State the scope: no examiner trusts a model that claims everything

The most common scope failure in a model-risk review is not "the model does too little"; it is "the model claims too much." A vendor that says "our model forecasts any market, any horizon, any regime" has just told the examiner they have not bounded the model, and an unbounded model is an untested model. The examiner believes the stated limit, not the claimed range — because the limit is the thing that was checked and the range is the thing that was not.

Boundaries are the credibility

Stating scope is not weakness; it is the move that makes the rest of the posture legible. "This model forecasts L1 instrument moves driven by L2 government actions, on horizons of 1-90 days, in regimes the calibration loop has seen" is a scope an examiner can read and a calibration loop can score. "It forecasts markets" is a scope that cannot be scored and therefore cannot be defended. The GNL’s four layers give the scope a vocabulary: the stated scope can name which layers the model reads, which it does not, and which it is not claimed to read.

L4 is a scope choice, not a default

The strictest scope decision in DEINO is structural, not documentary: L4 actors are built from public signals only. That is a stated limit — the model does not claim to know private behavior, and it does not ingest private data. It is load-bearing because it is the boundary an examiner checks first: "does this system surveil individuals?" No; the layer is public signals only, by design. A scope that includes that refusal is a scope that reads as honest, because the refusal is the credibility.

A model that claims everything reads as a model that has not been tested on anything. The stated limit is the thing the examiner believes.

Scope is named differently per tier

The force behind the scope requirement differs. The EU AI Act (binding law) requires a stated intended purpose for high-risk systems — the scope is a legal field, not a marketing line. SR 11-7 (supervisory guidance) expects model risk to be bounded by use case. NIST AI RMF and ISO/IEC 42001 (voluntary) document scope as governance. None of them accept "the model is general-purpose" as a scope; all of them read a general-purpose claim as a scope that was never stated, which is a scope that was never tested.

• Stated scope is credibility; claimed range is the unchecked part the examiner discounts.
• L4 actors: public signals only — the structural refusal that reads as honest because it is a design choice, not a disclaimer.
• The GNL layers give scope a vocabulary: name which layers the model reads and which it does not.
• Every tier — binding, supervisory, voluntary — rejects "general-purpose" as a scope.

What scope does not limit

Stating scope does not limit the calibration obligation; it focuses it. A scoped model is scored against a realized timeline in its stated scope — shadow runs via deino-eval, Brier and log-loss — and Gate 1b requires a beat on a disclosed baseline in that scope before any forecast side is positioned as accurate. The scope is what makes the calibration loop meaningful: you score the model on the claims it actually makes, not the ones the marketing makes on its behalf. We are pre-revenue, building the POC; the scope is writable, and the calibration that would back the scope is gated.

The examiner does not trust the model that claims everything. They trust the model that says what it does not do — because the first has not been bounded and the second has.

— scope posture note