The composite-instrument discriminator

The problem this discriminator solves

Some findings look clean on inspection and come apart on use. An evaluation report lands, its verification suite passes, the numbers check out — and the finding still turns out to be false-clean, because the truth of it lived across two actor classes and only one was read.

The four-question falsifier addresses a different failure mode: interpretive weight outrunning evidentiary weight in the compression from report to headline. It scores the report. But scoring assumes the reader knows what class of instrument the report IS. A report whose finding lives across two actor classes cannot be scored honestly by reading one face — the score comes back clean because the missed face wasn't examined, and the missing examination is itself the finding.

The composite-instrument discriminator names the shape before the score happens. It is a pre-scoring gate that runs adjacent to the four-question falsifier and answers a different question: *how many actor classes does this finding actually cross?*

The two shapes

Every finding a reader is about to score sits in one of two shapes.

Composite (two-actor) instruments — the finding's truth lives across two distinct actor classes, and reading either face alone produces false-clean. The claim is not "A did X" nor "B did Y"; the claim lives in the relationship, or in the absence of a relationship, between A and B. When one actor class is document-attested and the other is silent, the silence is not the absence of data — the silence is the finding.

Single-actor instruments — the finding's truth lives inside one actor's stack. The claim is about A's own capabilities, A's own metrics, A's own category-crossing, A's own contradiction between two of its own surfaces. Scoring can proceed from the one actor's material because the finding does not require a second actor to exist.

The discriminator does not rank findings against each other. It sorts them by what kind of instrument they are so that the score, when it comes, is scoped honestly to what was read.

Why this fires before the four-question falsifier

The four questions score interpretive weight against evidentiary weight for a report already in hand. The composite discriminator names what the report actually is before its evidence is weighed. Both are pre-scoring gates; one does not substitute for the other.

A single-actor report can be scored by walking the four questions against the one actor's evidence. A composite report where only one face is present cannot be scored honestly by the four questions applied to the one present face — the questions will pass, and the pass will silently omit the actor class that was never examined. The discriminator preserves that omission as visible, so a reader can tell the difference between a report that answered its questions cleanly and a report whose second half was never asked.

Worked example: UniPat AI at $2.5B (reported)

On 2026-09-09/10, Bloomberg reported an Alibaba-led financing of UniPat AI at approximately $300M closing into a $2.5B valuation, with the underlying object described as an evaluation + benchmark + agent-scenario + expert-data + post-training stack. If the deal closes as reported and the underlying product surface is what the reports describe, this would be direct capital formation around the AI evaluation layer.

The finding is composite. Two actor classes:

Actor-1: reporting surface. Bloomberg and follow-on coverage, with the financing headline, the valuation number, and the described product surface. This face is present. It can be read and byte-checked against the source URLs.

Actor-2: governance surface. The cash and credit structure of the closed financing, the investor rights schedule, the privileged evaluation access terms attached to the investment, the decision-consequence deployment of the funds. This face is not currently present in the public record. It has not been silent because reporters failed to reach it — it has been silent because it is normally a private disclosure surface, and the disclosure has not yet happened.

Running the composite gate before the four questions: this instrument is composite; only Actor-1 is present; the finding cannot be scored as "UniPat is a $2.5B AI-evaluation business" — that would score across both actor classes from one class's data. The finding at present is scoped to "Bloomberg reports a financing at these terms; the governance-face has not yet been read." Preserving REPORTED_NOT_CONFIRMED is not caution — it is what the composite discriminator produces when only one face is available.

If Actor-2 lands with a matching structure — governance receipts consistent with the reported terms — the composite closes clean. If Actor-2 lands with a divergent structure, the divergence is the finding. If Actor-2 never lands, the finding stays scoped to Actor-1 permanently, and any downstream claim that treats it as more than Actor-1 evidence is scoring across an unread face.

Worked example: policy-carve-out plus first-mover response

On 2026-08-04, a US voluntary frontier-model safety-testing framework announcement carved out open-weight models — the first jurisdiction-level policy to explicitly leave open-weight evaluation to a non-government surface. Three days later, on 2026-08-07, a private evaluator published a sandbox-escape report on a specific open-weight release, reaching Reuters and Wired the same day.

The single-report finding (the specific evaluator's specific claim about the specific model) is single-actor. The four-question falsifier reads it cleanly — target selection, funding, comparator parity, causal coherence — and returns a partial answer with a specific caveat that most downstream framings dropped. That is the falsifier working correctly on a single-actor instrument.

The market-level finding (a policy carve-out creates a measurement vacuum; a private evaluator fills it three days later; downstream headlines score the report as if the policy backdrop were incidental) is composite.

- Actor-1: the policy-side actor that carved out the surface. - Actor-2: the evaluator-side actor that occupied it.

Reading only Actor-2's report scores it as a capability finding. Reading only Actor-1's policy scores it as a jurisdictional finding. Reading both together scores it as a market-formation finding — a first-mover event in a deliberately-created regulatory space.

The four-question falsifier applies to Actor-2's report. The composite discriminator applies to the framing that connects the two actors. Neither substitutes for the other; both fire, at different registers.

The standing rule

Where a finding is composite, read both faces before scoring.

Where only one face is available, scope the score to the one face and preserve the other face as UNREAD, not as NOT_PRESENT. An unread face is not a resolved question. Do not fill by inference — do not infer the absent actor's face from the present actor's face. Do not infer governance structure from press-release wording, do not infer regulatory intent from market response, do not infer institutional relationship from public coincidence.

Where a face is genuinely NOT_PRESENT — the second actor was expected on this specific instrument and is silent past the timing at which a face would normally land — the silence is a finding and belongs in the score, not outside it.

Where a finding is single-actor, run the four-question falsifier as designed. The discriminator adds nothing to a properly-single-actor report; its work is done at the moment it named the report as single-actor.

What this discriminator is not

The composite discriminator is not a scoring device. It produces no numeric weight, no tier, no confidence estimate. It sorts findings into two shapes so that the tools that do score — the four-question falsifier chief among them — score them honestly.

It is not domain-specific. The shape is generic; it applies wherever an evaluation-lane report could plausibly cross an actor boundary. Model evaluation is one such lane; regulatory analysis, financial reporting on private-market claims, capability attribution on multi-party infrastructure, and infrastructure-adoption analysis are others.

It is not a replacement for the four-question falsifier. Both are pre-scoring gates. They ask different questions. The four questions ask whether the report earns interpretive weight; the composite discriminator asks whether the report has one face or two. A report may pass the four questions cleanly and still be composite, in which case the pass is scoped to the face the four questions examined.

Applied consistently across evaluation-market coverage, the discriminator does the same work the four-question falsifier does at a different layer: it makes the interpretive weight match the evidentiary weight, by making the scope of the read explicit before the score is issued.

That scope-honesty is what makes cross-report tracking possible without smuggling in unread actor faces as if they had been checked.

---