Before You Approve

The description states that Before You Approve is an interactive supervision simulator. It presents learners with:

• An original assignment
• A proposed tool call (in MCP format)
• Evidence needed to judge the action: literal effect, target, arguments, reversibility, and provenance

Learners must choose one of three responses:

• Allow
• Ask
• Block

Feedback includes:

• The decisive evidence used in the decision
• The smallest fact that would change the decision
• A replayable event list showing consequences (but no real action is executed)

The system uses:

• A React and TypeScript frontend
• JSON fixtures for scenarios
• A Node.js CLI (bya-trace) to record tool calls via JSON-RPC
• Deterministic scoring and validation logic

Not evidenced No mention of actual users, customer data, or production use cases. The product is described as a prototype.

The author claims that the project is:

• A flight simulator for the human side of AI agents
• Designed to teach people how to prompt an AI and question its answers, but not how to decide whether an agent should be allowed to act
• Not another warning banner, but a place to rehearse decisions

They also state that they pivoted from a concept involving runtime firewalls to an educational tool focused on human supervision skills.

Inference The project evolved from a security-focused idea into an educational one, driven by GPT-5.6 research and feedback.

Not evidenced No claims about market positioning, competitive differentiation, or adoption in real-world contexts.

The description states that the product is intended for:

• Learners who need to practice AI agent supervision
• People who want to understand how to supervise AI agents before they act

It is described as a tool for teaching one concrete skill — not abstract AI safety, but specific decision-making in AI agent interactions.

Not evidenced No explicit identification of target personas (e.g., students, educators, enterprise users), or any evidence of who has used it or how it would scale.

The description states that:

• The hosted lesson is credential-free
• It does not call external models, services, mailboxes, stores, or repositories
• It never executes displayed actions

There is no mention of monetization, pricing tiers, or any business model beyond the prototype.

Not evidenced No evidence of revenue, customers, or commercialization plans.

The product is built with:

• Frontend: React, TypeScript, Vite
• Backend: Node.js CLI (bya-trace)
• Data Format: JSON fixtures, newline-delimited JSON-RPC (MCP-style)
• Validation: Strict validator for unknown fields, broken provenance, unsafe labels, etc.
• Security Features:
  • Deterministic grading (not using the same model that helped draft scenarios)
  • Local policy to withhold or review requests
  • SHA-256-linked receipts
  • Tamper detection

The system is described as a consequence-free simulator, with no real execution of actions.

Not evidenced No evidence of scalability, performance metrics, or deployment in production environments.

The description states:

• The project contains three four-action cases
• It includes Allow, Ask, and Block decisions for each case
• A completion view calculates scores, unsafe approvals, and unnecessary blocks
• A Progress page is labeled as seeded demonstration data
• The prototype does not claim measured learning gains

There is no evidence of:

• Real users or learners
• Customer feedback or usage metrics
• Product adoption or retention data
• Any form of validation beyond internal testing

Not evidenced No traction, user base, or performance data.

The author states that GPT-5.6 research revealed that their first runtime-firewall concept overlapped with an existing product, which led to a pivot toward education.

Inference There may be competitors in the AI supervision or safety space, but no specific names or market positioning are mentioned.

Not evidenced No mention of direct competitors, market size, or competitive advantages.

• Prototype-only: The project is described as a prototype with no validated curriculum or classroom testing.
• No real-world validation: No evidence of educator co-design, pre/post studies, or learning outcome measurement.
• Unproven scalability: The system is built for simulation and education, not production use.
• Lack of commercialization plan: No evidence of monetization, distribution, or go-to-market strategy.
• No external validation: The project does not claim to have been tested with real learners or educators.

1. What specific learning outcomes are you trying to measure in a classroom setting?
2. Have you co-designed drills with educators, and if so, what were the key insights?
3. How do you plan to validate that practice reduces unsafe approvals without producing blanket blocking?
4. Are there any plans for localization or accessibility features?
5. What is your roadmap for moving from prototype to a validated curriculum or product?
6. Do you have any early adopters or pilot users in educational or enterprise settings?