From Neural Capability to Governed Authority: Evidence-Bearing Intelligence in EvoMind

Modern agent systems can call models, route among heterogeneous backends, retain memory, invoke tools, and act on external environments. The harder systems problem is proving where neural inference is actually wired, whether it is reachable on the canonical runtime path, what authority its output is allowed to exercise, what kind of evidence that output constitutes, what artifact or state it was grounded in, whether resulting state persists across process boundaries, and whether later behavioral change is causally attributable rather than merely correlated with stored state.

This technical white paper develops that problem as the central research thesis of EvoMind: intelligence should be treated as a governed, evidence-bearing software capability whose wiring, reachability, epistemic authority, provenance, persistence, and downstream influence are separately testable properties. The paper formalizes an inference-seam model, an evidence-state taxonomy for neural dependencies, an authority model, and a non-substitution rule for epistemic signals. It examines inference-seam auditing, TaskDecomposer authority boundaries, the VisualCritic proof-integrity case, artifact-grounded neural judgment, deterministic provenance, SaltInference as a controlled neural substrate, falsifiable architecture claims, and governed capability formation from persistent causal influence through admission, restart persistence, and transfer.

The contribution is intentionally narrower than a claim of novel local inference, model routing, persistent memory, or tool use. It is an architectural and evidentiary discipline for making intelligence injection points and cognitive state transitions falsifiable. The paper reports first-party systems evidence and explicitly does not claim external AGI validation or third-party replication. It concludes with a threat model, related-work comparison, quantitative evaluation framework, external replication protocol, limitations, and a research agenda.

Evidence cutoff: private EvoMind repository main at c76c28e9fbfa3068a1d8a70bc84f45e0d74e4d1f, committer timestamp 2026-09-20T22:36:23Z. The paper distinguishes E-MERGED evidence visible in the frozen repository/merged record, E-DEPOSIT evidence frozen in prior Zenodo research objects, and E-TRANCHE current-tranche claims whose exact final implementation was not independently visible on the inspected main snapshot. This is first-party systems research; independent third-party replication is not claimed.