JevPulse

A high-throughput qualitative consensus engine for YouTube comments. Instead of asking a chat model to summarize a comment section, it evaluates every comment individually with Jev and computes agreement deterministically in code.

A reasoning model is used once per video: Gemini Flash reads the title, description, a transcript excerpt and 20 sample comments and outputs a structured rubric. Jev then evaluates every comment against Layer 1 universal invariants and the Layer 2 video-specific rubric.

Layer 1 asks is_opinion (noul), comment_type (choice), specificity (score) and has_claim (noul). Layer 2 adds one noul per rubric criterion generated for that video. Answers are typed and calibrated probabilities rather than generated text.

Comments in a batch are bundled into a single shared state, and Jev evaluates all Layer 1 and Layer 2 questions across all batch comments concurrently in memory. The README notes that adding questions barely impacts latency because decisions evaluate in parallel over shared memory.

Silence is not dissent: a comment is classified as opposing only if it has a low affirmative score and actively exhibits critical communicative intent (criticism, disagreement or correction). Signal gating also filters low-entropy comments using calibrated isOpinion and specificity thresholds before consensus math.

The README's comparison table lists Jev at $0.042 per million input tokens with outputs free, against roughly $2.50–$3.00 per million for chat models. The per-3,000-comment row is cut off in the available excerpt, and all figures are the project's own.

A live demo is linked at jevpulse.vercel.app, and the repository includes a playground app. The README also documents confidence-calibrated provenance: supporting and dissenting evidence rank by absolute semantic confidence rather than social likes, preventing viral jokes or spam from distorting the report.