The architectureof Hivemind.

A session starts with a strategic question and two visible controls.sufficiency sets how many distinct answers must survive deliberation. feasibility sets the minimum score an answer must clear before it is accepted.

The run moves through three stages: the Theory Network explores, the Connectivity Agent synthesizes, and the Practicality Network scores. The answer is never separated from the trace that produced it.

The Theory Network is a panel of framework-conditioned agents. Each agent proposes an answer, critiques peer answers, and revises from a distinct strategic lens.

The agents retrieve from versioned framework indexes rather than a hidden prompt bundle. Citations resolve to works in the knowledge base, so the source of a claim stays inspectable.

The Connectivity Agent compresses the theory output into a smaller set of candidate answers. It clusters convergence, preserves provenance, and prepares the handoff to feasibility scoring.

The work is active synthesis, not passive observation.

The Practicality Network scores each candidate answer against real constraints: legal, financial, operational, and reputational by default. Deployments can add domain-specific constraints.

Each score carries a short rationale. If the run misses the feasibility threshold, Hivemind can regenerate with those constraints attached to the next deliberation.

Output is a small set of recommendations with scores, rationales, supporting frameworks, and links back to the underlying deliberation.

The audit trail records prompts, agent utterances, critiques, revisions, synthesis events, scores, and regeneration decisions. That trail is what makes the recommendation defensible.

Hivemind can run as a hosted web app or as a native deployment. For enterprise customers, the corpus and audit-trail store can stay inside the client's network boundary.