llm-d-sc-triage-gate

triage classifier for llm-d semantic classification. Labels: TRIVIAL, WORK.

Architecture: sequence-classification head (requires a runtime that reads logits), base sentence-transformers/all-MiniLM-L6-v2.

Accuracy

Read the real-traffic row first.

eval set	n	accuracy	95% CI	macro F1
real traffic, refined gold (high-effort re-adjudication)	376	0.9601	0.935 – 0.976	0.9391
real traffic (WildChat, unanimous 3-model jury)	418	0.9593	0.936 – 0.974	0.9375

Run-to-run variance. RECOMMENDED COMPLEXITY MODEL for llm-d-sc routing. Supersedes llm-d-sc-complexity-v2 (4-tier, 0.8963) and llm-d-sc-route-gate (2-tier SIMPLE+MEDIUM split, 0.9269) for backend selection.

Published seed is the FIRST one run (11), not the best. Both seeds score 0.9601 on refined gold -- not a duplicate: different parameter sums, logits differing by up to 1.48, 10 of 376 predictions disagree and happen to cancel.

accuracy 96.01% majority baseline 80.32% lift +15.69 p50 4.37 ms TRIVIAL recall 94.59%, precision 86.42% (TRIVIAL is 19.7% of the eval)

WHAT IT DECIDES. TRIVIAL (the 4-tier ladder's SIMPLE) versus WORK (MEDIUM, COMPLEX, REASONING): can a small model or a cache serve this, or does it need the main model?

WHY THIS FOLD RATHER THAN THE ONE PRAXIS CURRENTLY USES. Enumerating EVERY contiguous fold of the 4-tier ladder and ranking by three-juror agreement: SIMPLE | MEDIUM+COMPLEX+REASONING 86.9% agreement <- this model SIMPLE+MEDIUM | COMPLEX+REASONING 82.0% agreement <- the deployed split Their majority baselines are 80.32% and 80.59%, so the comparison is like-for-like: +3.32 points for choosing the split by measurement instead of by which one the router happens to implement. Tier-exact accuracy on the full 4-tier taxonomy is 0.8963 and is capped near 0.926 by inter-juror agreement; this decision is not.

TWO LIMITATIONS THAT MATTER MORE THAN THE HEADLINE.

1. CONTESTED ROWS. On the 176 rows where the three-model jury SPLIT, this model scores 80.11% against an 82.95% majority baseline -- BELOW chance -- with TRIVIAL recall at 43.33%. Contested rows are about 32% of real traffic.

2. ABSTENTION DOES NOT HELP IT. For the other gates in this family, model confidence tracks jury disagreement at 1.7-1.9x enrichment, so routing the least-confident slice to the large model fixes most of the contested-row problem. This model sits at 1.0x: its least-confident 5% contains contested rows at exactly the base rate. It is confidently WRONG on hard rows rather than uncertain about them, so no confidence threshold rescues it -- and at 99% recall it false-fires on 74% of traffic.

PRACTICAL GUIDANCE. Deploy at full coverage and accept 80.11% on the split rows, or send TRIVIAL predictions to the cheap path only when the cheap path degrades gracefully -- TRIVIAL precision is 86.42%, so about 5% of WORK prompts land there. A three-way TRIVIAL/STANDARD/HARD variant is being tested to see whether a middle tier fixes the blind-confidence problem, as it did for the egress gate; if it does, this card will point at it.

The eval has a measured ceiling

Gold labels were audited by blind paired adjudication in two strata — the rows this model got wrong, and a sample of the rows it got right — with the judge shown two candidate labels in random order and no indication of provenance. Roughly 4.9% of the gold labels are themselves wrong, so a PERFECT classifier scored against this eval would reach about 0.95, not 1.0.

Read the real-traffic accuracy against that ceiling, not against 100%. Auditing only a model's mistakes would move the number up artificially; sampling the correct rows too is what makes the estimate honest, and it revealed that on ~3.3% of "correct" rows the model agreed with a bad label — meaning measured accuracy is very slightly overstated.

How the eval was built

Real-traffic rows come from WildChat-1M (ungated real assistant traffic). Each prompt was labelled independently by three models (claude-opus-5, claude-sonnet-5, claude-fable-5-1) from the task rubric alone -- no labeller ever saw a proposed label, so agreement is evidence rather than assent. Only unanimous rows are scored.

Those three agree unanimously on roughly 70-74% of real prompts. The remaining prompts are published as a contested split rather than discarded: they measure how much real traffic this taxonomy does not resolve, which no single accuracy figure can express.

Training data

418 rows from triage-v2+triage-real+triage-active+triage-distill+triage-real-contested, mixing jury-labelled real traffic (register and class prior) with rubric-grounded synthetic data (coverage of tiers that are rare in real traffic). Training prior: None. Held-out eval prompts are excluded by content hash.

Latency

CPU single-request: p50 4.37 ms, p99 9.78 ms (Apple M-series, single thread). llm-d-sc serves the classifier on CPU, so model size trades directly against per-replica throughput.

Limitations

• WildChat is consumer traffic. For sensitivity it is ~93% PUBLIC and cannot measure the tiers that gate egress; the enterprise row above covers those.
• Labels come from LLM jurors, not human annotators. The rubric was validated by reproducing the project's hand-authored gold labels (complexity 0.9875, cost 1.000, sensitivity 1.000) before use.
• Not independently reproduced.