Miadi Review

System-1 Decision Engineering with Jev and TypeSafe AI

Sep 28, 2026 · model: manual-import · channel: How I AI

Watch source

In the chronicle

System-1 Decision Engineering with Jev and TypeSafe AI

Source: I’m using Jev more than Opus 5.5 or GPT-6. Here’s why. — How I AI https://youtu.be/-KIBgpGA_XI STATUS: I created my account and started a discussion on a tmux mia@gaia miadi-tide://miadi-jev and I have no idea if that was a success but for sure , we would integrate that layer into the miadi-factory to help the decisions (which imply questions: "Which decisions?", "which questions to ask to Jev and what is the context to send for having a decision?", "..." ; directly related (reviewing the same topic with 2 different presenters) ▥ review · ep 567 10 Levels of Jev: Architecting Fast, Cost-Effective Agentic Decisions

miadi-intent

decision-engine / structured outputs

  • definetly i’m getting that there really is something about decisions which basically we really need
  • LLM are not type-safe outputs!!! that explains why the miadi-chronicle got messy because LLM agent droped markdown files and are not forced at desired-output (matching a specific JSON type !!)
  • my God, message to all of my agent here, as you progress in the discussion it builds up something about are we ready to make a decision to call this tool? Do that in this and that so that really feels exactly so important. (08:20-09:00)

Summary

Claire Vo shows Jev, TypeSafe's "System One" decision model, built to classify rather than generate text. It returns typed primitives (a choice, a score, or a yes/no probability) in milliseconds. She uses it for pairwise clustering at scale, cheap pre-filtering in data pipelines, and interfaces that must react in under a second.

Academic Fields

Dual-Process Theory (Cognitive Psychology)

Kahneman's System 1 is fast, automatic judgement; System 2 is slow deliberation. TypeSafe names Jev after the first: it answers quickly and cheaply, and the frontier model deliberates on what remains. It is a metaphor for division of labour, not a claim that Jev reasons like human intuition.

Probability Calibration (Machine Learning)

A calibrated model's 0.8 is right about 80% of the time. Guo et al. showed modern networks are often overconfident. Noul returns exactly such a probability, and Choice gives one per option. Calibration holds over many predictions, so Claire's aggregate charts suit it better than any single verdict does.

Correlation Clustering (Algorithms)

Bansal, Blum and Chawla cluster items using only pairwise "same" or "different" judgements, the kind Claire gets from Jev when grouping pull requests. The field shows the result depends on which pairs are compared: ten comparisons per PR, not all pairs, means the clustering relies on how candidates are chosen.

Model Cascades and Routing

FrugalGPT and RouteLLM show that sending each query to the cheapest model able to answer it cuts cost sharply at similar quality. Claire's hybrid layering is a cascade: Jev filters and groups in bulk, and a frontier model summarizes the curated subset.

Mining Software Repositories (Software Engineering)

This field studies what version-control and issue data reveal about how software is built. Kalliamvakou et al. warn that GitHub data misleads easily; for example, many pull requests are merged outside GitHub's interface and look unmerged. Claire's effort-by-theme chart is an MSR analysis, and it inherits these pitfalls.

Affective Computing

Picard founded the study of systems that recognise and respond to emotion. Claire's voice demo detects emotional tone and changes the interface in real time. The field's standing question applies: labelling a voice as sad is an inference about a person, and it can be wrong.

Aspects

System 1 Fast Decision Architecture (02:50-04:30)

Traditional LLMs generate sequential text tokens with variable latency and high output costs. Jev takes unstructured text and returns a typed value with a probability, without generating tokens. Output is predictable in shape, not deterministic in content: code decides what to do with the probability.

TypeSafe Primitives (05:29-07:39)

Jev has three primitives: Choice picks one option from a list, with a probability for each; Score returns a number within a range, such as severity; Noul returns the probability that a yes/no statement is true. Calibration is claimed over many predictions, not for any single answer.

Sub-Cent Batch Pairwise Clustering (08:15-10:05)

By performing thousands of rapid pairwise similarity checks for fractions of a penny, Jev efficiently clusters vast unstructured corpora like pull requests or issue tickets before passing grouped summaries to larger frontier models for synthesis.

Hybrid Layering with Frontier LLMs (15:16-16:15)

Jev functions best as a frontline filter in multi-model architectures. It handles high-volume classification, deduplication, and routing tasks cheaply, freeing expensive frontier models to execute deeper analytical and generative reasoning on curated subsets.

End-results and Core Aspect

Core Aspect — High-Throughput Deterministic Decision Routing (04:50-05:28)

Jev decouples high-speed structured decision-making from conversational text generation. Developers can run very large numbers of classification, scoring and clustering calls at low cost, in background pipelines and in real-time interfaces alike.

Results

  • GitHub PR Categorization & Investment Tracking (07:48-10:38): As summarized from the video: over 1,700 merged pull requests compared in 17,000 pairwise checks (about ten per PR, not all pairs) in under two minutes for nine cents, then graphed as engineering effort by business theme. Not independently checked.
  • Developer Activity Meta-Analysis (11:37-12:55): Parsed thousands of local Claude Code and Codex session logs to categorize engineering work versus agent orchestration and content production across multiple months.
  • Real-Time YouTube Audience Signals Dashboard (19:10-21:39): About 4,500 YouTube comments classified for positive/negative sentiment and episode-topic requests, with sub-second semantic search over them (figures as summarized, not checked).
  • Sub-Second Voice Emotion to Quote Generator (22:52-25:17): Integrated OpenAI voice streaming with Jev to detect emotional sentiment instantly, dynamically shifting UI background colors and scoring appropriate reflective quotes in real time.

Related Review

10 Levels of Jev: Architecting Fast, Cost-Effective Agentic Decisions — IndyDevDan

Same tool, other use. Claire puts Jev over data: pairwise clustering, pre-filtering, sentiment. Dan puts it inside the agent harness: gating tool calls, timing compaction, screening files. In both, Jev decides first and cheaply; the frontier model works only on what survives.

Agentic Engineering Operating Level 260831 IndiDevDan: Dan puts contracts at L3, 'essential for long-term product stability' (4:30). This review's intent names what happens without one. Both stand behind rule I2 of ◇ foundation Bounded contexts and integration contracts.

Internal Usage

A proposal, not something Miadi does today.

  • T3 · Witness, the proposed inventory agent. It would read T1's binding line and each transcript, then record a session's mission, state and what needs William. Claire's session-log categorization is that sorting step: one cheap choice per session before a larger model writes the entry.
  • This review service. Finding today's related reviews took keyword searches that scored every match 0.5. Pairwise comparison of reviews, as Claire clusters pull requests, could propose Related Review links for a person to confirm.

Definitions

Team. In Miadi, a team is standing: a human lead and an agent lead, one desired outcome at one level. Building teams (T1, T2) make packages and practices others use; the witness team (T3) builds nothing and hears their work for William first (miadi-orchestration-kit#59).

System 1. Kahneman's name for fast, automatic judgement, used here as a metaphor. It does not mean deterministic: Jev returns probabilities. The decision becomes deterministic only when code applies a threshold to them.

Internal Usage. What a review could change for a named team or package; it proposes, it does not report.

Sources