FAQ

FAQ

Is this a framework?

The workspace crates provide folds, small host ports, and optional adapters. The host owns its executor and transcript; an adapter may manage seat sessions. See Architecture for what each crate provides.

Is this actually a hive mind, or is that just branding?

The mechanisms are the literal ones. Stigmergy is Grassé's: a deposit in a shared medium is the stimulus for the next unit of work, and no agent addresses another. Salience decays exponentially the way trail pheromone evaporates. Quorum sensing counts distinct grounded supporters inside a window rather than scoring an option, which is how a honeybee swarm settles a nest site. Cross-inhibition silences an advocate rather than debiting an option, which is the swarm's stop signal. The bid for the floor is the response-threshold model of division of labour, which is also Pandemonium's decision demon from the other direction. Further reading has the sources.

What it is not is an emergent-intelligence claim. See the question below about whether a room gives better answers.

Can a hive episode run turns concurrently?

Yes. HiveStep::Speak carries a bounded round of turns and one state to commit after all of them are appended. The default uses width four during the blind opening and width one after peers become visible. A host may use width one throughout for sequential turns. The bounds and visibility rules are in Hive episodes and ADR 0014.

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 module?

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.