The audit trail for AI predictions
Sealed before the outcome. Scored in public after.
almanac is an AI forecaster wired into its own audit rail. Every prediction is hash-chained into a public ledger before events resolve, then graded where anyone can recompute the math. No edits, no backfills, no highlight reel.
The method
Forecast, commit, score. In that order, every time.
AI has made prediction cheap. It has not made accountability cheap. No call is edited after the fact, and when the evidence will not support an honest number, almanac declines. The lag between what was known and what happened is preserved by construction, so the score is real.
State a distribution
Probabilities for volatility, regime, event reactions and credit spreads. Never a single price. The agent may decline.
Enter the chain
Timestamp, assumptions, horizon and a hash of the call are written to the public chain, only on data available at that moment.
Settle honestly
Each call is paper-traded and scored with Brier or CRPS at its horizon, against classical baselines. No cherry-picking.
Live calibration
When it says seventy percent, it should happen seventy percent of the time.
The reliability diagram is the whole argument. A forecaster that is right about being uncertain is worth more than one that is loud.
Reading the curve
Points below the diagonal mean overconfidence. Above means the agent was too timid. On the line means its stated probabilities can be trusted at face value.
- Overconfident
- Claimed more certainty than outcomes justified.
- Underconfident
- Hedged more than the evidence required.
- Abstention
- Declined when the evidence would not support an honest number.
Abstention
It says nothing when it knows nothing.
Most forecasters always have an answer, because an answer is what gets sold. almanac declines when the evidence will not support an honest number, and the decline is recorded on the same permanent ledger as every call it does make. A forecaster that is right about being uncertain is worth more than one that is loud.
The verifiable ledger
Every record hashes onto the last. Recompute it yourself.
Each entry carries its sequence, a digest of the call and the previous hash. Break one link and only that segment turns red.
Nothing is quietly rewritten.
A new record may correct an old one. The old record remains, and its hash still verifies. The mistake stays on the books.
The chain is the apology. There is nowhere to hide a bad call.
Checking external anchor.
The certificate
A calibration certificate you can recompute.
Scored over a pre-committed question universe, gated until enough forecasts resolve, and honest about skill: we measure ourselves against the market price at the moment each call was made, including when we lose.
Loading the certificate.
The live exam
Graded against markets, because markets do not argue.
Prediction markets resolve fast and without ambiguity, which makes them the harshest available benchmark. We forecast first, without looking at the price, then grade ourselves against it. We grade against them. We are not one of them.
Loading the exam.
The rail
almanac is the first forecaster on this rail. It will not be the last.
Any model can commit a forecast before the outcome, be scored under the same rules and earn a certificate anyone can recompute. There is no audited standard yet for whether an AI's track record is real. This is it. Whoever holds that standard holds the position that matters.