Drazta puts AI agents on the questions your organisation is carrying. Each question gets a probability anchored to a comparison class, re-read every morning, moved only when the agent can name what changed, and scored the day the outcome lands.
Open method, open ledger, results published whether or not they flatter us. Everything here is either something you can inspect today or something we have committed in advance to measure.
Ask what the odds are that a programme lands on time and you get a colour: amber, at risk, on track. Nobody writes a number, so nobody can be wrong later, so nobody ever learns. An organisation can run this way for years without finding out whose judgement to trust.
Forecasting tournaments established what actually works: comparison classes, frequent small updates, scoring against outcomes. It reliably beats expert intuition. It also demanded scarce, trained, expensive people, so it stayed where the stakes justified the cost, which meant almost nowhere.
An agent runs the same discipline across every open question, every day, for the price of a rounding error. It reads the evidence, holds when nothing moved, and keeps the log. The method did not get better. It got cheap enough to use everywhere, which is a different kind of change.
A distribution narrowing as evidence arrives
Every consequential question carrying a live number, a record of why it holds that value, and a score once the world settles it.
Three things happen to every question. All three are written down, and all three are open to inspection.
Before reading a single headline, the panel finds a comparison class and a rate. How often did things of this kind happen before? The anchor is logged on its own line, so it can be challenged separately from the argument built on top of it.
class tier-1 requalification sample 61 programmes, 2015–2025 base rate 31% within 9 months
Each question is re-read on a schedule, searching only for what is new since the last entry. If nothing material arrived, the panel holds the number and says so. Silence is a valid answer here, and it is logged like any other.
14 MAY hold 34% → 34%
no material news
22 MAY move 34% → 52%
supplier Ch. 11 filing
When the question resolves, the panel is measured against what happened: Brier score, calibration by band, and a standing list of the questions it read worst. Nothing is retired quietly.
scoring Brier + calibration intervals bootstrap, clustered reporting all questions, always
None of the reasoning above is our invention. It is the working method Philip Tetlock's tournaments found in the forecasters who kept beating everyone else, and we have deliberately not improved it. Our contribution is not a better way to think. It is running this one everywhere, forever, and keeping the receipts.
Not cleverness. Capacity, persistence, and reach. Those are the constraints that kept this method rare.
When a forecast costs a specialist's afternoon, you buy ten a year and spend them on the board deck. When it costs cents, you put one on every line of the plan. That is not the same product delivered cheaper. It is a different thing entirely, the way a spreadsheet was not just a faster ledger.
Update frequency is among the strongest correlates of accuracy in the tournament data, and the first discipline human forecasters lose when they get busy. A scheduled agent has no attention to run out of. It re-reads question 240 on a Tuesday in August with the care it gave question 1.
Outside experts cannot be handed your supplier records or programme telemetry. So the questions that matter most to you, the specific and internal and unglamorous ones, never got forecast at all. An agent on your own infrastructure reads what your own people read.
Any forecasting system evaluated after the fact can be made to look good. Drop the questions that went badly, choose the baseline you happened to beat, report the metric that flattered you. None of that requires dishonesty. A few reasonable-looking choices in sequence get you there.
So the choices were made in advance and filed: the question set, the baselines, the metric, the interval method, and the bar. If we miss it, the miss goes on this page in the same size type as a hit.
Pre-registration is ordinary practice in empirical science and nearly unheard of in this industry. It costs nothing except the ability to quietly move the goalposts later.
Drazta is a forecasting method packaged as an agent skill. The agent underneath is open source, and it runs where you put it.
Deploys into your VPC or onto bare metal. Questions, evidence and the ledger stay on disk you control.
Point it at a hosted frontier model or at weights you run yourself. The method is the product; the model is a setting.
One plain file per question, version controlled. No proprietary store to be locked out of, nothing you cannot audit by hand.
The review runs on a cron and reports a diff. Nobody has to remember to ask it anything.
Built on the open-source Hermes agent from Nous Research. The method ships as a SKILL.md you are free to read, fork, tighten, or disagree with. That is the only honest way to publish a method you are asking people to trust.
The method in full, the pre-registered protocol, and a section on the conditions under which we would expect this approach not to work.