Expand the frames you bring to a choice.

You ask a question with the options you can articulate. The question fans out to two parallel panels: a curated set of AI models, and a network of other humans. Each panelist reasons independently, and you see how they distribute across your options along with a measure of how aligned they were.

Two outputs come from this. The first is expanded framing: the disagreement among panelists tells you which considerations matter to which kinds of minds, which framings are in tension, and where your question was incomplete. The second is prediction: when a question's window closes, the asker marks what turned out accurate and the community ratifies. Over time, this builds a track record — on questions, on ratifiers, and on which classes of question are genuinely tractable.

Framing and prediction reinforce each other. A well-framed question is one where the panel's confidence reliably predicts the eventual outcome. A poorly-framed question is one where strong consensus turns out wrong, or where the panel can't converge at all. Tracking prediction is how we test whether framings work.

The first approach is asynchronous and built around human input. You write a question and the options you can articulate, and it fans out to two parallel panels — a curated set of AI models, and a network of other humans. Each panelist responds independently by selecting from your options, and a confidence number summarizes how aligned they were.

You do not wait for it live. Responses accumulate while the question is open; you come back to a distribution. When the window closes, you mark what turned out accurate and the community ratifies — building a track record on questions, on ratifiers, and on which kinds of question are tractable at all.

This is the prediction layer. Its value is in the spread and the outcome: where the panels converge, where they split, and whether that confidence held up against reality.

The second approach is synchronous. Instead of waiting on a distribution, you hold a live conversation with a panel of distinct voices — each a different temperament and way of reasoning. You bring something you are turning over, and they respond in real time, in their own registers, building on and pushing against one another the way a good group of advisors would.

An unseen orchestrator decides who speaks to what you have said, so the right voice answers the right moment rather than all of them at once. You can shape the panel — add or remove voices, including raw frontier models that answer as themselves — and you can let the panel carry the discussion among itself while you watch, stepping back in whenever you like.

Where Ask the Hive gives you a measured distribution to come back to, Dialogue gives you the live texture of minds working a problem — the questions you were not asking, voiced as they occur. The two are different tools for different moments: one to poll and predict, one to think out loud with company.

The current implementation borrows from several traditions: decision theory (formal frameworks for choice under uncertainty), Bayesian inference (how priors update on evidence), behavioral economics and psychology (known biases in human judgment), and the older literature on collective deliberation and shared cognitive substrates. None of these is the dominant frame. Each catches a different way an individual's framing of a decision is incomplete.

This stack is what we are using now to design how the panels reason and how the aggregation works. It is not what we permanently commit to. As Hivepsyche evolves — local models trained on different corpora, new aggregation schemes, additional ratification signals — the intellectual stack will expand or shift.

Hivepsyche surfaces inputs you did not have and disagreements you did not see. It does not replace your judgment. We will not tell you what to do.

Where the panels split, we will not paper over the split. Where community ratification points one way, we will report it; we will not promise it is true. The system produces a richer view of a question. The decision is still yours.

We are also explicit about what the panels are. The AI panel is a curated set of language models, not oracles. The human panel is other Hivepsyche users, not vetted experts. Each panel has characteristic strengths and characteristic failure modes, and the ratification system exists in part to expose both over time.

The signals we are tracking:

1. Whether panel disagreement reliably predicts decision difficulty — confident panels right at a meaningfully higher rate than split panels.
2. Whether community ratification converges with expert ratification on a sampled subset.
3. Whether a synchronous dialogue, held before asking the Hive, produces measurably better framings — measured by whether panel confidence becomes a tighter predictor of ratified outcomes when a dialogue precedes the poll.
4. Whether perspective-specific fine-tuned models outperform generalist baselines on both prediction accuracy and the framing-improvement signal.

Each of these is testable. We will publish what we find, including the failures.

This page will be updated as the data comes in.