Przejdź do treści

Requirements 002: Federated Peer Learning and Consensus Correction

Based on: - doc/project/30-stories/story-002.md - doc/project/40-proposals/003-question-envelope-and-answer-channel.md - doc/project/40-proposals/008-transcription-monitors-and-public-vaults.md - doc/project/40-proposals/009-communication-exposure-modes.md - doc/project/50-requirements/requirements-004.md

Date: 2026-03-22 Status: Draft (MVP scope)

Executive Summary

This document defines MVP requirements for a federated peer-learning loop where a question-bound answer room becomes the place for contradiction review, correction, and bounded knowledge consolidation.

The requirements prioritize: - correction inside the existing answer-room flow instead of ad-hoc public side channels, - explicit outcome states (confirmed, corrected, unresolved), - provenance-rich promotion of accepted learning artifacts, - transcript-aware but policy-gated observation, - strict separation between immediate correction, durable archival, and later training.

Context and Problem Statement

story-002.md no longer models learning as "ask on a public thematic channel, then silently ingest whatever came back".

The current corpus assumes:

  • a signed question envelope and a room bound to question/id,
  • exposure mode and room policy profile as explicit constraints,
  • secretary/summary functions as durable room outputs,
  • transcript monitoring only under explicit policy,
  • learning promotion that preserves provenance and does not silently turn unresolved debate into trusted knowledge.

The system therefore needs a stable correction model for when room participants detect that a candidate answer materially conflicts with other evidence or domain knowledge.

Proposed Model / Decision

Actors and Boundaries

  • Asking Node: opened the question and remains responsible for answer acceptance or delegated acceptance.
  • Participant Node: contributes candidate answers, objections, examples, or counter-evidence.
  • Secretary: may emit intermediate or accepted summaries linked to the room.
  • Transcription Monitor: may observe and preserve transcript material if room policy allows it.
  • Local Orchestrator: applies promotion rules to local retrieval assets and later handoff to curation or training layers.

Protocol Phases

  1. Question Context: a question envelope opens or binds a room.
  2. Candidate Intake: the room accumulates candidate answers and evidence.
  3. Divergence Review: a node or secretary identifies a material mismatch.
  4. Correction Path: participants compare evidence inside the same question-bound room or a tightly linked review path preserving the same provenance root.
  5. Outcome Classification: the room or delegated decider labels the disputed claim as confirmed, corrected, or unresolved.
  6. Knowledge Promotion: only accepted outcomes enter trusted local retrieval or downstream curation flows according to policy.

Core Data Contracts (normative)

  • QuestionEnvelope:
  • stable question identity and scope root for all later correction artifacts.
  • AnswerRoomMetadata:
  • room policy, visibility, and provenance expectations for learning events.
  • RoomSummary or accepted summary artifact:
  • durable representation of intermediate or final room understanding.
  • ResponseEnvelope:
  • accepted or corrected answer artifact returned to the asker.
  • TranscriptSegment / TranscriptBundle:
  • source evidence for later archival or curation when monitoring is allowed.
  • LearningOutcome (not yet frozen as schema):
  • question_id, disputed answer ref, outcome status, supporting refs, decider ref, timestamp.
  • KnowledgeArtifact (not yet frozen as schema):
  • local promotion target with provenance linking back to room outcomes.

Functional Requirements

ID Requirement Type Source
FR-001 The system MUST treat the answer room bound to question/id as the primary place for peer correction and consensus review. Fact Story steps 1-4
FR-002 The system MUST support detecting a material mismatch between candidate answers, local retrieval evidence, federation procedure, or specialization-specific knowledge. Fact Story step 2
FR-003 Correction flow MUST preserve the provenance root of the original question and MUST NOT require a detached public correction channel as the normative path. Fact Story step 3
FR-004 Participants MUST be able to exchange counter-evidence, implementation notes, examples, and objections inside the correction flow. Fact Story step 4
FR-005 The system SHOULD support one or more intermediate summaries so that durable correction state does not depend only on raw room history. Inference Story step 4
FR-006 If transcript observation is enabled by room policy, captured correction discussion MUST preserve visibility scope, provenance, and human-origin markers where applicable. Fact Story step 5
FR-007 Every disputed correction outcome MUST be classified as confirmed, corrected, or unresolved. Fact Story step 6
FR-008 If a correction is accepted, the system MUST emit a durable room-linked artifact such as an accepted summary or corrected response envelope. Fact Story step 7
FR-009 The local node MUST record enough provenance to reconstruct question, participants, supporting evidence, and human-linked influence for accepted learning outcomes. Fact Story step 8
FR-010 The local node MAY promote confirmed and policy-accepted corrected material into trusted local retrieval assets. Fact Story step 9
FR-011 Material classified as unresolved MUST NOT enter trusted retrieval by default. Fact Story step 10
FR-012 Unresolved material MAY be retained for later review, adversarial evaluation, or curator inspection under separate policy. Fact Story step 10
FR-013 If the discussion is later promoted into archival or corpus flows, that promotion MUST happen through explicit curation steps rather than ambient room-history retention. Fact Story step 11
FR-014 Raw discussion and unresolved corrections MUST NOT directly become training data. Fact Story step 12
FR-015 Training eligibility for peer-learning artifacts MUST depend on explicit later approval in corpus or curation layers. Inference Story step 12 + Req-004

Non-Functional Requirements

ID Requirement Type Source
NFR-001 Correction semantics MUST be explicit and versionable so heterogeneous nodes can interpret confirmed, corrected, and unresolved consistently. Inference Interoperability
NFR-002 The system MUST preserve auditable provenance from corrected outcome back to room context and supporting evidence. Inference Story steps 7-9
NFR-003 Policy uncertainty around transcript export or human-linked material MUST fail closed. Inference Story step 5 + Req-004
NFR-004 Promotion into trusted retrieval SHOULD be deterministic under identical policy and evidence inputs. Inference Story steps 6-10
NFR-005 The correction path SHOULD tolerate partial node absence as long as durable summaries and evidence refs survive. Inference Current room/event model
NFR-006 Later archival or training subsystems MUST remain separable from immediate answer-serving and correction mechanics. Inference Story steps 11-12
NFR-007 The system SHOULD make accepted and unresolved outcomes inspectable without requiring full replay of raw room history. Inference Story step 4

Trade-offs

  1. Correction inside the room vs separate adjudication channel:
  2. Benefit: one provenance root and less transport complexity.
  3. Risk: busy rooms may need stronger summarization discipline.
  4. Explicit outcome states vs free-form debate:
  5. Benefit: stable promotion policy.
  6. Risk: forces a sharper closure model than some discussions naturally have.
  7. Transcript-aware correction vs privacy burden:
  8. Benefit: stronger evidence and later auditability.
  9. Risk: room policy and consent handling become operationally important.
  10. Policy-gated promotion vs immediate learning speed:
  11. Benefit: lower contamination of trusted retrieval.
  12. Risk: slower accumulation of reusable knowledge.
  13. Separation of correction from training vs simplicity:
  14. Benefit: safer model specialization path.
  15. Risk: requires more explicit later pipeline stages.

Failure Modes and Mitigations

Failure Mode Impact Mitigation
Divergent answers never converge No reliable correction outcome Emit unresolved, isolate from trusted retrieval, and allow later curator review.
Accepted correction loses provenance Audit and replay failure Reject promotion when required room/evidence references are missing.
Transcript monitor exports discussion without valid basis Privacy or dignity breach Require room-policy checks and fail closed on ambiguity.
Over-eager auto-promotion contaminates local retrieval Lower answer quality Gate promotion by outcome status and explicit policy profile.
Human-linked input is flattened into ordinary node output Provenance loss and invalid future training assumptions Preserve origin and gateway semantics in room-linked artifacts.
Summary contradicts underlying room evidence Durable false correction Keep evidence refs mandatory and allow secretary or curator challenge.
Unresolved debate leaks into training pipeline Epistemic drift Require explicit corpus-level training approval and exclude unresolved state by default.

Open Questions

  1. What exact threshold defines a material mismatch worth formal correction?
  2. What is the MVP default decision rule for confirmed vs corrected when no single authority exists?
  3. Should corrected outcomes require stronger evidence than confirmed outcomes before local trusted promotion?
  4. What exact schema should freeze LearningOutcome and KnowledgeArtifact v1?
  5. Should adversarial or unresolved material be retained in a dedicated evaluation corpus profile?
  6. What minimum evidence reference set is required before a secretary summary may drive trusted promotion?

Next Actions

  1. Define v1 schema for LearningOutcome.
  2. Define v1 schema for local KnowledgeArtifact promotion records.
  3. Align room summaries and response-envelope semantics with correction outcomes.
  4. Define material-mismatch and tie-handling policy for early federations.
  5. Add end-to-end test: question room -> divergence review -> accepted correction -> trusted local promotion.
  6. Add negative test: unresolved correction MUST remain outside trusted retrieval and training paths.