Calibrated forecasts, on the record
Decisions rest on forecasts, and most forecasts cannot be defended: the question was loose, the reasoning was not recorded, and the outcome was never checked against a source agreed in advance. A number without a fixed question, a stated base rate and a resolution rule is a guess with a decimal point.
The SolutionFireball formalizes the scientific and probabilistic method as a procedure that code enforces. Every forecast carries its exact question, its base rate and where it came from, the evidence that moved it and by how much, and the rule that will decide it. The reasoning is transparent while the forecast is open, auditable after it resolves, and outcomes are published to a timestamped public record that is never rewritten.
Each stage produces a recorded artifact before the next begins, and every probability on the page is arithmetic done by code from what the models propose. What each stage produces:
The question does not move once it is set. The source that will decide it is declared before any forecasting happens. The numbers are computed, never estimated by a model. More than one model vendor is involved, and resolution needs independent agreement. Open forecasts are re-run on a schedule that carries their reasoning forward, and every open forecast is checked daily for resolution. Nothing is overwritten: a correction sits beside what it corrects, and the record of every run is kept with the cost of producing it.
Questions enter through Amazon API Gateway. The forecasting engine runs on AWS Lambda with durable execution, checkpointing each of its seven stages and calling model providers from more than one vendor, while Amazon EventBridge schedules drive the daily rerun sweep, resolution checks and the daily report. Every run is appended to the forecast record in Amazon DynamoDB. Table streams drive a projection function that writes run reports and client-safe records to Amazon S3, which Amazon CloudFront delivers to the Client Portal and to Fireball's own product pages. Forecasts are published to Etched In Stone, the public record.
Forecasts are published to Etched In Stone, Watchfire's public record: each entry is timestamped by independent authorities and anchored to the Bitcoin blockchain, and nothing published is ever rewritten. A series runs from the first forecast through every update to its resolution, with the outcome graded against the final forecast in the open. Errata are appended, never substituted.
Deliberately absent from this page: how each stage does its work, the reference-class procedure, the evidence and update rules and their constants, prompts, model names, and infrastructure identifiers. The method is described to the level at which its discipline and its record can be judged, not reproduced.
Discuss how Fireball can put calibrated, defensible forecasts inside your decisions — on your own questions, under contract, or through curated topics in the Client Portal.