DOSSIER // FORECAST SPEC // PROBABILITY STATED BEFORE OUTCOME // NON-PRICE · SYBIL-RESISTANT // WE PRE-REGISTER HOW WE COULD BE WRONG // DOSSIER // FORECAST SPEC // PROBABILITY STATED BEFORE OUTCOME // NON-PRICE · SYBIL-RESISTANT // WE PRE-REGISTER HOW WE COULD BE WRONG // 
DOSSIER REQUEST CLEARANCE ▸
FILE // FORECAST-SPEC   CLASS // PUBLIC   REV // v0.1
// DOC · HOW SIGNAL FORECASTS ARE SCORED

FORECAST SPEC

When the Federation issues a signal forecast, it commits in advance to a falsifiable outcome and scores itself after the fact. This page locks the three parameters that govern that scoring — and pre-registers how the central claim could fail, with a commitment to publish that test either way. The companion Forecast Record holds the results; this page holds the method.

// 01 · DOCTRINE
“A forecast that cannot be wrong is worthless. We state a probability before the outcome is known, fix the resolution rule in advance, and score the result whether it lands or not — and where a claim could be explained by something other than skill, we say so and test it.”
▾ GALACTIC FEDERATION OF FINANCE · RESEARCH DOCTRINE

A convergence signal forecasts one thing: that independent accumulation will follow. This spec defines how that is measured, scored, and checked against a control — without ever forecasting price. Every parameter below is committed as method + provisional value + recalibration trigger, consistent with the standing rule that no number is published as final before the data supports it.

// 02 · THE MEASURE

Qualified-holder growth — not price.

The outcome is behavioral and verifiable on-chain by anyone, with no trust in a private resolution.

For a token a signal fires on, count its qualified holders at signal time and again at the horizon. A signal resolves correct if that count grows past a fixed threshold:

outcome = 1 if (qhc_at_H − qhc_at_signal) / qhc_at_signal > threshold

This is non-price (a fraction of supply, not dollars), publicly recomputable from chain data, and roster-private — the count is total, never the watched wallets.

// 03 · PARAMETER ONE

The qualifying floor — F.

A wallet counts as a qualified holder only if it holds ≥ F% of circulating supply at the snapshot. The floor is the sybil defense, and it is self-defending: to count, a wallet must hold a real, non-trivial position — which is precisely the behavior being measured. Faking the metric requires doing the real thing.

// LOCKED · F

F = 0.01% of circulating supply, fixed across all tokens (not scaled by market cap), so there is one published number anyone can verify. Recalibration trigger: after the first 10 resolved forecasts, the observed holder-balance distribution is reviewed; if 0.01% systematically yields unworkable counts, F is adjusted once and republished with the supporting sample. Until then F is provisional.

// 04 · PARAMETER TWO

The broadening threshold.

The threshold defines what counts as “broadening.” It is a definitional constant, not a performance claim — the same bar applies to signal tokens and to the control. Skill is claimed only through the control comparison below, never through the threshold itself.

// LOCKED · THRESHOLD

+15% growth in qualified-holder count, applied identically at the H7 and H30 horizons. Because holder counts naturally grow more over 30 days than 7, +15% is a harder bar at H7 — so H7 hit-rates are expected to read lower than H30. That is a property of a fixed bar, not a weaker H7 signal, and it is disclosed here in advance.

// 05 · PARAMETER THREE · THE SKILL TEST

The control universe — and a pre-registered confound.

This is where the record’s credibility lives. The claim is: tokens our signals fire on broaden past the bar more often than comparable tokens with no signal do.

Phase 1 (launch) — unmatched control. The control is every eligible token in the same window on which no signal fired, selected by a mechanical, reproducible rule (no hand-picking), N = 20–30 per resolved batch, scored at the same horizons against the same +15% bar.

// DISCLOSED CONFOUND

Convergence signals often fire on tokens already in motion. An unmatched control compares signal tokens against all eligible tokens and therefore does not isolate predictive skill from pre-existing momentum. Phase 1 results are reported as suggestive, not isolated skill, and this page says so in plain language. We would rather state this ourselves than have it pointed out.

Phase 2 (rigor) — matched control. A second control is added, matched on holder velocity (qualified-holder growth rate in the window before signal time), holding momentum roughly constant on both sides. Both base rates — unmatched and matched — are reported side by side.

// PRE-REGISTERED COMMITMENT

The Phase 2 matched-control result is published on the record regardless of outcome. If matching shrinks or eliminates the measured edge, that is published exactly as a positive result would be. This commitment is the reason the comparison means anything — and pre-registering it, before any data exists, is the point. A separate corrections-log entry records this commitment so it cannot be quietly dropped later.

// 06 · SCORING

Brier, against two honest baselines.

Each resolved forecast is scored by Brier (squared error of the stated probability against the binary outcome). In Phase 1 the probability is pinned at p = 0.50 by design, so the Brier sits at the 0.25 no-skill line until enough history accrues — the model only earns a calibrated probability once forecasts have actually resolved. Edge exists only if the mean Brier beats both the coin-flip line (0.25) and the universe base rate from the control above. Reported split by horizon, never blended into a single headline.

// 07 · DISCIPLINE CHECK

What this spec is bound by.

No committed numbers without a spec. F and the threshold are provisional values with explicit recalibration triggers; no final figure is claimed before the data supports it. No price. Every parameter — supply fraction, holder counts, holder velocity — is non-price; price never enters resolution. Pre-registration over flattery. The disclosed confound and the publish-regardless commitment are the core: the method states how its own central claim could be wrong, and commits to testing and publishing it. This spec is versioned and append-tracked alongside the corrections log.