Preregistered Belief Revision Contracts

Dimensions overall score 7.0

• Deliberative multi-agent systems, including recent LLM-based agent societies, allow agents to ex- change messages and revise beliefs over time. While this interaction is meant to improve performance, it can also create dangerous conformity effects: agreement, confidence, prestigeclaim 82%
  Evidencepartial

  Deliberative multi-agent systems, including recent LLM-based agent societies, allow agents to ex- change messages and revise beliefs over time. While this interaction is meant to improve performance

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

• 2 X h∈H |b(h)−b ′(h)|. 3.2 Belief state as a router-side artifact (LLM instantiation) The formal object bt i ∈∆(H) is aprotocol stateused for deliberation, enforcement, and auditing. In deployedSource proofclaim 82%
  Evidenceverified

  2 X h∈H |b(h)−b ′(h)|. 3.2 Belief state as a router-side artifact (LLM instantiation) The formal object bt i ∈∆(H) is aprotocol stateused for deliberation, enforcement, and auditing. In deployed

  Implicationmissing

  Implication not extracted yet.

  Verificationverified

  verified

• or that self-reported probabilities are trustworthy. PBRC does not require these assumptions: a router/orchestrator can maintain an authoritative belief state and compute contract updates itself (Section 4.4), while agents contribute candidate hypotheses, arguments, and evidence tokens. Finite candidate setsclaim 82%
  Evidencepartial

  or that self-reported probabilities are trustworthy. PBRC does not require these assumptions: a router/orchestrator can maintain an authoritative belief state and compute contract updates itself (Section 4.4)

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

• (triggered by evidence) expands H and reinitializes b over the expanded set. All of our “social-only cannot force substantive revision” guarantees apply within each stage, and the introduction of new hypotheses becomes auditable and evidence-gated. Takeawayclaim 82%
  Evidencepartial

  (triggered by evidence) expands H and reinitializes b over the expanded set. All of our “social-only cannot force substantive revision” guarantees apply within each stage

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

• deterministic mapping from structured tool outputs to hypotheses. For retrieval tokens, a simple auditable pattern is: (i) canonicalize h into a natural-language claim claim(h); (ii) run a pinned NLI model on (premise=snippet,hypothe- sis=claim(h))claim 82%
  Evidencepartial

  deterministic mapping from structured tool outputs to hypotheses. For retrieval tokens, a simple auditable pattern is: (i) canonicalize h into a natural-language claim claim(h)

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

• 2. each Ui,j is a (possibly partial) revision operator that maps a current belief state and the current event to a next belief state, Ui,j : ∆(H)×Evt⇀∆(H), withEvtthe set of events (multisets of messages); 3.≺ i is a strict priority order on {1, . . .claim 82%
  Evidencepartial

  2. each Ui,j is a (possibly partial) revision operator that maps a current belief state and the current event to a next belief state, Ui,j : ∆(H)×Evt⇀∆(H), withEvtthe set of events (multisets of messages); 3

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

Competitive landscape

This work introduces a protocol for multi-agent systems that prevents dangerous conformity by strictly separating communication from belief revision, ensuring auditable evidence for all changes.