Stream 1: Evidence aggregation

Uncertainty-aware fusion of model outputs, retrieval signals, and structured priors.

Purpose

Stream 1 is the entry point of the core scoring pipeline. It takes heterogeneous evidence — model logits, retrieval scores, structured priors, policy checks, and any caller-supplied features — and reduces them to a single normalized contribution.

The defining choice in this stream is how the reduction is done. Naive averaging and naive Bayes both produce systematically overconfident aggregates when source quality varies. Stream 1 instead applies an uncertainty-aware fusion that down-weights low-quality signals before combination.

What the stream does

1. Normalize each evidence source

   Inputs of different types (probabilities, scores, booleans, priors) are mapped to a common $[0, 1]$ scale with documented conversion rules.

2. Estimate per-source quality

   Each source is associated with a quality estimate — either a fixed prior from configuration or a runtime-derived estimate (e.g. retrieval consensus, model entropy).

3. Fuse with quality-weighted aggregation

   The fused contribution is the quality-weighted aggregate of the normalized sources, producing the stream’s $c_1$ output.

4. Mark abstention if no usable evidence

   If every source is missing, untrusted, or out-of-range, the stream sets its abstention indicator $a_1 = 1$ and contributes nothing to the aggregate.

Why not Dempster-Shafer or Bayesian fusion

The framework has been benchmarked head-to-head against both Dempster-Shafer and Naive Bayes; results are reported in Head-to-head comparison. Both alternatives are mathematically appealing but produce overconfident aggregates under realistic source-correlation conditions.

Stream 1 trades that mathematical elegance for an aggregator that is well-calibrated under correlated sources, validated empirically rather than asserted axiomatically.

Configuration surface

Operators can adjust:

• Per-source quality priors — fixed values used when runtime estimates are unavailable.
• Normalization rules — how non-probability inputs (booleans, raw scores) are mapped into $[0, 1]$.
• Abstention threshold — how much usable evidence is required before the stream is willing to contribute at all.

See Hyperparameters for the configurable surface and recommended defaults.

What this stream does not do

• It does not interpret the prediction. Stream 1 scores evidence quality, not the semantic content of the prediction.
• It does not detect adversarial inputs. That is the job of Stream 6.
• It does not produce the final calibrated score. That is the job of Stream 5, applied after the full composition.