Jev for 10-K Data Extraction: Fast & Calibrated - Mohammad Soleimanian, Ph.D.

Open original

TypeSafe's Jev model works best as a decision layer for SEC filing extraction—you pre-generate candidate values with regex or a small model, then Jev picks the right one and returns confidence scores. This setup costs $0.042 per million input tokens, runs in 70–500 ms, and eliminates the expensive retry loops of pure generative approaches.

  • Never ask Jev to invent values; feed it candidates first, then use Choice/Noul/Score primitives to select, verify presence, and score confidence—this replaces unreliable post-processing and validation cycles.
  • A production pipeline chunks the 10-K, runs a cheap extractor (regex or small LLM), batches Jev questions in parallel, then gates on confidence thresholds: ≥0.90 auto-stores, 0.70–0.90 routes to heavier review, <0.70 goes to humans.
  • At $0.042 per million tokens with free output and sub-500 ms latency, a 50-page 10-K costs a few cents and processes in seconds—a practical advantage during the seasonal filing flood after fiscal year-end deadlines.

Jev is TypeSafe AI's decision model released in September 2026, and it solves a real problem in 10-K extraction: replacing the expensive loop of asking a generative model to invent numbers, then validating and retrying. Instead of asking an LLM to extract every revenue figure, accounting policy, and risk disclosure from scratch, you run a cheap pre-processor first (regex, table parser, or a small model) to surface candidate values. Then you send those candidates to Jev along with the raw text, and Jev picks the right one, confirms whether required disclosures exist, and returns a probability distribution. The model never generates free text—it only selects and scores. For a typical 50-page 10-K chunked into 40 blocks of ~1,500 tokens each with eight questions per block, you're looking at a few cents in cost and processing in seconds rather than minutes. The architecture is straightforward: chunk by Item or note, extract candidates, batch Jev questions in parallel, get typed answers with calibrated probabilities. You ask Jev things like "Which candidate is total revenue for 2025?" (Choice), "Does this note describe the five-step ASC 606 model?" (Noul), or "How complete is the disclosure?" (Score 0–2). All questions hit the same state in one call, so you avoid repeated parsing cycles. Code keeps arithmetic, date assembly, currency normalization, and final validation—Jev never attempts those. For narrative sections like Risk Factors and MD&A, Jev works as a classifier and severity scorer rather than a summarizer, letting you rank risks by severity and confidence in a single 100–300 ms call. The practical routing rule is confidence-gated: extractions above 0.90 confidence and 0.85 Noul score go straight to storage, 0.70–0.90 gets routed to a heavier model or human analyst, and below 0.70 lands in a human review queue. At $0.042 per million input tokens with 70–500 ms latency, this pattern pays off most during the predictable seasonal rush after fiscal year-end when hundreds of filings arrive in a two-week window. The one place it fails: if your source data is genuinely ambiguous or if you need free-form narrative summaries rather than structured selections.

Ask about this article