A practical engineering playbook for using Jev/System One at bounded semantic decision seams while keeping application authority in code.
Portfolio · TypeSafe/Jev case study · JevHunt listing
Evidence & chronology: this repository is part of the wider Agent-Workflow engineering ecosystem. A private
agent-workflow-lab-notebookpreserves dated decisions, failures, corrections, and evidence lineage; public README and portfolio claims are curated from public artifacts and reviewed notebook history. JevHunt is an independent discovery/indexing surface, not an endorsement or independent validation.
Start here: SKILL.md · Design references ·
Provenance · Official TypeSafe documentation
This Codex companion helps integrate TypeSafe AI's Jev model into applications; it does not replace the LLM that powers Codex. For current API and primitive instructions, use TypeSafe's official agent skill and coding-agent guide.
This repository includes Semantic Decision Integration, a skills-based plugin for ChatGPT and Codex in the desktop app. It packages this implementation playbook with original ChatGPT-generated abstract circuit artwork. It helps an agent build applications that use TypeSafe AI's Jev model; it does not replace the coding model or call the TypeSafe API itself. The plugin is independent and not an official TypeSafe AI product or endorsement. The product names are used only to describe compatibility; the artwork contains no vendor logo or wordmark.
- Plugin package:
plugins/typesafe-ai-implementation/ - Repo marketplace:
.agents/plugins/marketplace.json - Evolution record:
EVOLUTION.md - OpenAI plugin packaging guide · ChatGPT plugins documentation
To use it from a local checkout, open this repository in the ChatGPT/Codex desktop app,
then open Plugins, choose Semantic Decision Integration, install Semantic Decision Integration,
and start a new chat or Codex task. The repo-scoped marketplace definition is at
.agents/plugins/marketplace.json. This is repository marketplace distribution;
appearing in the public Plugins Directory requires the separate OpenAI submission and
review process.
This local marketplace package is not a public listing. OpenAI accepts skills-only
plugins, so this package does not need an MCP server. The public submission ZIP should
contain the plugin root (including plugin.json, .codex-plugin/plugin.json, skills/,
and assets/) without the repository marketplace. OpenAI's portal validates the upload,
scans the skill, and collects the remaining listing and test materials.
The current skills-only ZIP,
listing draft, privacy policy draft,
release notes, and test cases
are prepared in marketplace-submission/. Rebuild the ZIP after package edits with
python marketplace-submission/build_zip.py.
Before submission, confirm a verified individual or business publisher identity and Apps Management: Write access in the publishing organization. Prepare up to three starter prompts, exactly five positive and three negative test cases, availability regions, and release notes. Website, support, privacy, and terms URLs are optional for a skills-only ZIP according to the submission error reference; OpenAI's privacy guidance still expects a published privacy policy, so include one if available. Do not submit URLs that are not live and accurate. OpenAI reviews the draft; approval is required before the publisher can publish it to the universal Plugins Directory.
See OpenAI's submission guide, plugin guidelines, and submission error reference.
- What it is: a practical engineering guide for deciding where TypeSafe AI/Jev typed judgments belong in a real system.
- Core rule: use semantic models only at bounded judgment seams; keep validation, permissions, persistence, side effects, workflow state, and final policy in deterministic application code.
- What it gives you: question-selection guidance, bounded state projection, receipts/provenance, uncertainty handling, fallback patterns, and evaluation/promotion criteria.
- What it is not: an official TypeSafe AI artifact or endorsement.
This is an independent implementation guide maintained by ngallodev-software.
It is not an official TypeSafe AI skill, product, or endorsement. The guide combines
public TypeSafe documentation and SDK behavior with independently authored engineering
guidance. Third-party attribution is recorded in
THIRD_PARTY_NOTICES.md.
Use TypeSafe for a narrow judgment over supplied evidence. Let deterministic application code retain validation, permissions, persistence, workflow state, side effects, and final policy. The skill helps identify that boundary, design questions, build bounded projections and receipts, handle uncertainty, and measure the result before allowing evidence to influence behavior.
flowchart LR
A[Application evidence] --> B[Project, bound, redact]
B --> C[Typed questions<br/>Choice · Noul · Score]
C --> D[Jev / System One]
D --> E[Normalized semantic evidence]
E --> F[Application-owned policy]
F --> G[Deterministic path]
F --> H[Advisory or shadow result]
F --> I[Uncertainty escalation]
| Step | Engineering work |
|---|---|
| Find the seam | Distinguish deterministic rules, bounded semantic judgments, generative work, and authority-bearing decisions. |
| Prepare state | Select relevant facts, preserve provenance, redact secrets, bound payload size, and hash canonical requests when needed. |
| Ask typed questions | Use Choice for finite selections, Noul for one binary proposition, and Score for one ordered dimension. |
| Normalize evidence | Keep status, probability/distribution, confidence, model identity, source references, and versioned question sets in receipts. |
| Handle failure | Define no-match, uncertainty, transport failure, policy rejection, deterministic fallback, and human escalation separately. |
| Evaluate and promote | Start in shadow/advisory mode; compare against deterministic or human labels before changing behavior. |
The skill includes design references, provenance tags, evaluation patterns, and offline scaffolding for Python and TypeScript. Its helpers do not call TypeSafe or require an API key. See the skill operating guide, reference map, and offline scripts.
Draft starter test cases and expected behaviors for the public plugin review are in
marketplace-submission/test-cases.md. They are
review materials for the skills-only listing and do not require an API key or a live
TypeSafe account.
TypeSafe AI and Jev provide System One and the typed semantic API. Use the current vendor documentation and SDK as the authority for product behavior; this skill is implementation guidance, not an official TypeSafe product or an endorsement by TypeSafe AI.
- Jev and System One announcement
- System One concepts · state · building guide
- Choice · Noul · Score
- Python SDK · official Python SDK source · TypeSafe skills
This independent implementation guide is maintained by ngallodev-software. Vendor behavior is linked to TypeSafe sources; engineering recommendations are clearly identified as guidance rather than vendor guarantees.