Jev Benchmark / Evaluation Report

Jev Spam Detection & Moderation Benchmark

A task-level evaluation isolating false positives, stealth marketing evasion, inference latency, and quarantine queue trade-offs.

Conversational & Forum Spam Slice (CFS-300)v1.0.4
Gold Samples: 300Run Date: 2026-09-18
Double-blind ground truth & consensus cross-validated
Macro F1
88.5%
Harmonized spam & clean F1
False Positive Rate
1.1%
Only 2 / 180 benign blocked
P95 Latency
320 ms
Ultra-fast inline chat gate
Cost / 1k Msgs
$0.016
Viable for free-tier users
Protocol

Expensive Mistake Isolation Methodology

Spam detection has asymmetric error costs: falsely blocking a legitimate paying user (False Positive) is far more damaging than letting a promotional flyer slip through (False Negative). Our benchmark explicitly segregates legitimate banter, gray-hat affiliate spam, and malicious phishing payloads.

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

This benchmark serves as a calibrated architecture baseline. Real-world spam exhibits distribution shifts weekly; production implementations must maintain active threshold tuning.

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 (Noul / Choice Primitive)
92.3% Overall / 88.5% F188.5%320 ms$0.016Inline synchronous gate suitable for WebSocket chat messages and form submissions.
GPT-5.6 Terra
93.0% Overall / 89.2% F189.2%8,400 ms$1.420Slightly higher nuance on contextual sarcasm; too slow for real-time messaging gates.
Claude Sonnet 5
94.3% Overall / 91.0% F191.0%18,200 ms$4.250Best detection of stealth cloaked affiliate links; best deployed in asynchronous post-publish sweep.
Sources: Synthesized from TypeSafe 711-case workflow benchmark & Every.to independent auditRegion: US West / Sept 2026
Dataset Slices

Gold Dataset Slices & Coverage

Constructed from synthesized user community forums, direct chat logs, and public anti-spam corpora, carefully balanced with tricky benign messages (job announcements, discount sharing).

Legitimate User Chats180 cases

Normal conversational questions, greetings, account troubleshooting, product feedback

60% share
Gray-Hat Affiliates & SEO Drops60 cases

Sneaky link drops, keyword-stuffed discount codes, off-platform redirect hints

20% share
Malicious Scams & Phishing60 cases

Crypto wallet drainers, credential harvesting, fake customer support phone numbers

20% share
Confusion Matrix

Cross-Category Misclassification

Rows = Predicted class, Columns = Ground truth class:

Pred \ TruthCleanGray AffiliateMalicious
Clean178111
Gray Affiliate2453
Malicious0456
Error Insight: Malicious phishing was intercepted with 93.3% recall (56/60). The primary borderline slippage occurred in gray affiliate promos that masquerade as genuine product recommendations. False positive on clean messages was kept under 1.2%.
Threshold Tiers

Three-Tier Quarantine Architecture

Never execute binary hard drops on ambiguous spam. Use Jev estimated probabilities to partition traffic into three deterministic pipelines.

Spam Prob < 0.2062.0% of messages
Instant Pass (Post to feed)
Accuracy: 99.4% clean accuracy
0.20 <= Spam Prob < 0.8514.0% of messages
Quarantine (Hidden pending mod review)
Accuracy: Human triage safe
Spam Prob >= 0.8524.0% of messages
Hard Drop + Rate Limit Warning
Accuracy: 98.6% precision
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: Benign User Sharing a GitHub Repo Link

Accurate
"Hey guys, I wrote an open-source backup tool for Postgres here: https://github.com/alice/pg-safe. Free and MIT licensed, hope it helps!"
PredictionClean (Non-spam)
Ground TruthClean
Confidence0.94

Many naive keyword filters block all messages containing URLs. Jev correctly recognized genuine developer utility and community sharing context without triggering false flags.

Case 2: Cloaked Telegram Channel Recruitment

Accurate
"Great post admin! By the way, whoever wants to make 300 daily from home check my bio or t.me/fast_cash_2026. Only 5 slots left."
PredictionMalicious
Ground TruthMalicious
Confidence0.98

Classic social engineering template with urgency triggers. High confidence (0.98) triggers instant silent suppression.

Reproduction Harness

Run This Benchmark in Local Node.js

Run this script as a pre-commit filter or in an Edge Worker before writing comments to your persistent database.

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

export async function detectSpam(messageText: string) {
  const result = await experimental_evaluate({
    model: 'jev-latest',
    state: { message: messageText },
    questions: {
      spamCategory: {
        type: 'choice',
        instructions: 'Determine if this message is legitimate, gray-hat advertising, or malicious spam.',
        criteria: {
          clean: 'Genuine organic user content, constructive questions, or relevant links.',
          promotional: 'Unsolicited marketing, coupon codes, aggressive self-promotion.',
          malicious: 'Phishing URLs, financial scams, malware distribution, or impersonation.',
        },
      },
    },
  });

  return {
    category: result.answers.spamCategory.value,
    confidence: result.answers.spamCategory.confidence,
    probabilities: result.answers.spamCategory.probabilities,
  };
}
FAQ & Clarifications

Benchmark & System One FAQs

Why not use a standard regex or keyword blocklist for spam?

Spammers continuously mutate character encodings (homoglyphs, zero-width spaces, leetspeak). Jev operates over semantic sentence embeddings, identifying deceptive intent regardless of superficial text obfuscation.

How do we protect legitimate users from being banned accidentally?

Adopt the 3-tier quarantine pattern: never ban on a single borderline score. Route messages with 0.20–0.85 spam probability to a shadow-review queue where moderators can approve or confirm with one tap.

Explore More Jev Benchmarks & Tools