Execution Model

An intelligence evaluation is a recorded argument, not an opaque score. It starts with a decision question and candidate set, binds the relevant evidence, applies declared policy, actively searches for reasons to withhold confidence, and emits a reviewable recommendation or refusal.

flowchart TD
    Q[Decision question] --> V[Validate candidate universe]
    V --> E[Bind evidence snapshot and posture]
    E --> S[Evaluate support]
    E --> C[Search contradictions]
    E --> F[Construct falsifiers]
    S --> B{Challenge burden met?}
    C --> B
    F --> B
    B -->|yes| P[Apply named scenario and policy]
    B -->|partly| H[Hold or downgrade]
    B -->|no| X[Refuse and request evidence]
    P --> R[Rank, compare, and test sensitivity]
    R --> A[Advisory recommendation]
    H --> O[Review packet]
    X --> O
    A --> O

Evaluation stages

Candidate validation establishes identity, comparable fields, quality, and lifecycle state. Evidence posture records what kinds of support are present and how directly they bear on the decision. Contradiction and falsifier passes expose adverse evidence before ranking rather than appending caveats afterward.

Judgment then applies an explicit policy in a named scenario. Ranking is meaningful only within that policy: weights, gates, uncertainty treatment, and ties are part of the result. A recommendation must retain the candidate fingerprint, evidence references, policy identity, decisive criteria, sensitivity, and reasons alternatives were not selected.

Decision state	Minimum retained reason	Permitted reader conclusion
recommend	challenge burden met; preferred action stable enough under declared policy	the policy prefers this action under this evidence snapshot
hold	a named dependency, review, or observation is unresolved	no action is justified until the condition closes
downgrade	an action remains discussable but support or stability is weaker	confidence and claim strength must be narrowed
refuse	a gate, evidence boundary, or validity condition failed	the requested recommendation is not available

None of these states independently authorizes laboratory execution.

Review path

sequenceDiagram
    participant Caller
    participant Eval as Candidate evaluation
    participant Challenge as Skeptical challenge
    participant Policy as Judgment policy
    participant Review as Review projection
    Caller->>Eval: question, candidates, evidence refs
    Eval->>Challenge: validated comparison
    Challenge->>Policy: support plus adverse evidence
    Policy->>Review: recommendation or refusal
    Review-->>Caller: brief, audit, next steps

Decision briefs and outsider packets make the same reasoning available at different review depths. Independent-rerun and benchmark modules test whether the outcome can be reproduced and whether policy behavior remains calibrated across blinded cases, counterfactuals, sensitivity changes, and regret analyses.

Learning without silent drift

Outcome feedback may propose adaptation, but it cannot rewrite an active decision retroactively or mutate policy invisibly. Learning records connect reviewed outcomes to a declared policy change. Existing packets retain the policy and evidence snapshot under which they were produced, so later improvements do not erase the historical basis of a decision.

Policy comparison must produce a new record rather than overwrite the old recommendation. A reviewer can then distinguish a changed evidence snapshot, a changed candidate universe, and a changed policy—three different causes of recommendation movement.