Story 002: Federated Peer Learning and Consensus Correction

Current Baseline Used by This Story

This story no longer assumes a public thematic channel plus an ad-hoc side channel. It is grounded in the current Orbiplex corpus where:

• a signed question-envelope opens or binds a live answer room,
• the live room is the main place where peers compare, challenge, and refine answers,
• exposure mode and room policy profile constrain who may observe, join, summarize, or export later traces,
• disagreement handling produces explicit outcomes (confirmed, corrected, unresolved) rather than ambient chat residue,
• local learning is policy-gated and provenance-rich instead of being an immediate, silent side effect of receiving peer messages,
• later archival and training paths remain separate from immediate answer serving.

This story follows the learning and correction loop that may happen after the question flow described in story-001.md and story-004.md has already opened a room.

Sequence of Steps

1. A local node opens a swarm question through the current baseline flow and receives one or more candidate answers in the answer room bound to the question/id.
2. The asking node, a participating peer, or a secretary notices that one answer materially conflicts with:
3. another candidate answer,
4. local retrieval evidence,
5. federation policy or known procedure,
6. or the node's own specialization-specific knowledge.
7. Instead of opening a separate public correction channel, participants continue the review inside the same question-bound room or in a tightly linked review path that preserves the original question/id, room scope, and participant provenance.
8. Peers post counter-evidence, implementation notes, examples, and objections. A secretary or other designated node may emit one or more intermediate summaries so the room does not have to treat raw message history as the durable source of truth.
9. If room policy allows transcript observation, a transcription monitor captures the relevant discussion with explicit provenance, visibility scope, consent or policy basis, and human-origin markers where applicable.
10. The asking node or delegated secretary classifies the emerging outcome for the disputed claim as one of:
11. confirmed,
12. corrected,
13. unresolved.
14. If the room converges on a better answer, the accepted correction is emitted as a new durable outcome linked to the original question, such as:
15. a signed room summary,
16. a corrected response envelope,
17. or another accepted answer artifact allowed by current room policy.
18. The local node records the outcome together with enough provenance to reconstruct:
19. which question triggered the learning event,
20. which nodes participated,
21. which evidence or summaries influenced the correction,
22. whether any human-linked input affected the accepted result.
23. The node MAY promote confirmed or policy-accepted corrected material into local retrieval assets such as:
24. vector memory,
25. indexed files,
26. a local knowledge-artifact queue.
27. Material classified as unresolved is kept isolated from trusted retrieval by default. It may still be preserved for later review, adversarial testing, or curator inspection, but it is not silently promoted as trusted knowledge.
28. If the discussion is judged culturally or operationally valuable and policy allows it, later layers MAY turn the transcript and its accepted summaries into curated corpus entries or vault material. That promotion happens through explicit curation, not through the mere fact that peers debated in a room.
29. If training is desired, the node or federation routes only explicitly approved corpus material into later specialization jobs. Raw discussion and unresolved corrections do not directly become training data.

Open Continuation

• Exact divergence threshold for opening a correction path.
• Default consensus rule and tie handling for confirmed vs corrected.
• Whether adversarial or unresolved material should feed separate debate-style evaluation datasets.
• The exact schema set for divergence signals, consensus outcomes, and local knowledge-artifact promotion.