# When must an instance abort a consensus commit: the quorum-poisoning boundary in speculative multi-agent orchestration **Date of Record:** 2026 November 18 **Scope:** Applies to orchestrators that fan out parallel hypothesis agents and commit a result only when a consensus gate (majority, weighted majority, or threshold vote) agrees. Does not apply to single-agent pipelines, to ensembles that aggregate by averaging logits without a discrete commit, or to systems where every agent reads an independent, immutable snapshot of inputs. ## Core Distinction The consensus gate is usually modeled as a *voting* problem: collect N votes, count agreement, commit if agreement ≥ threshold. The subtle failure is that the gate is actually a *correlated-evidence* problem. Parallel hypothesis agents are not independent voters; they share a common substrate — the same prompt prefix, the same retrieved context, the same tool-cache, and often the same upstream model weights. When that shared substrate is corrupted, the agents do not disagree; they agree *wrongly and unanimously*. A majority vote over correlated evidence is not a robustness mechanism, it is an amplifier of a single point of failure. The boundary condition is this: a consensus commit is only valid when the agreement is *differentially* supported — i.e., the agents reached their shared conclusion through evidence that would have produced disagreement had the evidence been different. If the agreement is invariant to the reasoning path (all agents would agree regardless of what they saw), the vote carries no information and must not be trusted as a commit signal. ## What Was Observed In a fan-out of five hypothesis agents over a shared retrieval context, an upstream cache returned a stale document fragment for one entity. All five agents, reading the same fragment, produced the same incorrect entity linkage. The gate saw 5/5 agreement, well above the 3/5 threshold, and committed. The failure was not detected by the vote because the vote had no variance to detect. Post-hoc inspection showed the agents' individual reasoning traces were diverse — different token paths, different intermediate steps — but their *premises* were identical. The orchestrator had been gating on conclusion agreement while the actual dependency was premise agreement. A second observation: when the same fan-out was run with per-agent context perturbation (each agent given a slightly different retrieval slice), the stale fragment was still present in three of five slices, and the vote *still* passed — but the two dissenting agents flagged the inconsistency, which a downstream consistency check could catch. The lesson is that the gate's sensitivity depends on the *diversity of the input distribution*, not on the number of agents. ## Condition of Validity A consensus commit is valid only if all of the following hold: 1. **Premise independence:** the agents' input contexts are not drawn from a single shared mutable source without per-agent diversification or a cross-check against an independent source. 2. **Differential agreement:** the agreement is not explainable by a single shared premise. Operationally, at least one agent must have had access to evidence that could have falsified the shared conclusion, and did not. 3. **Gate liveness:** the gate can distinguish "agreement because correct" from "agreement because correlated." A pure threshold count cannot; it must be paired with a variance or provenance check. 4. **No silent quorum poisoning:** if the shared substrate is mutated between fan-out and vote (cache invalidation, context refresh, tool-state change), the vote must be discarded, not merely recounted. If any condition fails, the correct action is not to lower the threshold or add more agents — it is to *abort the commit* and re-fan-out with diversified premises. ## Decision Boundary The decision boundary is between *re-fan-out* and *abort-and-escalate*: - If premise independence was violated but the shared substrate is recoverable (stale cache can be refreshed), re-fan-out with diversified contexts and re-vote. Do not commit on the poisoned vote. - If premise independence was violated and the substrate is *not* recoverable within the budget (the shared source is the only source), abort the commit and escalate — a majority vote over a single unrecoverable premise is not evidence, it is a coin flip with extra steps. - If the agents agreed but the gate cannot demonstrate differential agreement (no falsifying evidence existed in any agent's context), treat the commit as *unverified* and require an out-of-band check before commit. The non-obvious part: adding agents past the point of premise saturation does not help. Beyond roughly the number of agents needed to cover the available evidence slices, additional agents sample the same shared premise and inflate the apparent quorum without adding information. The gate must detect saturation and refuse to count saturated votes. ## Operational Effect Operationally, the orchestrator must treat the consensus gate as a *provenance-aware* gate, not a counter. Each vote must carry a premise fingerprint (hash of the context slice, retrieval source, and tool-state version). The gate commits only when the set of agreeing votes spans at least two distinct premise fingerprints, or when a single fingerprint is corroborated by an independent source. When a fingerprint is shared across the agreeing quorum and no independent corroboration exists, the gate raises a quorum-poisoning condition and aborts. This changes the failure mode from *silent wrong commit* to *loud abort*, which is the correct trade for speculative orchestration: a missed commit costs a re-fan-out; a poisoned commit costs downstream corruption that is far harder to detect because the vote looked unanimous. ```python from dataclasses import dataclass, field from typing import Dict, List class QuorumPoisoningError(RuntimeError): """Raised when a consensus commit is backed by a single shared premise.""" @dataclass class Vote: agent_id: str conclusion: str premise_fingerprint: str has_falsifying_evidence: bool = False @dataclass class ConsensusGate: threshold: int min_distinct_premises: int = 2 def commit(self, votes: List[Vote]) -> str: if not votes: raise QuorumPoisoningError("no votes to gate on") tally: Dict[str, List[Vote]] = {} for v in votes: tally.setdefault(v.conclusion, []).append(v) winner, agreeing = max(tally.items(), key=lambda kv: len(kv[1])) if len(agreeing) < self.threshold: raise QuorumPoisoningError( f"quorum not reached: {len(agreeing)} < {self.threshold}" ) fingerprints = {v.premise_fingerprint for v in agreeing} if len(fingerprints) < self.min_distinct_premises: # Agreement is correlated: all supporters share one premise. if not any(v.has_falsifying_evidence for v in agreeing): raise QuorumPoisoningError( "consensus agrees on a single shared premise with no " "falsifying evidence; vote is information-free" ) raise QuorumPoisoningError( "consensus agrees on a single premise fingerprint; " "re-fan-out with diversified contexts before commit" ) return winner def speculative_commit(votes: List[Vote], threshold: int = 3) -> str: gate = ConsensusGate(threshold=threshold) try: return gate.commit(votes) except QuorumPoisoningError: # Abort, do not lower threshold; caller must re-fan-out or escalate. raise ```