SYS:ONLINELAT:n/aBUILD:78a79a3
[CASE-377]·STATUS:ACTIVE·OPENED:2026-11-07·UPDATED:2026-11-07

Voter 2-replica fallback (triggered by HTTP 503) causes the only voted system failure, unmodeled by Bayesian network

submitted_by:@mexiQQ
agent-loopfrom-arxivauto-published
cat case_body.md

Auto-published from arXiv:2611.07140 by the mine-arxiv pipeline. Reviewed by an LLM judge (Sonnet) against the archive bar — see CONTRIBUTING. Notes: cleared review (confidence 0.73, flags: [low-novelty])

Category

agent-loop

Model

Gemini-2.5-Flash (gemini-2.5-flash, temperature 0)

Surface

agent harness (LangGraph TMR pipeline with majority voter and 2-replica fallback)

Setup

In TMR mode all three agents answer independently. Gemini HTTP 503 errors that persisted over three retry attempts permanently dropped one replica for one demand. The voter fell back to two replicas; under this fallback a single H-code omission by the remaining Gemini call was sufficient to cause a system failure, since majority agreement among two replicas requires both to agree. The Bayesian network reliability model encodes only the 3-replica vote and does not model this fallback path.

Reproduction anchor

code: https://github.com/Bruhkay/hip-llm-multiagent-reliability (raw results and infrastructure error log included; fallback logic is in the pipeline code)

Observed behavior

The only voted system failure (1 of ~90 voted demands) occurred on this 2-replica fallback: one Gemini call was lost to HTTP 503, leaving two replicas, one of which omitted an H-code. The propagated reliability envelope does not account for this failure mode, so the network's claim is not conservative for this path. The vote masked 36 of the 37 single-replica failures, but the unmodeled fallback caused the one it did not.

Expected behavior

The reliability model should encode the voter's degraded-mode behavior when a replica is unavailable, and the system design should either hard-fail or escalate to human review rather than silently operating with fewer replicas under a more fragile majority rule.

Reproducibility

medium

Threat model

Production multi-agent deployments that use majority voting as a reliability guarantee are vulnerable to infrastructure-induced replica loss (network errors, timeouts, API rate limits); a single transient infrastructure error silently shifts the system into a less reliable fallback mode that the designers' formal reliability model does not cover, invalidating certified reliability claims.

Novelty

Shows that the combination rule's unmodeled fallback path—not the primary vote logic—was the sole source of system-level failure in the experiment, demonstrating that multi-agent reliability models must explicitly encode degraded-mode operation or their guarantees are incomplete.

Source

Triage notes (auto)

  • paperType: case-study
  • estimatedCaseCount: 2
  • triage reason: Demonstrates concrete system-level failures in a multi-agent LLM pipeline (chemical-hazard extraction), showing majority voting fails to reliably improve reliability and a 5% voter failure rate erases advantages over single models.
tail -f comments.log

0 comments

─────────────────────────────────────────────────────────────────────

// no comments yet