Skip to content
HVMactive

Hivemind

Strategic analytical software for groups making consequential decisions.

Hivemind is software that helps teams reason through hard strategic decisions. It breaks a question into testable sub-claims, attaches weighted evidence to each, and rolls the result up into a structured brief — citations, confidence intervals, dissents — that a team can argue with and update without redoing the work. It exists because the memos behind consequential decisions usually hide their assumptions; a graph that lays them out is one a team can attack.

Hivemind, briefly

phase 1 of 4

What Hivemind is

Hivemind is strategic analytical software for operators reasoning under uncertainty — founders, analysts, and decision-makers who have to commit to a course of action when the evidence is incomplete and the stakes are real.

The one-line pitch: instead of a memo full of confident-sounding prose, you get a structured argument you can defend, attack, and update — with the assumptions named and the evidence attached.

What is Hivemind

Hivemind is strategic analytical software for operators who must reason under uncertainty. It decomposes a strategy question into a tree of evaluable sub-claims, attaches weighted evidence to each branch, and aggregates probabilistic posteriors back to the root. The output is a structured brief — citations, intervals, dissents — not a chat transcript.

The thesis is plain. Decision-grade analysis benefits from explicit structure rather than freeform prose. A memo that hides its assumptions is one nobody can argue with. A graph that lays them out is one a team can attack, defend, and update without redoing the work.

The product

The working surface is a hypothesis tree. An operator drops a strategy question in plain prose; Hivemind parses for entities and decisions and returns an initial frame. From there the workflow runs six steps: intake, decompose, gather evidence, score, red team, render brief. Each step has its own artifact — frame, tree, ledger, posterior, critique, memo — and each is auditable independently of the others.

Capabilities currently shipping include hypothesis decomposition (CAP-01) and the evidence ledger (CAP-02), where sourced citations are weighted by provenance and recency. Scenario trees (CAP-03) and counterfactual simulation (CAP-04) are in alpha: branch probability over outcomes, hold inputs constant, vary one assumption at a time. Adversarial review (CAP-05) and the briefing render (CAP-06) are scoped and not yet shipping.

The graph compounds. The longer an operator works inside Hivemind, the harder the working knowledge is to redo elsewhere; the moat is the accumulated structure, not the model behind it.

Why coherence matters

Hivemind sits on a coherence engine. Coherence here is not “does the text sound confident” — it is a layered, falsifiable score over a structured argument. Five layers contribute under fixed weights: contradiction (S1, 0.30), argumentation (S2, 0.20), embedding similarity (S3, 0.20), compression (S4, 0.15), and structural connectivity (S5, 0.15). Each layer returns a number a human can inspect: which proposition pairs were flagged, the size of the grounded extension, whether a fixed-point iteration found a cycle, the longest support chain.

Anti-gaming is a separate term. Template overlap, prior-corpus echo, contradiction denial, repetitive filler, and fluency-without-content are weighted and subtracted from a clamped floor. The composite is the layered score multiplied by the anti-gaming score; the brief that sounds smartest is not the brief that wins.

The product layer above this is comparative. A pitch is scored against the median coherence of incumbents in its primary domain. The interesting number is domain-relative: a pitch can be coherent in absolute terms and still under-perform the incumbents in its own market. Operators making a real bet should know which one they are looking at.

What it isn’t

Hivemind is not a chat assistant with citations bolted on. The primitive is the hypothesis, not the message. There is no transcript to scroll; there is a graph to walk, and every node is independently rebuttable.

It is not a forecasting market. Posteriors come from evidence and structure, not from crowd betting; the surface that matters is the one a single operator can defend in writing.

It is not a single model. The coherence engine combines a contradiction-detection backend, a graph extension procedure, an embedding similarity check, a compression ratio, and a structural connectivity score. Swapping any one of those backends does not change the contract the engine makes with its caller.

Status

Hivemind is in active development; the company is at seed stage as of 2026-04. Hypothesis decomposition, the evidence ledger, and the workflow scaffold are usable end-to-end. Scenario branching and counterfactual simulation are in alpha against a small set of operator partners. Adversarial review and the briefing render are scoped and not yet shipping. A recorded walkthrough is pending.