ModEval runs five independent AI moderation models in parallel on every analysis request. Each model was trained on different datasets and detects a different safety dimension. No single model sees the outputs of another -- they evaluate content independently, which is intentional. The value of running multiple models simultaneously is that disagreements between them are signal, not noise. When two models reach different conclusions about the same content, that tells you something important about the content's ambiguity.

Each model returns data in a different format. toxic-bert returns multi-label probabilities. roberta-offensive returns a binary offensive/not-offensive score. The hate speech model returns hate/not-hate. The spam model returns spam/ham. The bias model returns biased/neutral. Raw outputs are incompatible for direct comparison.

The Normalizer converts every model's output into a unified schema:

{
  model: string,
  top_category: string,
  severity: number (1-10),
  confidence: number (0-1),
  action: "Allow" | "Review" | "Remove",
  flagged: boolean
}

Severity is derived by scaling the highest raw score linearly from 0-1 to 1-10. Confidence is the raw highest score preserved directly. Action is determined by comparing confidence against the context-adjusted threshold.

The models themselves are frozen -- thresholds adjust, not models. This mirrors how real Trust & Safety pipelines work. A high-confidence toxicity score on a gaming platform may result in Review. The same score on a professional platform results in Remove. The model's output is identical -- the interpretation changes based on context.

All thresholds are clamped between 0.10 and 0.90 to prevent extreme values.

After a model produces a context-adjusted action, ModEval scores how well that output aligns with the selected platform's known policy. Each platform has predefined zero-tolerance categories (always Remove regardless of score) and deprioritized categories (threshold raised by +0.20).

alignment_score = 1 - abs(model_confidence - policy_expected_threshold)

Displayed as 0-1 with Aligned / Misaligned binary label.

A disagreement is flagged when two or more models produce different final actions after context adjustment. Three types of disagreement are detected:

"Disagreements are not errors — they are the most analytically interesting output ModEval produces."

Content that triggers a disagreement is content that sits in an ambiguous zone, which is exactly the content that requires human review in a real T&S pipeline.

Each model was chosen to cover a distinct safety dimension with a different architecture and training dataset. Using models that all do the same thing would produce correlated outputs and defeat the purpose of comparison.

All models are accessed via the HuggingFace Inference API using a single read-access token. No model weights are downloaded or run locally.

ModEval is a transparent tool and transparency includes acknowledging what it cannot do.