VI · Multi-AgentEmerging★

Epistemic Fault-Domain Quorum

also known as Epistemic Sybil Resistance, Independence-Counted Quorum, Evidence-Root Quorum

Credit a multi-agent quorum only with the resilience its distinct upstream roots justify, admitting the coalition when the number of independent root faults needed to compromise it clears a declared threshold.

Context

A team routes a high-stakes decision through several agents and treats their agreement as evidence that the decision is sound. The members may run on different models, different prompts and different vendors, and the count of members that agreed is what gets written into the approval record. Behind those members sits shared upstream material: the same retrieved documents, the same telemetry feed, the same tool backend, the same base model, sometimes the same pre-training corpus. The dependency graph from members back to those origins is rarely written down anywhere.

Problem

Members that descend from a common upstream root do not supply independent observations; they relay one observation several times, and the tally cannot tell the difference. Holding a single evidence root fixed while report multiplicity rises from 1 to 32 collapses naive posterior coverage from 0.940 to 0.263, so confidence grows precisely where information does not. The structural version is worse: an arbitrarily large quorum can still have a minimum root-fault cut of one, meaning a single stale feed or a single corrupted document flips a unanimous verdict. Adding voters at a fixed threshold cannot raise that cut, so the usual remedy for a shaky decision buys nothing.

Forces

  • Adding another member is cheap and visibly raises the vote count, while adding a genuinely distinct evidence root is expensive and changes no number anyone currently reports.
  • Vendor and architecture labels are the available proxy for independence, yet measured across 38 models it is inferred generative-process diversity rather than vendor difference that tracks correlated failure, at a cross-benchmark partial rank association of -0.216 with a 95% interval of [-0.309, -0.122], negative on all ten benchmarks.
  • Shared roots frequently sit below the model layer entirely — one telemetry feed, one retrieval corpus, one tool backend — so decorrelating by model family leaves the cut exactly where it was.
  • Letting members read one another restores agreement fast: proposals from different model families converge within a single round, spending on coordination the diversity the coalition was assembled for.
  • Recognising a shared ancestor can only lower the credited resilience, never raise it, so honest accounting is locally unattractive to whoever is seeking the approval.

Example

A bank asks five review agents to sign off on a change to its fraud rules, and all five approve, which the change log records as unanimous. Four of them read the same nightly risk report, and that report was built from a feed that had gone stale a day earlier. When the change misfires, the postmortem finds the five approvals rested on two sources of information rather than five, and that the stale feed on its own was enough to flip the verdict.

Diagram

Solution

Therefore:

Treat independence as something measured at admission rather than assumed from headcount. For each member, enumerate the roots it depends on — the source documents it read, the retrieval corpus behind them, the telemetry feed, the tool backends it called, the base model and the generative process that produced it — and record the result as an exposure map built from what ran, not from declared configuration. The resilience the coalition may be credited with is then the smallest set of root faults that would compromise the decision, not the number of members that agreed. Compare that cut against a threshold declared in advance for the class of decision. A coalition below the threshold is repaired by buying a distinct root — a second corpus, a second tool backend, a separate model lineage — or by downgrading the claim to what the evidence supports; enlarging the membership at a fixed threshold is provably not a repair. Keep members from reading each other's complete outputs before they commit, because exchange collapses the very diversity being counted. Store the map and the cut with the decision so an auditor can recompute both.

What this pattern forbids. A quorum cannot be credited with more resilience than its smallest root-fault cut, a coalition whose cut is below the declared threshold must not be admitted, and adding members at a fixed threshold never raises the cut.

And the patterns that stand alongside it, or against it —

  • complementsVoting-Based Cooperation★— Finalise a decision across multiple agents by collecting and tallying their votes on candidate options, so the joint output reflects collective rather than single-agent judgement.
  • complementsHeterogeneous-Model Council with Synthesis Judge★— Three or more role-specialized personas run on different model architectures in parallel; a synthesis judge — given only their structured JSON, not the original input — produces the final verdict.
  • complementsQuorum on Mutation·— Require multiple consecutive ticks (or runs) to agree before a mutation to durable state lands.
  • complementsSelf-Consistency★★— Sample the same question multiple times at non-zero temperature and aggregate by majority or judge to mitigate hallucination.
  • complementsConsensus-Averaging Over Expertise✕— Anti-pattern: a self-organising LLM team pursues integrative compromise, averaging expert and non-expert views instead of weighting the known expert, so team output falls below the best member and degrades further as the team grows.
  • complementsHidden Distributed Monolith (Multi-Agent)✕— Anti-pattern: a multi-agent system is presented as decoupled, independently deployable agents, but at runtime they share context, run in synchronous chains, and have no failure isolation, so it behaves as a tightly-coupled distributed monolith.
  • complementsClone Fan-Out Research·— Spawn 100 or more identical, full-capability agent instances in parallel — each a complete general agent rather than a role-specialised worker — and aggregate their independent outputs into a single answer.
  • complementsModality-Conflict Arbitration★— Have each observation channel emit its own recorded claim, treat disagreement between channels as a detected event, and resolve it against a pre-declared per-claim-type modality authority rather than an implicit fused guess.

Neighbourhood

Click any neighbour to follow the language. Scroll to zoom, drag to pan.

References

Provenance