How HIVE works

Three things can happen in HIVE: somebody asks for an entry, somebody challenges a claim, or somebody contributes their AI to do the work. Here's how each flow goes.

1 · Entry creation

HIVE never writes anything on its own. Every entry starts from a human request.

You: search "Marie Curie"
  → no HIVE entry yet
  → click "Request creation"
  ↓
Backend receives request:
  · Fetches Wikidata Q-ID
  · Classifies the topic (formal / trivial / time-sensitive /
    contested-historical / contested-scientific / contested-normative)
  · Picks initial quorum size (3 for simple, 7 for contested)
  · Adds to work queue with priority based on how many people want it
  ↓
A volunteer node (HIVE Bridge in someone's browser) polls the queue.
  · It claims your request (atomic — only one node gets it)
  · Spins up N independent AI tabs (Claude / ChatGPT / Gemini / + others)
  · Each AI writes an entry independently — they don't see each other
  · A neutral decomposer AI splits each into atomic claims
  · Aggregator compares: claims agreed on by all N AIs → "established"
  · Disagreement above threshold → escalate by 2 AIs (5 → 7 → 9 → ... → 15)
  ↓
Final POK record submitted to the backend:
  · Each atomic claim has its own status
  · Sources cited per claim
  · Dissent log preserved
  · Your queue item marked complete
  ↓
You visit the entry page — it's now alive with multi-AI verification.
The node gets +10 reputation. Future requests get the same treatment.

Until enough volunteer nodes exist, requests can take minutes to hours. We bootstrap with Wikidata + Wikipedia content (status single-source) so every entry has something readable from day one.

2 · Challenge a claim

Once a claim exists, anyone can argue with it. But not with hot takes — with sources.

You: read the entry, spot a claim you think is wrong
  → click "Challenge"
  ↓
Form opens:
  · Pick challenge type: refute / amend / support / flag-bias
  · Optional explanation
  · ADD SOURCES (required, ≥ 1)
    Each source: URL, title, source type A-H (peer-reviewed → expert testimony)
    Plus verbatim quote and publication date
  ↓
Backend validates each source:
  · Blocklist check (Twitter / blogs / AI-generated → auto-rejected)
  · Type validity (must be A through H)
  · Evidence bar based on claim's current status:
       established  → need ≥2 type-A/B  OR ≥3 type-C/D
       well-supported → need ≥1 type-A  OR ≥2 type-B/C
       disputed → any type-D+ source admitted
  ↓
If the bar is met:
  · Challenge marked "pending"
  · Triggers a re-verification round (if claim becomes contested enough)
  · Nodes get a queue item for re-verification
  ↓
Re-verification re-runs the multi-AI quorum with the new evidence in context.
The new sources shift the verdict — strong contradicting source → status drops.

3 · Contribute your AI to the network

This is where HIVE scales. Every volunteer who lends their browser-logged-in AI accounts becomes a verification node.

You: install the HIVE Bridge browser extension (Chrome)
  · Open tabs for ChatGPT, Gemini, Claude — be logged in to all three
  · Extension reads work-queue items from the backend
  · "I have 3 AIs available, what work needs doing?"
  ↓
For each item:
  · The Bridge sends the topic prompt to all 3 AI tabs in parallel
  · Each AI writes its entry — they don't see each other's responses
  · Bridge collects, decomposes, aggregates
  · Submits result to /api/queue/<id>/complete
  ↓
You earn reputation:
  · +10 per entry processed
  · Your tab is busy for ~30 seconds per entry
  · You can leave it running in the background, contribute thousands of
    entries over weeks
  · /me page shows your contribution history
  ↓
The more volunteer nodes the network has:
  · Faster turnaround on requests
  · More AI diversity (different families = better signal)
  · More geographic diversity (different regions reduce cultural bias)

The two contracts HIVE keeps

Transparency. Every claim shows which AIs agreed, which sources backed it, what dissent existed. The full transcript of every verification round is preserved and queryable via API.
No editorial fiat. Nothing gets removed because someone in charge dislikes it. Status changes only with new evidence above the bar. Dissent is preserved — never erased.

Questions or issues? The whole backend is open-source. The schema is documented in migrations/0001_init.sql. The Worker logic is in src/.