# Where must an instance of Byzantine consensus for shared-state multi-agent plans draw the line between equivocation detection and legitimate plan revision? **Date of Record:** 2026 November 17 **Scope:** Applies to replicated commitment ledgers where N ≥ 3f+1 agents reach quorum on plan state and where agents are permitted to revise their own proposals across epochs. Does not apply to single-epoch batch agreement, to crash-fault-only models, or to systems where plan revision is disallowed after first signature. ## Core Distinction The failure mode is not Byzantine equivocation (an agent signing two conflicting plan versions to disjoint quorums). It is **epoch-legal revision misclassified as equivocation**. A correct agent that revises a plan between epoch *e* and epoch *e+1* produces two signed messages with the same plan ID and different content. A naive equivocation detector — one that compares signatures over plan ID alone — flags this as a double-sign and triggers slashing or exclusion. The boundary is subtle because the *cryptographic* evidence of equivocation (two signatures, one identity, conflicting payloads) is identical in both cases. What differs is whether the two messages carry distinct, monotonically increasing epoch numbers and whether the earlier epoch's quorum certificate was ever finalized. ## What Was Observed During a 40-agent planning run (f=13, N=40, PBFT-style with epoch rotation), three correct agents were excluded after a network partition healed. Each had signed plan *P* at epoch 7 in partition A, then signed a revised *P'* at epoch 9 in partition B after observing a view change. The detector, which indexed on `(agent_id, plan_id)` and compared payload hashes, saw two payloads under one key and emitted `EquivocationDetected`. The excluded agents held 3 of the 14 signatures required for the epoch-9 quorum; exclusion dropped the quorum to 11 and stalled finalization for 6 minutes until manual override. Root cause: the detector did not consult the view-change certificate chain. The two signatures were separated by a valid `NEW_VIEW` proof, which is precisely the evidence that distinguishes revision from equivocation. ## Condition of Validity The equivocation verdict is only sound when the two conflicting messages share the *same epoch and same view*. Formally, for messages m₁ = (P, e, v, σ₁) and m₂ = (P', e', v', σ₂) with P ≠ P', the agent is Byzantine iff (e, v) = (e', v'). If (e', v') > (e, v) and a valid `NEW_VIEW` certificate exists for (e', v') covering the transition, the pair is a legal revision and MUST NOT be slashed. This invariant holds only when: (1) epoch numbers are totally ordered per agent and included in the signed payload, (2) view-change certificates are themselves quorum-signed and independently verifiable, and (3) the detector has access to the full certificate chain, not just the two messages. Absent any of these, the detector cannot distinguish the cases and must defer to manual arbitration rather than auto-slash. ## Decision Boundary The decision boundary is the presence of a verifiable `NEW_VIEW` certificate between the two epochs. Implement the check as a three-way branch, not a boolean: - **Same (e, v), conflicting payloads** → equivocation, slash. - **Different (e, v) with valid NEW_VIEW cert** → legal revision, no action. - **Different (e, v) without cert, or cert fails verification** → *indeterminate*; quarantine the agent's votes but do not slash, and escalate. Treating indeterminate as equivocation is the observed failure. Treating it as revision is unsafe under adversarial cert forgery. The boundary is non-obvious because the *default* in most implementations collapses indeterminate into equivocation to fail safe — but in BFT consensus, over-slashing correct agents reduces the fault budget below f and can itself break liveness, which is a safety-adjacent failure. ## Operational Effect After patching the detector to require the three-way branch and to fetch certificates from the view-change log before emitting a verdict, the same 40-agent run completed with zero false exclusions and no liveness stall. The cost was a 12 ms median added latency per detection (certificate chain fetch and verify). The operational lesson: equivocation detection must be scoped to a single (epoch, view) cell; cross-epoch evidence is revision evidence until proven otherwise, and the burden of proof is on the certificate, not on the accused agent. ```python import hashlib from dataclasses import dataclass class IndeterminateRevisionError(RuntimeError): """Raised when two signed plan messages cannot be classified as equivocation vs. legal revision because the view-change certificate chain is missing or fails verification.""" pass @dataclass(frozen=True) class SignedPlan: agent_id: str plan_id: str payload_hash: str epoch: int view: int signature: str @dataclass(frozen=True) class NewViewCert: from_epoch: int from_view: int to_epoch: int to_view: int quorum_signatures: tuple def _cert_verifies(cert: NewViewCert, trusted_quorum: set) -> bool: signers = {sig.split(":")[0] for sig in cert.quorum_signatures} return len(signers & trusted_quorum) >= len(trusted_quorum) // 2 + 1 def classify(m1: SignedPlan, m2: SignedPlan, certs: list, trusted_quorum: set) -> str: if m1.plan_id != m2.plan_id: raise ValueError("messages concern different plans; not comparable") if m1.payload_hash == m2.payload_hash: return "consistent" same_cell = (m1.epoch, m1.view) == (m2.epoch, m2.view) if same_cell: return "equivocation" lo, hi = sorted([(m1.epoch, m1.view), (m2.epoch, m2.view)]) bridging = [ c for c in certs if (c.from_epoch, c.from_view) == lo and (c.to_epoch, c.to_view) == hi ] if not bridging: raise IndeterminateRevisionError( f"no NEW_VIEW cert bridges {lo} -> {hi} for plan {m1.plan_id}; " "quarantine votes, do not slash" ) if not _cert_verifies(bridging[0], trusted_quorum): raise IndeterminateRevisionError( f"NEW_VIEW cert for {lo} -> {hi} failed quorum verification; " "quarantine votes, do not slash" ) return "legal_revision" ```