# Word and deed — the method · v0.1 draft, Nov 16, 2026

What counts as a call, how a call is extracted and judged, how it is scored, what is
immutable, and how names are handled. Status: plan — the engine implements these rules
when the prototype is merged (engine task §150); every score will name the version of
the rules it was computed under. A new version of this file carries its own date.

## 1. What counts as a call
A dated, checkable claim by an author about the future move of an asset, a rate or an event.
Not calls: opinions about the past, general words, irony, questions, advertising, retelling
someone else's call, and conditional scenarios without a position by the author.

## 2. How a call is extracted
1. Two models from different companies extract calls independently and blindly — neither
   sees the other's answers, the document's date (in its fields, its text or its URL),
   or what happened next.
2. Code compares the two extractions by asset, direction and horizon.
3. A call found by only one extractor, or read two ways, is checked before it is scored;
   until the reconciliation rules are measured on hand-labeled data it goes to a manual
   queue, not into the score.
4. The quote must appear verbatim in the source, or the call is not recorded.
5. A call without a stated horizon is kept apart and flagged; how it is scored is set by
   a dated version of this file before the trial.
6. A conditional call ("if the central bank hikes…") is closed on the fact of the condition.

## 3. How a call is judged (code only, no models)
- A single stock up/down counts only if the stock beat (trailed) its own market's index
  over the same period, in the asset's currency. No local index or no price → "unresolved".
- An index, commodity, crypto asset or yield up/down: by the sign of the move.
- A target "above/below X": by the close on the horizon date.
- "No change": a move inside a band fixed before the trial.
- Events: by the official source — a central-bank decision, a closed prediction market.
- The starting price is the last close known before the moment the call was published.

## 4. Scores
- Hit rate: the share that came true.
- Weighted hit rate: by horizon, after Bailey, Borwein, Salehipour and López de Prado,
  "Evaluation and Ranking of Market Forecasters" (2018); the weights in use are listed in
  each version of this file.
- Brier score and calibration: only where the author stated a probability. Verbal
  confidence is not converted into a probability.
- A public score appears only after the minimum sample set before the test is reached; the
  count is shown beside the score while it builds.
- Archive (collected later) and live (recorded at the moment of the call) are kept apart
  and never mixed in one score.

## 5. Immutability
Each finished UTC day joins the Ledger's hash chain and its outside time anchor. Changing
any past record breaks the chain, and anyone can check. A correction is a new record
pointing at the old one — never an edit. A call its author later deletes stays on file,
labeled "deleted by the author on [date]".

## 6. Names
Public figures — bank strategists, central bankers, fund managers — are named, with a short
verbatim quote, its source and its date. A person is linked to a fund's action only by a
confirmed role and a primary document, never by the employer's name alone. Filings name the
filer: a fund's 13F shows the fund's holdings, not one person's trades. Commentators who are
not public figures appear under a type label ("video commentator 7") until they choose to
be named.

## Parameters
Numeric thresholds (target and horizon tolerances, the "no change" band, the default
horizon, the minimum sample) are candidates: fixed before the trial, published here with
the version, and changed only by a new dated version. A scoring rule enters the public
score only after it is measured on hand-labeled data.

_Not investment advice. The same content for every user._
