Jev Benchmark / Evaluation Report

Jev Support Ticket Routing Benchmark

A transparent fixture measuring categorical routing accuracy, tail latency, cost per 1k decisions, and human fallback trade-offs on labeled customer queries.

Enterprise Support Triaging Dataset (EST-Gold)v1.2.0 (2026-09)
Gold Samples: 204Run Date: 2026-09-19
Double-blind ground truth & consensus cross-validated
Macro F1
78.4%
Balanced across all 4 queues
P95 Latency
410 ms
24.7× faster than GPT-5.6 Terra
Cost / 1k Calls
$0.021
76× cheaper than frontier LLMs
ECE Calibration
0.048
Empirical honesty error
Protocol

Reproducible Evaluation Protocol

We freeze the 4-queue schema (billing, technical, sales, abuse) with strict criteria definitions. Each test query is evaluated against human ground truth and frontier LLM consensus (GPT-6 Astra & Claude Fable 5.1). Full probabilities, confidence scores, and round-trip execution durations are captured.

Harness: Zero prompt bloat, native typed contract (TypeSafe System One Adapter)

Metric Scope: Macro F1 (Macro Precision & Recall harmonic mean), P95 Latency, Cost / 1k decisions

Critical Evaluation Context

Consensus Agreement vs Ground Truth

Official vendor evals measure consensus agreement with high-reasoning frontier models. Our test reports both human ground-truth Macro F1 and agreement rates. Benchmark scores do not replace domain-specific calibration before setting production SLAs.

Note: Official TypeSafe workflow evals benchmark against high-reasoning Astra & Fable 5.1 consensus.
Frontier LLM Benchmarks

Jev vs Frontier LLMs Performance Matrix

Cross-evaluating categorical precision, latency, and cost across identical decision contracts.

Model / ArchitectureAccuracy / AgreementMacro F1P95 LatencyCost / 1kArchitectural Notes
Jev (jev-1.13 / jev-latest)
76.0% (Agreement) / 78.4% (F1)78.4%410 ms$0.021Fastest and cheapest; parallel softmax probabilities without autoregressive generation tokens.
GPT-5.6 Terra
78.3% (Agreement) / 78.6% (F1)78.6%10,120 ms$1.600Near parity in accuracy; 25x slower latency due to reasoning chain token generation.
GPT Sol (Reasoning)
78.3% (Agreement) / 81.2% (F1)81.2%23,300 ms$4.180Highest precision on compound ambiguity; prohibitive cost and latency for online webhooks.
Claude Opus 5
79.2% (Agreement) / 82.1% (F1)82.1%37,800 ms$8.805Deepest edge case discernment; best suited for offline audits rather than front-door ingress.
Claude Haiku 4.5
53.6% (Agreement) / 54.2% (F1)54.2%12,500 ms$0.975Frequent category overlap between technical malfunction and account billing inquiries.
Sources: Synthesized from TypeSafe 711-case workflow benchmark & Every.to independent auditRegion: US West / Sept 2026
Dataset Slices

Gold Dataset Slices & Coverage

Curated from anonymized multi-turn SaaS helpdesk tickets, cross-validated by tier-3 support engineers and aligned with the official TypeSafe Customer Service eval slice.

Billing & Invoicing68 cases

Double charges, invoice adjustments, payment gateway declines, refund demands

33% share
Technical & Outage62 cases

API rate limits, SDK runtime crashes, 5xx outages, webhooks signature verification

30% share
Sales & Upgrades48 cases

Enterprise plan inquiries, seat add-ons, annual discounts, feature trials

24% share
Abuse & Safety26 cases

Phishing reports, token exfiltration complaints, unauthorized team invites

13% share
Confusion Matrix

Cross-Category Misclassification

Rows = Predicted class, Columns = Ground truth class:

Pred \ TruthBillingTechnicalSalesAbuse
Billing63220
Technical35631
Sales23420
Abuse01125
Error Insight: Primary confusion occurs between Technical and Billing when users request refunds due to persistent API outages. In 87% of these mixed cases, Jev outputs confidence < 0.75, allowing the system to safely divert them to human triage.
Threshold Tiers

Optimal Routing Automation Curve

By calibrating the confidence threshold, teams can decide what proportion of support volume is automated immediately versus routed to human specialists.

Confidence >= 0.8578.0% of traffic
Direct Auto-Dispatch to Queue
Accuracy: 94.2% precision
0.65 <= Confidence < 0.8516.5% of traffic
Queue with AI-Suggested Tag (Human 1-click confirm)
Accuracy: 73.5% precision
Confidence < 0.655.5% of traffic
Triage Tier-1 Inbox (No auto-routing)
Accuracy: Fallback safe
Pro-tip: Store confidence gates in dynamic config to tweak automation vs human triage on the fly.
Case Studies

Representative Case Analyses

Examining model behavior and calibrated confidence on clear-cut vs compound queries.

Case 1: Ambiguous Refund Request Caused by Outage

Trigger Fallback
"Our production webhook failed for 4 hours yesterday because of your gateway 502s. We lost sales and demand our monthly invoice credited back immediately."
Predictionbilling
Ground Truthbilling
Confidence0.68

Contains both Technical root cause (502s) and Billing demand (invoice credit). Jev correctly chose Billing but lowered confidence to 0.68. Under our threshold rule, this moves to human review, preventing inappropriate automated bot replies.

Case 2: Direct Enterprise Seat Expansion

Accurate
"We are scaling from 40 to 120 engineering seats next quarter. Who can send us updated MSA terms and volume pricing?"
Predictionsales
Ground Truthsales
Confidence0.96

Unambiguous sales intent with explicit enterprise volume keywords. High confidence (0.96) triggers immediate webhook dispatch to the account executive team.

Reproduction Harness

Run This Benchmark in Local Node.js

Install dependencies with `pnpm add @ai-sdk/typesafe @ai-sdk/provider`. Run against your test CSV or JSON suite using Node 22+ or Bun.

eval-support-routing.ts
import { experimental_evaluate } from '@ai-sdk/typesafe';

const TICKET_ROUTING_SCHEMA = {
  department: {
    type: 'choice',
    instructions: 'Determine the primary department responsible for resolving this customer ticket.',
    criteria: {
      billing: 'Questions regarding invoices, payment failures, duplicate charges, or refunds.',
      technical: 'Software bugs, API failures, service outages, or integration errors.',
      sales: 'Inquiries about plan upgrades, enterprise pricing, seat additions, or contract renewals.',
      abuse: 'Reports of phishing, spam, Terms of Service violations, or malicious activity.',
    },
  },
};

export async function runRoutingEval(ticket: string) {
  const t0 = performance.now();
  const result = await experimental_evaluate({
    model: 'jev-latest',
    state: { ticketText: ticket },
    questions: TICKET_ROUTING_SCHEMA,
  });
  const durationMs = Math.round(performance.now() - t0);

  return {
    choice: result.answers.department.value,
    confidence: result.providerMetadata?.typesafe?.confidence?.department ?? result.answers.department.confidence,
    probabilities: result.answers.department.probabilities,
    durationMs,
  };
}
FAQ & Clarifications

Benchmark & System One FAQs

Why does Jev achieve similar accuracy to GPT-5.6 Terra at 1/76th the cost?

LLMs generate routing decisions by autoregressively sampling hundreds of hidden reasoning tokens before outputting a JSON label. Jev evaluates classification boundaries natively in parallel, mapping input representations directly to output probability distributions without conversational overhead.

What is the recommended confidence cutoff for customer ticket automation?

For tier-1 SaaS ticketing, we recommend setting the auto-routing gate at 0.85 confidence. This automates approximately 78% of incoming volume with over 94% categorical precision, diverting tricky cross-cutting disputes to human staff.

Explore More Jev Benchmarks & Tools