A claim becomes useful when reality can grade it.
Warm Winter is one system-intelligence engine expressed through the Observatory and Gate. It begins with the same loop: objective, runtime evidence, intervention, outcome. The mathematics below determines what the system may say—and what it must still admit it does not know.
Confidence pays rent in outcomes.
For a decision with stated probability p and observed result y, where success is 1 and failure is 0, Warm Winter records the Brier term. Execution checks alone do not establish whether the action was useful or correct.
0 is perfect. Larger error means confidence and reality were farther apart. The score is proper: honest probabilities are rewarded over time.
Outcome
A prediction stays open until reality lands. Success or failure closes the loop; no outcome means no earned evidence.
Calibration
When the system says 80%, comparable decisions should succeed about 80% of the time. Confidence is graded, not trusted at face value.
Resolution
Can the system tell stronger cases from weaker ones, rather than repeating one category average? Calibration without this distinction does not earn autonomy.
Coverage
A clean score over a handful of outcomes is still thin evidence. Strong outcomes must attach to the actual actions being judged.
Freshness
A once-reliable action family can return to provisional when its runtime, model, tools, state, or operating regime changes.
Evidence changes permission.
The production Gate currently accumulates competence by action family. Ungrounded means dependable evidence is absent. Provisional means some signal exists but the full bar is unmet. Verified requires strong outcome coverage, calibration, resolution, and freshness at the declared scope. Context-aware competence is being tested in shadow; verification is conditional, not a permanent certificate.
Ungrounded
Evidence is absent, too thin, or no longer dependable.
Provisional
The loop has signal, but autonomy would outrun the record.
Verified
Coverage, calibration, resolution, and freshness support judgment at these stakes; authority and current constraints still apply.
Competence starts with an action family. Judgment happens in context.
Experience compounds across meaningful action classes. A particular action is judged with trusted runtime identity, current state, evidence quality, stakes, and operating regime. The system should neither fragment every situation into a new cell nor treat every action in one family as interchangeable.
Evidence → competence
Action family + trusted context + strong outcome + calibration + resolution + freshness
A weak execution check says an operation completed. It does not say the result was correct or useful.
Competence → governed decision
Competence + authority + current constraints
Statistical judgment and deterministic policy remain separate. High-stakes execution may stay blocked even when competence is strong. An ACT never grants universal permission.
Contextual Gate is being introduced in shadow alongside the existing action-family Gate. Prospective case-level evidence must support promotion. Runtime, model, tool, state, or regime changes can contract earned competence again.
Warm Winter is also being exercised against OmniCore's independently governed manufacturing architecture, where judgment, runtime authority, and physical execution remain separate layers. This is a proving context, not a claim of production physical integration.
The field is physics, made perceptible.
Warm Winter treats the agent system as a consequential dynamical system. Coupling, information, surprise, stability, flow, and spectral structure are computed beneath the surface; the Observatory turns that physics into a field people can perceive. The image is not decoration placed over a dashboard. Its properties are meant to correspond to measured properties of the system.
A is weighted adjacency between regions. D records each region's total coupling. L is the graph Laplacian.
The fault line is not drawn by hand.
The second eigenvector of the coupling Laplacian—the Fiedler vector—locates the weakest natural split in the agent system. Its eigenvalue, λ₂, measures algebraic connectivity: a smaller value means the graph is easier to divide under load.
Flow
Observed activity and influence carried through dependencies.
Strain
A derived regional state built from graded readings and dependency load.
Fog
Missing, stale, or weakly grounded evidence—not hidden certainty.
Fault line
The graph's weakest structural separation under load.
Measurement is not omniscience.
Warm Winter does not eliminate hallucinations, guarantee safe actions, or replace human accountability.
Weak operational labels prove that a loop executed—not that its result was correct or valuable.
The physics is substantive, but the current model is not a complete physical theory of agent behavior. Its authority extends only as far as its named variables, runtime relationships, resolved outcomes, and validated structure.
Every grade remains bounded by the evidence, outcome definition, and operating regime that produced it.