When several AI checkers vote on an answer, what hidden things they share could make them all wrong together?
Which shared channels cause the strongest correlated validator failures?
This explores which things validators in an AI quorum have in common (prompts, data sources, infrastructure, model lineage) are most likely to make them wrong together, and the corpus can name the suspects but hasn't ranked them.
This explores which things validators in an AI quorum have in common are most likely to make them wrong together, and the corpus can name the suspects but hasn't ranked them. The Honest Quorum paper lists eight inputs that agentic validators can share, from model weights all the way to provider infrastructure. A single fault in any shared input can hit several validators at once. A quorum vote then counts agreement as independent confirmation when it may be one error echoed many times Can a quorum of validators really provide independent judgment?.
The corpus is clearest on which channels the obvious fix doesn't reach. Switching to different model families addresses only two of the eight channels. Correlated errors likely persist through shared prompts, shared retrieval sources, and shared provider infrastructure Does model diversity actually reduce validator agreement failures?. Suppose you hire validators from three different labs, give them the same prompt and the same retrieved documents, and run them on the same cloud. You've diversified the channel that was easiest to change and left the others untouched. The paper's proposed test is to compare error correlation across validators that differ on one channel at a time. That experiment would settle the ranking, and the corpus doesn't report having run it, so nothing here can call one channel the strongest.
The stakes come from how the guarantees split. Agreement holds from protocol assumptions alone, but semantic correctness holds only statistically, and those statistical bounds depend on how validators behave, which the protocol can't enforce Can validator consensus guarantee both agreement and semantic correctness?. A quorum can authenticate correctly, sign the right messages, and reach consensus while endorsing an invalid transition. Agreement shows the protocol ran, not that the answer is right Can a quorum of honest validators certify an invalid transition?. My reading is that shared channels are what quietly break the independence those statistical bounds assume.
A neighboring finding suggests why the correlated failures would be hard to spot. Frontier models silently corrupt about 25 percent of document content over long workflows, and the corruption never plateaus and goes unnoticed in spot checks Do frontier LLMs silently corrupt documents in long workflows?. Weaker models tend to fail visibly by deleting content, while stronger ones fail by subtle corruption that keeps the surface intact Does model capability change how documents degrade?. That work isn't about validators. But if strong validators share an input, their shared mistakes could look clean, so nothing would prompt anyone to question a unanimous vote. The open question is whether the channels that are cheapest to share, prompts and retrieval sources, turn out to be the most damaging.
Sources 6 notes
The Honest Quorum paper identifies eight things agentic validators may share—from model weights to provider infrastructure—allowing one fault to affect multiple validators simultaneously. Quorum voting then counts agreement as independent confirmation when it may simply reflect a common cause.
While different model families address two of eight shared fault channels, correlated epistemic errors likely persist through prompts, retrieval sources, and provider infrastructure. Measuring error correlation across validators differing on one channel at a time could quantify the effect.
Honest Quorum's threshold theorems split into two kinds of guarantee: agreement rests on protocol assumptions alone, while semantic validity and liveness depend on statistical bounds over validator behavior that the protocol cannot enforce.
A quorum of validators can authenticate correctly, sign expected messages, and achieve consensus while endorsing a semantically invalid transition. Agreement proves the protocol ran, not that the answer is right.
Even the strongest models (Gemini 3.1 Pro, Claude 4.6 Opus, GPT 5.4) degrade documents by ~25% over long relay workflows across 52 domains. Degradation decelerates but never plateaus, and errors compound silently, remaining undetected in spot-checked outputs.
Show all 6 sources
DELEGATE-52 shows weaker LLMs degrade documents through visible deletion, while frontier models degrade through subtle corruption that preserves surface integrity. This shift makes frontier failures harder to detect and potentially more dangerous at workflow scale.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- The Honest Quorum Problem: Epistemic Byzantine Fault Tolerance for Agentic Infrastructure
- Can AI Agents Agree?
- Psychologically Enhanced AI Agents
- LLMs Corrupt Your Documents When You Delegate
- The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs
- LLMs Get Lost In Multi-Turn Conversation
- RRSI: Regularized Recursive Self-Improvement of Agent Harnesses
- When AIs Judge AIs: The Rise of Agent-as-a-Judge Evaluation for LLMs