§02f Jupiter — Post-Hoc Forensics — Inter-AI analyst orchestration
Mars® Spec › §02f Jupiter — Post-Hoc Forensics › Inter-AI analyst orchestration
← The evidence trail · Section index · Causal attribution →
3. Inter-AI analyst orchestration
The post-hoc forensic layer may invoke one or more other AI systems as analysts over the evidence trail or its retained witnesses. Where it does, the following structural requirements apply.
Per-analyst attribution. Each AI analyst’s contribution to a forensic finding is attributed to that analyst and recorded as an order-typed, evidenced item. The forensic record states which analyst contributed which finding element. Per-analyst attribution is not optional; an analysis that produces a finding without attributing it to the analyst that contributed it does not satisfy this requirement.
Model-to-model divergence as a forensic signal. Where two or more AI analysts evaluate the same order-typed question over the same evidence and diverge, the divergence is recorded as a measurable forensic signal — an inter-analyst divergence record. The divergence record is distinct from the analysts’ individual attributed contributions; it is a first-class forensic artifact carrying the analysts’ identities, the question evaluated, the diverging results, and the divergence measure. The divergence is not resolved into an aggregate; it is retained as the evidenced record of the analysts’ disagreement.
Recomputability. Each analyst’s contribution and the recorded divergence are recomputable: an independent inspector re-derives each analyst’s contribution and the divergence from the retained witnesses and the analysts’ recorded derivation traces, without trust in the deployment. A mismatch between the re-derived result and the recorded result flags an unsound contribution.
Order-governed composition of analysts. Where the forensic layer composes the outputs of multiple AI analysts, it does so under the same composition discipline (§02a §3a) that governs all other composition acts. The composed finding inherits the contributing analysts’ soundness anchors. A composed finding that does not declare the soundness anchors it inherits is structurally incomplete.
Inter-analyst divergence and the divergence signal. The inter-analyst divergence record is a species of the cross-method divergence signal (§02a §4) instantiated at the forensic layer. It is governed by the same order-shape contract and witness obligations. An implementation that records per-analyst attribution and inter-analyst divergence without grounding them in the §02a composition discipline and witness chain is structurally incomplete at the forensic layer.
Hand-off to causal attribution. An inter-analyst divergence record is itself a forensic finding eligible for causal attribution under §4. A persistent divergence between analysts on the same order-typed question is attributed to a causal dimension — for example, divergence on a descriptive-order quality question attributed to the model dimension (a drifting analyst) or divergence on a prescriptive-order question attributed to the prompt dimension. The divergence record is the input to the §4 attribution act; the attribution act is not separable from the divergence-recording act. An implementation that records inter-analyst divergence but does not route it to causal attribution is a structurally incomplete post-hoc forensic layer.