Jev: Fast, Typed AI Decisions with System One Models

Jev introduces noul, a probability score from 0 to 1 for propositions, where the number is both the answer and the certainty guiding four steps and per-question thresholds.

Learn to use a one-dimensional score with ordered levels from level zero to level two, describing observable behavior rather than labels, using fractional scores and weighted composites for spectrum reasoning.

Master fast primitive selection by identifying propositions as nouls, exclusive buckets as choices, and ordered positions as scores, while avoiding fake scales, overlapping options, and noul sprawl.

Learn when and how to use json structures to define instructions, options, levels, and criteria, with worked examples and uniform option sets for cleaner model outputs and clearer probabilities.

Demonstrates one call, three primitives—Choice, Score, and Noul—driven by a Python routing rule, with policy thresholds guiding ticket routing and escalation decisions.

Define per-action cost thresholds to govern ai behavior, using three tiers: act, confirm, and human, and set action-specific safety bars based on consequences like read-only, money moves, and irreversible actions.

Pin the exact model version for production, verify on every call, and guard against drift when new versions ship; refuse actions on mismatched models and re-tune thresholds with each update.

Discover speculative fan-out and parallel sampling to speed AI decisions by asking all needed questions in one call, then branch in code for routing and confidence-based decisions.

Decompose a complex judgment into atomic scores for Python experience, leadership, and system design. Normalize and combine them with code-based weights to reveal the composite scoring decision.

Adopt the cascade to route intent swiftly through four tiers—Jev classification, plain code, frontier model, and human—reducing cost and latency while preserving accuracy.

Explore how retrieve-then-judge, fan-out, composite scoring, confidence gating, and cascade form a complete production system, and learn anti-patterns like sequential chaining and cross-type thresholding that hurt latency and explainability.

Jev demonstrates high-volume classification by splitting generation from classification, measuring 1,018 papers for $0.08 with 256 ms latency. Summarisation costs ~ $3.99, illustrating relevance filtering before expensive steps.

Re-score the shortlisted results after vector search by asking a relevance question per candidate and using a Score to boost top-1 accuracy. Over-retrieve to boost recall, then re-rank.

Screening RAG passages re-ranks retrieved text by three questions—relevance, contradiction, and presence of model instructions—reducing padding, cost, and exposure, while enforcing sanitization and layered safeguards.

Apply guardrails for llm input-output by implementing inbound and outbound checks with a severity score, tagging results by family (prompt injection, data leakage, unsafe content), and maintaining a latency budget.

Explore hierarchical classification across large taxonomies and learn how beam search fixes greedy errors by carrying uncertainty across top branches to rank complete paths.

This lab demonstrates screening and re-ranking in a RAG pipeline, reducing token spend and filtering out stale policies and injections to improve accuracy with fewer passages.

Compare sequential, fully concurrent, and capped concurrency using an async client to process twenty tickets, measure latency, respect rate limits with a semaphore, and see identical results.

Log every decision as structured data and query the log to reveal model version, full distribution, confidence, latency, cost, and a traceable id.

Build a gold set you can measure against by sampling real traffic, labeling by hand, over-sampling hard cases, and freezing a 200-row csv or jsonl in version control.

Compare model accuracy with calibration to reveal when predictions mean anything. Understand how to use reliability curves and expected calibration error to detect overconfidence and guide automated decisions.

Identify the nine jaggedness failure modes in fast, typed ai decisions and apply concrete workarounds, from literal reading and maths to dates, indirection, and generation.

Capstone brief: build Fernway's end-to-end support-desk autopilot and report per-ticket cost versus a three-cent baseline. Gate actions by justified thresholds, using deterministic code for routine cases and llm for tail.