FAQ

FAQ

Is this a framework?

No. There is no runtime to adopt, no executor to inherit, no trait object graph to fill in. It is three crates of folds plus three ports, and the widest of those ports has one method.

Why is there no fan-out?

Because the failure mode of a naive hive mind is a message that starts N turns with no approval anywhere in sight. HiveStep::Speak carries exactly one HiveTurn and there is no variant that carries two, so the rule holds as a type invariant.

There is also an empirical reason. Sparse communication topologies match or beat fully connected ones in multi-agent debate at much lower cost, and parallel fan-out wins only on genuinely decomposable work. The one thing fan-out buys that matters here is independence, and that is bought as Visibility::Blind, a filter on the projection, for the cost of a flag. See ADR 0002.

Does a room of agents actually give better answers?

That is not a claim this repository makes. Almost every positive multi-agent result in the literature is confounded by compute: the multi-agent arm just spent more. The benchmark here includes the control that settles it, an independent vote at a matched budget, and the deliberation beats it by 1.2 to 3.2 points across desk sizes while spending about half the turns.

That is a real margin, and it is a small one. What the protocol reliably provides is a bounded episode that terminates for a reason you can audit. See Benchmarks, including the section on what the numbers do not show.

Why can't it own a database?

Because messages are addressed by sequence number across surfaces the host owns: reactions, board cards, run rows. A second append-only log could not be made consistent with the first, so there is not one. The host lends its log through SessionLog and keeps ownership of it.

Why no floating point in the hive crate?

So every payload derives Eq and every fold reproduces exactly. Scores are fixed point in thousandths. A deliberation you cannot replay is a deliberation you cannot audit.

Can I use just the pure part?

Yes. Depend on tinyhivemind-core directly and you get desks, rosters, mentions, conversation identity, and the responder ladder's decisions, with no ports to implement and nothing to await. tinyhivemind re-exports it for hosts that want the paging walk too.

How do I get a version?

You pin a commit. Every crate is publish = false, and a tag would be a weaker second name for the commit you already pinned. See Host integration.

What quorum threshold should I use?

A majority of the desk, and never the whole of it: min(agents / 2 + 1, agents - 1). Below a majority, rooms deadlock because two options can both carry. At unanimity, one grounded objection makes quorum unreachable for the rest of the episode. Episode policy has the numbers on both.

My agents write malformed markers.

Expected, and it is a prompting problem rather than a parsing one. The four specific failures real models produce, and the fix for each, are in Trace grammar.

Where did this come from?

It is being extracted from opencompany, where the same layer had two defects: a first-person-collapsed transcript, so agents read their peers' replies as their own, and a transcript that was only re-read on a rebind, so an agent missed a peer's interleaved reply. Both are fixed here. See ROADMAP.md.