How a forecast becomes a record

2different models must agree before a Call resolves

A Call is scored only if it was recorded before its outcome. After its deadline, two different models each read a stored copy of the source registered with the Call, neither seeing the other's answer. The Call resolves only when both give the same answer and code finds the passage each one quotes, word for word, in that copy. The record says it was machine-reviewed, and an editor's decision may override it. Records are appended, never edited or deleted. Ambiguity stays open; it never becomes a NO.

Every ARCANE article carries a read: a thesis, the observable that would show it, a horizon, what would confirm it and what would break it. A Call is that read in a form that can be checked later against an exact source and deadline.

Once a Call is past its deadline, the source registered with it is retrieved again. The exact bytes are stored with their digest, the time they were retrieved and the HTTP status. When that source is an index page, up to three pages on the same site that it links to are captured the same way; code chooses them, not a model.

Two different models then read what was captured. One proposes yes, no or unresolved; the other decides on its own, without seeing that answer. Each must quote the exact passage it relies on, and a passage that reports only another forecast does not decide anything. The Call resolves only when both give the same yes or no, both passages appear word for word in the stored copy, and that copy was captured after the deadline. The decision is appended once, and the record says it was machine-reviewed.

If the models disagree, either answers unresolved, or any check fails, nothing is written and the Call is tried again the next day. An editor can still resolve a Call by hand, bound to the same registered source, and an editor's decision may override the machine's.

We do not create a historical forecast, choose a missing deadline, or change an outcome once recorded. The absence of confirming text is not a NO. Every attempt, refusal and unresolved case is kept, so the record cannot be improved by quietly dropping the misses.

A first prospective window for rates and credit Calls is planned, with held-out enrollment from 25 November to 25 December 2026 and review no earlier than 25 January 2027, and only with at least 30 independently resolved cases. These are planned windows, not evidence that anything has run. Public timestamping of each run is under evaluation and not yet switched on.

• Source passages and receipts are private ledger material; the method is public, the retained copies are not redistributed.
• PDF and other binary sources stay unresolved until a checked text path exists.
• Two models can agree and both be wrong. The deciding passages are kept with the record so an editor can check them.

Published

the resolution rules and the planned evaluation windows.

Withheld

which models resolve Calls, their instructions, storage, internal schemas and source copies.

1. 21 Sep 2026Resolution custody rules adopted.
2. 22 Sep 2026First public edition on Labs.
3. 25 Sep 2026A figure drawn from the published numbers added. No wording or number changed.
4. 25 Sep 2026Rewritten for automatic resolution, adopted on 24 September 2026: two different models must agree, and the record says it was machine-reviewed. The earlier text said only an editor could approve a resolution.

@techreport{arcane2026howa,
  title       = {How a forecast becomes a record},
  author      = {{ARCANE Research}},
  institution = {ARCANE},
  type        = {Method card},
  year        = {2026},
  url         = {https://arcaneintel.net/research/how-a-forecast-becomes-a-record}
}