeps working — every operator already has the code, the config, and the data schema. Vizzor was built to outlast its hosting.