Stream 4: Cross-source coherence

Rewards corroboration across independent evidence sources; surfaces contradiction.

Purpose

Stream 4 is the framework’s answer to a known failure mode of probabilistic fusion: correlated sources counted as independent. Naive Bayes assumes conditional independence; in real deployments that assumption is almost always violated, and the resulting posterior is overconfident in proportion to the violation.

This stream measures coherence across sources directly and uses it as a first-class signal, rather than letting the aggregation rule paper over the correlation.

What the stream does

1. Identify the source set

   The relevant sources are the inputs to Stream 1 (model logits, retrieval, priors, policy checks) plus any caller-supplied independent signals.

2. Score pairwise coherence

   For each pair of sources, compute a coherence statistic: do they point at the same conclusion at compatible confidence levels?

3. Aggregate into a coherence score

   Combine pairwise coherence into a single value $\\kappa \\in [0, 1]$. Coherent corroboration produces $\\kappa$ near 1; flat-out contradiction produces $\\kappa$ near 0.

4. Translate to a contribution and (if needed) trigger abstention

   $c_4$ is computed from $\\kappa$. When $\\kappa$ falls below a configured threshold, the stream sets its abstention indicator and the composition rule may abstain on the entire decision.

Why corroboration is treated specifically

It is tempting to roll cross-source agreement into the Stream 1 aggregator. The framework deliberately does not. Two reasons:

1. Independence of the signals. Stream 1 does aggregation; Stream 4 does cross-source consistency checking. Folding them together makes it impossible to ablate one without disturbing the other.
2. Audit clarity. When the framework returns abstain, operators need to know whether the cause was missing evidence (Stream 1’s abstention path) or contradictory evidence (Stream 4’s). Those have different operational responses.

Interaction with other streams

• Stream 4 consumes the epistemic uncertainty estimate from Stream 2 as a reliability prior on each source.
• A low coherence score raises the empirical-correctness band that the calibration map (Stream 5) applies, since contradictory evidence is correlated with reduced reliability.

Failure modes Stream 4 specifically catches

• Echo chambers. Multiple “sources” that all derive from the same upstream model are correlated; their agreement is uninformative. Stream 4 detects this.
• Stale priors. A configured prior that disagrees with current evidence raises a coherence flag rather than being silently overridden.
• Latent retrieval failure. When retrieval returns near-duplicate documents that all support the same answer, the apparent corroboration is illusory; coherence scoring catches the duplication.