llm-d-sc-cost-v2

cost classifier for llm-d semantic classification. Labels: MINIMAL, LOW, MODERATE, HIGH.

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)	371	0.8976	0.863 – 0.924	0.8059
real traffic (WildChat, unanimous 3-model jury)	427	0.8735	0.839 – 0.902	0.8097
legacy hand-authored held-out	60	0.9667	0.886 – 0.991	0.9665

Hand-authored minus real traffic: +0.093. Hand-authored held-out prompts are written in the same clean register as the anchors; real users send truncated pastes, fragments and roleplay preambles. The real-traffic row is the one that predicts production behaviour.

Run-to-run variance. Three seeds of this recipe give 0.8868 / 0.9003 / 0.8976 on refined gold (median 0.8976, spread 1.35 pts). This is the MEDIAN run. Trained on 43,717 jury-agreed plus 15,805 contested real-traffic rows — 6.5x the previous corpus, which is worth +2.16 points over the smaller-corpus median and is the only change in this project that exceeded cost's measurement noise.

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

85942 rows from cost-v2:5970+cost-real:43717+cost-real-contested:15805+cost-distill:30000, 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: {'MINIMAL': 15361, 'LOW': 34295, 'MODERATE': 33322, 'HIGH': 2964}. Held-out eval prompts are excluded by content hash.

Latency

CPU single-request: p50 2.93 ms, p99 3.24 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.