Visual Contract

The description states that Visual Contract is a four-step pipeline with:

1. A CLI and small web UI.
2. Input of a production brief.
3. Compilation into a YAML-based "Visual Contract" using GPT-5.6, containing up to 15 assertions across 10 types (e.g., exact text, object count, spatial relations).
4. Editing and re-validation of the contract.
5. An audit process where deterministic checks are done in Python and visual checks go to GPT-5.6 Vision.
6. A rule engine issues an assessment per assertion: pass, fail, or uncertain.
7. Output includes JSON, JUnit XML, HTML reports, and annotated images.
8. Exit codes plug into CI/CD pipelines.

The system is described as fully replayable without API keys, using recorded fixtures to simulate model responses.

Inference This appears to be a proof-of-concept or prototype tool for automating image validation in creative workflows, likely targeting teams producing AI-generated assets such as banners, mockups, or game assets.

The author claims the product is inspired by software testing practices, aiming to bring deterministic verification to image creation and review. It positions itself as a way to avoid human error ("squinting at a screen") and replace subjective feedback with structured checks.

Inference The positioning is early-stage, focused on solving a niche problem in AI-generated content workflows. It does not claim to be a commercial product or platform but rather a tool for internal use or experimentation.

The description states that the author ships AI-generated images (e.g., ad banners, product mockups, game assets) and was frustrated by manual review processes.

However, no explicit ICP or customer segmentation is provided beyond the author’s personal use case.

There is no evidence of a business model or pricing structure. The description does not mention:

• Revenue streams
• Subscription plans
• Licensing fees
• Paid features
• Monetization strategy

Inference The tool appears to be non-commercial, possibly an open-source prototype or hackathon submission, with no indication of monetization.

The project is built using:

• Codex for scaffolding and implementation.
• GPT-5.6 via OpenAI Responses API with Structured Outputs.
• FastAPI, Pydantic, OpenCV, Pillow, Typer, Jinja, pytest, GitHub Actions.
• A CLI and small web UI.

The author reports:

• Use of structured outputs to enforce type safety.
• Replay mode for testing without API calls.
• Security hardening through multiple rounds of review.
• Prompt injection resistance tested via adversarial examples.
• Holdout test sets to validate honesty of evaluation.

Inference The technical stack and delivery approach suggest a highly technical prototype, likely built by one person, with strong attention to correctness and reproducibility. It is not yet a scalable SaaS offering.

There is no evidence of:

• Customers or users
• Revenue or monetization
• Product-market fit
• Adoption beyond the author’s own use case
• Any form of product release or distribution outside GitHub

The project is described as a hackathon submission, and the only “demo” is a GitHub repo with replayable fixtures.

Inference This is an early-stage prototype, likely in pre-product-market fit or pre-commercialization phase. No traction or adoption data are evident.

The description does not mention any competitors or direct market comparisons.

Inference No competitive context is provided. The tool may be addressing a niche within AI-generated image validation, but there is no evidence of existing solutions in the space.

• Single-person team: The project has only one contributor, which raises questions about scalability and long-term maintenance.
• No commercial traction or customers: The tool is described as a personal solution, not a product for sale.
• Self-reported performance data: All metrics (e.g., 43/43 assertions correct) are from internal testing, with no external validation.
• No pricing or monetization strategy: No indication of how the tool would be sold or used commercially.
• Unverified model outputs: The system relies heavily on GPT-5.6, which is not independently verified for consistency or reliability.

1. What is your actual workflow with AI-generated images? Is this a personal tool or part of a larger team process?
2. Have you tested the system with real-world inputs from multiple users or teams?
3. How do you plan to scale beyond a single developer’s use case?
4. Are there any plans for monetization or commercial deployment?
5. What are the limitations of GPT-5.6 in this context, and how do you handle model hallucinations or inconsistencies?
6. Is there any intention to open-source or release this as a product?