GPT-5.5-class answers in about 4 seconds.
Don't predict which questions are hard — observe agreement between fast agents, and only pay for intelligence when they disagree.
Agreement is the router.
No difficulty classifier, no learned router for this preset. Four cheap voters answer in parallel; if they agree, that's the answer. If they don't, and only then, a frontier model is paid for.
Tier 0 runs four voters in parallel on Cerebras wafer inference: two gpt-oss-120b
single-pass calls plus two gemma-4-31b agents running an RLM loop
(reason → run_python → observe, capped at 6 rounds, sandboxed execution). Extracted answers are
normalized and compared. Tier 1 only fires on disagreement — GPT-5.5 re-solves the
problem from scratch, never seeing what the voters produced.
-
Consensus rule
If ≥3 of 4 extracted answers agree after normalization, return immediately — median 3.0s, covering 83% of AIME traffic at 98% precision. No difficulty prediction, no classifier: agreement itself is the router.
-
Why arbitration is "clean"
An early version showed the arbiter the voters' reasoning on disagreement. It anchored on the majority's (wrong) answer 7 times out of 9, against a 98% solo baseline for the same model with no anchor. The fix: on disagreement, GPT-5.5 gets the original problem only — the voters' outputs are discarded, not summarized, not shown as "context."
Where the seconds and the accuracy actually go.
Every number on this page comes from timed runs, not estimates. Sample sizes are small (see honest limits) — treat single-point comparisons as directional.
Accuracy vs. median latency, AIME 2025, latency on a log axis. The shipped config sits closest to the top-left corner of any config that isn't a raw frontier or Fable solo call.
Fraction of queries that reach the frontier model. cascade-rlm-conv spends more frontier calls than the plainer wafer-only cascades in exchange for closing most of the remaining accuracy gap — still a 6× reduction from calling GPT-5.5 on every query.
gemma-4-31b gains 12–16 points from running as an RLM agent (reason → run_python → observe) instead of answering solo. gpt-oss-120b shows no such benefit — the working rule so far: externalized reasoning helps models that don't already reason internally, and doesn't help ones that do.
The exact preset, verbatim.
cascade-rlm-conv is a preset shipped with hermes moa serve —
nothing here is bespoke glue code.
# preset: cascade-rlm-conv mode: cascade cascade: min_consensus: 3 clean_arbiter: true reference_models: - provider: custom:cerebras model: gpt-oss-120b - provider: custom:cerebras model: gpt-oss-120b - provider: custom:cerebras model: gemma-4-31b agent: rlm - provider: custom:cerebras model: gemma-4-31b agent: rlm aggregator: provider: openai-codex model: gpt-5.5 reference_max_tokens: 8000 reference_quorum_grace: 1.0
# start the server
hermes moa serve --port 8655
# any OpenAI-compatible client works — this one's Python from openai import OpenAI client = OpenAI(base_url="http://localhost:8655/v1", api_key="not-needed") resp = client.chat.completions.create( model="moa:cascade-rlm-conv", messages=[{"role": "user", "content": "..."}], )
The command above is the general-purpose invocation — pick any port for your own deployment. The pinned production instance for this preset runs as the cascade endpoint (localhost :8655 on the host). Any OpenAI-compatible client or agent harness can point at it, including client-side tool calls — they're forwarded straight through to the underlying models.
Where this doesn't apply.
⚠ read before you route production traffic here
- Competition-math-class evaluation only. AIME, HMMT, GPQA-numeric — exact-answer domains where a consensus vote can be checked by string/value normalization. This is not evidence for open-ended or prose tasks.
- ~92–95% is the wafer-only ceiling. The remaining errors are systematic, not random: all four Tier 0 voters can agree on the same wrong answer, and no consensus rule catches a unanimous mistake. That failure mode is exactly why the frontier arbiter tier exists.
- Prose and code lanes route elsewhere. Advisory context from other agents measurably hurts code-editing accuracy in our tests — the cascade pattern here is scoped to exact-answer math, not applied uniformly.
- Small samples. AIME 2025 is n=60; the GPQA-numeric RLM-lift comparison is n=25; several other evals cited here are n=20. Treat single-point deltas as directional, not final.