Forecasting as Confidence Input

Mechanism

The forecasting engine accepts inputs as typed records. Each record carries a value, a source identifier, a method-of-acquisition descriptor, a timestamp, and a confidence value bounded on the unit interval. The confidence value is supplied by the source, not inferred by the engine. A source asserting a measurement also asserts the credence it places in that measurement under its own stated method. The method descriptor enables downstream verifiers to evaluate whether the asserted confidence is plausible given the acquisition procedure, but the engine itself does not silently re-rate the supplied confidence.

When the engine computes a forecast, it does so by composing typed input records through declared computation steps. Each step is a function whose signature includes an explicit confidence-propagation rule. The propagation rule is selected from a closed set defined in the policy reference; permissible rules include independent conjunction, dependent conjunction with declared correlation, disjunction, weighted aggregation with declared weights, and bounded transformation with a declared maximum confidence ceiling. Each rule produces an output confidence as a function of input confidences alone. The engine refuses to execute steps for which no propagation rule is declared.

The engine maintains a structural invariant: the confidence of any derived value is bounded above by the joint confidence of the inputs from which it was derived under the propagation rule that produced it. The bound is computed and recorded at each step, not approximated at the end. If the joint input confidence is 0.6, no transformation, however many composition stages it traverses, can produce an output asserting confidence above 0.6 unless additional independent evidence is introduced through a declared evidence-injection step that itself carries a source and confidence.

Forecast outputs are typed records mirroring the input format: each forecast carries a value, a producing-engine identifier, a method-of-derivation descriptor, a timestamp, and a confidence value reflecting the joint-input bound. The forecast also carries a derivation lineage enumerating the inputs and the propagation rules that produced it. A relying party consuming the forecast can independently verify that the asserted output confidence is consistent with the declared inputs and rules; the engine's claim is auditable rather than asserted by authority.

Sources that supply low-confidence inputs cannot recover high-confidence outputs by routing their inputs through additional engine stages. Each stage applies the propagation rule and the bound is preserved. Attempts to introduce confidence uplift without corresponding evidence injection are rejected by the engine as policy violations and recorded in the lineage as rejected operations. The structural prohibition against confidence laundering is enforced at the data-path level, not by reviewer convention.

Operating Parameters

Confidence values are represented as bounded real numbers on the unit interval, with policy-defined precision. The policy reference specifies the closed set of admissible propagation rules and, for each rule, the parameter space within which it operates: declared correlation coefficients for dependent conjunction, weight vectors for weighted aggregation, ceiling values for bounded transformation. Sources are required to register their method-of-acquisition descriptors with the engine; unregistered methods cause input records to be admitted at a policy-defined default confidence ceiling rather than at their asserted value.

Evidence-injection steps are governed by an injection-rate parameter that limits how rapidly confidence can be raised through new evidence within a single derivation chain. Freshness parameters bound the age of inputs admissible to a given forecast; stale inputs are either excluded or admitted at a degraded confidence per a policy-defined decay function. Forecast outputs carry a confidence-floor parameter below which the engine emits a structured insufficiency rather than a forecast, preventing the system from presenting near-zero-confidence values as actionable outputs.

Source-registration parameters require each source to commit to a method-of-acquisition profile before any of its inputs are admitted at the source's asserted confidence. The profile names the instruments, procedures, calibration intervals, and known failure modes of the source, and it is itself a typed artifact stored alongside the source's input records. Changes to the profile produce a new profile version; inputs supplied under an obsolete profile are admitted at the policy-defined default ceiling rather than at the source's asserted value until the profile is updated. This parameterization closes the loophole by which a source could quietly relax its acquisition standards while continuing to assert unchanged confidence values.

Alternative Embodiments

In a single-source embodiment, all inputs originate from a single source and confidence propagation reduces to bounded transformation. The engine still maintains the joint-input bound but the propagation rules collapse to a simpler family. This embodiment is suitable for closed sensor networks under unified custody.

In a multi-source independent embodiment, inputs originate from sources whose methods of acquisition are demonstrably independent. Independent conjunction and disjunction rules dominate, and the engine can in principle produce output confidence exceeding any single input's confidence through corroborating evidence. The embodiment is suitable for open-world fusion such as cross-vendor sensor aggregation.

In a multi-source correlated embodiment, sources share method or substrate dependencies. The engine requires declared correlation parameters and applies dependent-conjunction rules. Output confidence is more conservatively bounded, reflecting the structural fact that correlated sources cannot corroborate one another as strongly as independent sources.

In a streaming embodiment, the engine operates over a continuous flow of input records and emits forecasts on a rolling basis. Each forecast carries the joint-input bound computed over the inputs active in its derivation window. Late-arriving high-confidence inputs may revise prior forecasts; revisions are recorded as new forecast records with their own lineage, not as silent updates to prior records.

In a federated embodiment, multiple instances of the engine operate under separate custody and exchange forecasts as inputs to one another. The receiving engine treats incoming forecasts as input records with their own source-asserted confidence and applies its own propagation rules. The federation does not collapse into a single trust domain; each engine remains responsible for its own bound enforcement.

Composition

The confidence-input mechanism composes with the conflict-resolution mechanism: when independent forecasts disagree, the conflict-resolution protocol consumes their per-forecast confidences as part of the disagreement record. It composes with the lineage mechanism: every propagation step is an entry in the forecast's lineage, enabling external auditors to recompute the bound. It composes with the policy-reference mechanism: the closed set of admissible propagation rules and the parameter envelopes for each rule are declared in the policy reference and versioned alongside the engine.

On the consuming side, downstream decision mechanisms accept forecast records as typed inputs and apply their own confidence-aware logic. A decision threshold expressed as a minimum acceptable confidence is enforced against the forecast's reported confidence; the decision mechanism does not re-derive or override the bound.

The mechanism also composes with the identity-layering infrastructure: a forecast's source identifier resolves through the identity-layer stack of the producing entity, exposing which scope of the source's authority the forecast was issued under. A forecast asserted under a professional-licensure layer carries different downstream weight than the same forecast asserted under a contextual-session layer, even when the numeric confidence value is identical. The engine does not collapse this distinction; it preserves the issuing layer in the forecast's lineage so that consumers may apply scope-aware decision logic without re-querying the source.

Distinction from Prior Art

Conventional forecasting systems treat confidence as a presentation property: a single confidence value is computed at the output by a model whose internal computations do not preserve a bound between input credence and output credence. Probabilistic graphical models and Bayesian networks propagate probabilities through declared structure, but they typically assume calibrated input distributions and do not enforce a structural prohibition against confidence uplift through unevidenced transformation. Ensemble and meta-learning approaches aggregate predictions from multiple models but treat the aggregated confidence as a property of the ensemble's calibration rather than as a bounded function of input source confidence.

Sensor-fusion frameworks track per-source uncertainty but generally do so within a fixed mathematical framework (Kalman filters, Dempster-Shafer combinations) without exposing the propagation rule as a typed, policy-declared parameter. The fusion's behavior is implicit in the framework rather than explicit in a closed rule set that a relying party can audit.

The disclosed mechanism differs structurally: per-source confidence is a typed input-side assertion rather than an inferred output-side estimate; propagation rules are declared, closed, and policy-bound rather than implicit in a model architecture; and the structural invariant that output confidence is bounded above by joint-input confidence under the declared rule is enforced by the engine itself, not relied on as a property of the chosen model. Confidence laundering — the production of high-confidence outputs from low-confidence inputs through opaque transformation — is structurally prohibited rather than discouraged.

Disclosure Scope

A further structural distinction concerns auditability. In conventional systems, recovering the chain of reasoning that produced a given confidence value typically requires access to the model's internals: weights, intermediate activations, training data, or proprietary calibration tables. Even where such access is granted, the trace is rarely a typed artifact and rarely supports independent recomputation. In the disclosed mechanism, the propagation lineage is itself a typed first-class output, sufficient on its own to permit any third party to recompute the bound and verify that the engine's stated confidence is consistent with its declared inputs and rules. This shifts confidence assertions from a matter of vendor representation to a matter of structural verifiability, which is the property that licensing and regulatory regimes increasingly demand.

This article describes the confidence-input mechanism of the forecasting engine as disclosed in the cognition patent. Implementations in which forecasting input fields carry typed per-source confidence, in which a closed set of declared propagation rules governs the derivation of output confidence, and in which output forecast confidence is structurally bounded above by joint-input confidence under the producing rule fall within the disclosed scope. The scope encompasses single-source, multi-source independent, multi-source correlated, streaming, and federated embodiments, and is independent of the specific numerical representation of confidence, the specific catalog of propagation rules elected by a given policy reference, and the specific transport mechanisms used to convey input and forecast records.