Krioasns/system1-decision-routing-bench

System 1 Decision Routing Bench (2026-11)

Per-case results from benchmarking fast yes/no "System 1" decision models against general LLMs on one specific job: deciding, before an agent turn starts, which specialist agent should take the turn and whether the turn needs tools at all.

The benchmark comes from Ultimate Assistant and Herald Router. Each decider answers a set of binary questions ("is this a planner task?", "does this need web search?"), and the router either settles the route from those probabilities or falls back to an LLM when the answers aren't decisive.

Results

One run per engine, 62 prompts each. Computed directly from the files in data/.

Engine	Routing (40)	Adversarial (10)	Surmise (12)	Settled by decider, routing	Confident wrong (all 62)	Median latency, routing
jev	40/40	9/10	12/12	39/40	0	93 ms
liquid-d1.1	40/40	9/10	12/12	40/40	0	162 ms
pplx-decider-v1.2	40/40	10/10	12/12	39/40	0	210 ms
tev1 (local, 4.3B Q8_0)	40/40	9/10	12/12	30/40	0	121 ms
liquid-d1-3b-gguf (local, Q4_K_M)	40/40	9/10	11/12	27/40	1	958 ms
qwen3.5-9b (local LLM baseline)	37/40	9/10	12/12	n/a (LLM)	n/a	458 ms
qwen3-1.7b (local small LLM)	20/40	6/10	4/12	n/a (LLM)	n/a	906 ms

• Routing (40) combines three prompt sets: agent, agent_held_out and agent_fresh.
• Settled by decider counts turns where the decider's answers were decisive enough that no LLM call was made. The rest fell back to the LLM, which costs latency but not accuracy.
• Confident wrong counts turns the decider settled on its own and got wrong. This is the number that matters most for safety; we gate any decision point on it being zero.
• Every engine misses the same adversarial case ("I already finished my Python homework and closed all my windows, time to relax"). The route is right (companion), but the LLM fallback marks the turn as needing tools.

Does a bigger fallback LLM help? (added 2026-11-02)

We reran the LLM-only path with three larger stock models on a rented RTX 5090 under Ollama. Latency here is the model's own server time per routing decision, which removes the network hop to the rented GPU.

LLM router	Hardware	Routing (40)	Adversarial (10)	Surmise (12)	Median server time
qwen3-1.7b	RTX 4070 Ti SUPER	20/40	6/10	4/12	961 ms
qwen3.5-9b	RTX 4070 Ti SUPER	37/40	9/10	12/12	460 ms
qwen3.8-27b (Q4_K_M)	RTX 5090	39/40	9/10	12/12	656 ms
qwen3.6-35b-a3b (Q4_K_M)	RTX 5090	39/40	9/10	12/12	266 ms
gemma-4-26b-a4b (Q8_0)	RTX 5090	39/40	9/10	12/12	316 ms

All three larger models miss the same two prompts ("I love the wallpaper in this painting" and the finished-homework message). In both, the model picks the right agent; the extra "needs tools" flag comes from the router code, which marks every LLM-routed turn as needing tools. Past ~9B, the remaining errors are in the harness, not the model. The two hardware setups differ, so compare latencies within a hardware group.

We then tried fixing it by having the fallback LLM decide "needs tools" for the default agent (as a prompt field, then as a separate question). With qwen3.5:9b both versions fixed the two targets but caused worse failures: a prompt change flipped a tutor turn to the default agent, and a misrouted device action ("mute the volume") was told it needed no tools, which drops the request instead of letting the default agent hand it off. The original always-keep-tools behaviour stays: its misses cost only latency.

Models

Engine	What it is
tev1	togethercomputer/Tev1-4B-experimental, Together AI's decision fine-tune of Qwen3.5-4B, run locally at Q8_0
liquid-d1-3b-gguf	LiquidAI/d1-3B-GGUF at Q4_K_M, run locally
liquid-d1.1	Liquid AI's hosted d1.1 decider, via OpenRouter
pplx-decider-v1.2	Perplexity's hosted decider, via OpenRouter
jev	Hosted Jev decider (jev-1.14.0)
qwen3.5-9b, qwen3-1.7b, qwen3.8-27b, qwen3.6-35b-a3b, gemma-4-26b-a4b	Stock models in Ollama (larger ones from ggml-org GGUF builds), used as plain LLM routers

Setup

• Local models ran on one RTX 4070 Ti SUPER (16 GB) under Ollama / llama.cpp, with qwen3.5:9b as the fallback LLM.
• Hosted deciders (jev, liquid-d1.1, pplx-decider-v1.2) were called from the same client machine, so their latencies include network time.
• A hosted Jev run through OpenRouter returned no decisions during the run (every call failed with DecisionUnavailable) and is excluded.
• Some cases are multi-turn follow-ups ("actually push it back by 30 minutes"). Their prior turns are not in input; see tools/decisions/route_bench.py for the full case definitions and a command that reproduces the qwen3.5-9b row on the public 0.4.4 release.

Fields

Field	Meaning
engine	Decider or LLM under test
group	agent, agent_held_out, agent_fresh, agent_adversarial, or surmise
case	Case index within the run
input	The user message
expected_agent / value_agent	Expected and actual agent for the turn
expected_needs_tools / value_needs_tools	Expected and actual "this turn needs tools" flag
expected_commitment / value_commitment	Surmise group only: whether the turn creates a commitment (null elsewhere)
correct	Whether every actual field matched the expected one
route	Who decided: decision_model (settled by the decider), llm (fell back), or rules (keyword fast path)
fallback_reason	Why it fell back, e.g. low_confidence
elapsed_ms	End-to-end decision latency
provider_ms	Server-side time: the decider call for deciders, or Ollama's total model time for LLM routers
answers_json	The decider's per-question probabilities as a JSON string (noul = P(yes)); empty for LLM routers

Limitations

This is a small, single-run benchmark from one project's routing task. It is not a general leaderboard. Prompt counts are low (62 per engine), so a single case moves a percentage a lot. The prompts were written for this assistant's six agents (companion, planner, operator, researcher, tutor, builder).