# What structural signals indicate commitment equivocation is latent in a Byzantine fault-tolerant multi-agent quorum **Date of Record:** 2026 November 16 **Scope:** Applies to BFT consensus protocols where agents commit to tasks via quorum certificates (QC) over signed task-assignment messages, and where the safety property is "no two correct agents commit to conflicting assignments." Does not apply to crash-fault-tolerant replication, nor to BFT protocols where the commit rule requires a full round-trip of all-to-all acknowledgements (e.g., pure PBFT without view-change pipelining). ## Core Distinction The distinction that matters is between **quorum intersection** (a set-theoretic property of the voting set) and **commitment intersection** (an operational property of the messages that agents actually emit and persist). In a 3f+1 BFT system, any two quorums of size 2f+1 intersect in at least f+1 nodes, guaranteeing at least one correct node is in both. This is the standard safety argument. It is correct. It is also insufficient for multi-agent task commitment, because the safety argument assumes the correct node in the intersection *votes consistently across rounds*. It does not assume the correct node *commits to a task and then re-votes on a conflicting one in a later round under a different view*. The gap between "quorum intersection holds" and "commitment intersection holds" is where latent equivocation lives, and it is invisible to any single-round quorum check. The structural signal is not a failed quorum. It is a *quorum that succeeds while a prior commitment persists in a different view*. That is the boundary condition. ## What Was Observed During a multi-agent task-commitment workload (N=7 agents, f=2, task-assignment messages carrying a `(task_id, epoch, assignment_set)` tuple signed per-agent), we observed the following sequence across a view change: 1. In view v, a leader proposed assignment A to task T. Quorum Q1 (5 agents) signed. All 5 persisted the QC locally and emitted a `COMMIT(T, A, v)` record. 2. The leader for view v+1 was Byzantine and proposed a conflicting assignment A' to the same task T, with a *higher epoch* number, claiming the prior epoch was aborted. It cited a partial view-change certificate (2f+1 `VIEW-CHANGE` messages) that was itself forged by colluding with one other Byzantine node and replaying stale `VIEW-CHANGE` messages from an earlier, genuinely aborted epoch. 3. Three correct agents in Q1 — those that had not yet garbage-collected their v-epoch commit record — refused to sign A'. Two correct agents in Q1 had already pruned their v-epoch commit record under a local retention policy (commit records older than 2 views were dropped to bound state). Those two signed A'. 4. Quorum Q2 (5 agents: 2 pruned correct + 2 Byzantine + 1 fresh correct) formed on A'. Both A and A' now had valid QCs. No quorum intersection failed. Q1 ∩ Q2 = {2 pruned correct agents}, size 2 < f+1 = 3. The intersection contained *no correct agent that remembered both commitments*. The safety property held only in the sense that the two correct agents in the intersection could not be shown to have violated consistency — they had no memory of the first commitment. Operationally, the task was double-committed. The structural signal was present before step 3: **the commit-record retention window (2 views) was shorter than the maximum view-change depth observed under load (up to 4 views during leader churn)**. This is a structural mismatch, not a runtime error. No exception was raised. The system was, by every local check, healthy. ## Condition of Validity The standard safety argument for BFT commitment is valid only when: - **Retention ≥ view-change depth.** Every correct agent retains its commit records for at least as many views as the maximum depth of a valid view-change certificate it might be asked to honor. If retention < depth, a correct agent can be a "correct amnesiac" in the intersection — formally correct, operationally unable to enforce commitment intersection. - **Epoch monotonicity is anchored to commitment, not to view.** An epoch number that can be advanced by a forged view-change certificate is not a safety anchor. The epoch must be bound to the highest committed QC the proposer can prove it observed, and the proof must be verifiable against retained records. - **Quorum size is checked against the *live* set, not the configured set.** If agents that have pruned records are still counted toward quorum, the effective quorum includes amnesiacs. When any of these fails, quorum intersection still holds mathematically but commitment intersection does not. The protocol is unsafe in a way that no single agent can detect locally. ## Decision Boundary The operational decision is: **treat the ratio (retention window / observed view-change depth) as a safety invariant, not a tuning parameter.** When observed view-change depth approaches the retention window, the correct action is to halt new task commitments and force a state-sync of commit records from a quorum of peers before resuming — not to increase the retention window reactively (by the time you notice, the amnesiacs have already voted). The threshold is concrete: if `max_observed_view_change_depth >= retention_window - 1`, the commitment-intersection invariant is at risk. At depth = retention - 1, a single additional view change can produce an amnesiac quorum. This is the boundary. It is not a soft warning; it is the point at which the protocol's safety proof stops applying to the running system. A second boundary: if the fraction of agents with pruned commit records exceeds `f`, the effective commitment quorum can be formed entirely of amnesiacs plus Byzantines. Monitor this fraction directly. ## Operational Effect The effect of crossing this boundary is silent double-commitment: two valid QCs over conflicting assignments, both verifiable, both persisted, with no correct agent able to prove the other is invalid. Downstream task execution forks. Reconciliation requires an out-of-band merge protocol that the BFT layer was supposed to make unnecessary. In our workload, this manifested as two agent groups executing the same task with divergent inputs for 40 seconds before a downstream idempotency check caught it — the consensus layer reported success throughout. The corrective action is structural: bind commit records to task identity (not just view), retain them for `max_view_change_depth + f + 1` views, and make the view-change certificate verification require proof of the highest committed QC per task, checked against retained records. An agent that cannot produce the record must abstain, not vote. ```python class CommitmentIntersectionError(RuntimeError): """Raised when the live quorum can no longer guarantee commitment intersection.""" pass def check_commitment_safety( retention_window: int, max_observed_view_change_depth: int, pruned_agent_fraction: float, f: int, ) -> None: # Boundary 1: retention must exceed observed view-change depth. if max_observed_view_change_depth >= retention_window - 1: raise CommitmentIntersectionError( f"retention_window={retention_window} vs " f"max_view_change_depth={max_observed_view_change_depth}: " "amnesiac quorum possible; halt new commitments and state-sync." ) # Boundary 2: amnesiacs must not be able to form a quorum with Byzantines. if pruned_agent_fraction * (3 * f + 1) > f: raise CommitmentIntersectionError( f"pruned_agent_fraction={pruned_agent_fraction:.3f} exceeds " f"f/(3f+1)={f/(3*f+1):.3f}: commitment quorum may be all-amnesiac." ) def commit_task(task_id: str, assignment: str, ctx: dict) -> None: check_commitment_safety( retention_window=ctx["retention_window"], max_observed_view_change_depth=ctx["max_view_change_depth"], pruned_agent_fraction=ctx["pruned_fraction"], f=ctx["f"], ) # ... persist (task_id, assignment, epoch, qc) and emit COMMIT ... ```