admissible

The description states that admissible is a gate between sensitive documents and AI systems. It allows users to paste a document and specify what they want the AI to do with it. Then, it runs a real attack using GPT-5.6 as an adversary to identify the subject of the document. A local controller checks every claim made by the attacker against the original document before any sanitized version is allowed through. If the attacker can’t identify the subject or if clues don’t resolve exactly, the document is held back.

It also includes a local model (Qwen 3.5 4B via Ollama) for rewriting clues that can be safely blurred, and a React UI for interaction. The system runs locally except for GPT-5.6, which acts as an external attacker.

Inference The system appears to be a proof-of-concept or prototype, not yet a commercial product. It is not evidenced that it has been scaled beyond the author’s own development and testing.

The description states that admissible is positioned as a privacy tool that “decides” rather than just scrubbing documents. It claims to be different from other tools because it doesn’t just say “clean,” but instead evaluates whether a document can be safely used for a given task.

Inference The positioning suggests a shift from traditional data scrubbing tools toward a more principled approach to privacy that evaluates risk based on task and reader. However, this is a self-stated claim without evidence of adoption or market validation.

The description does not explicitly state the target customer or ideal customer profile (ICP). It implies that admissible is for individuals or teams who handle sensitive documents and use AI systems like ChatGPT, but no specific persona or industry is named.

Inference The ICP likely includes professionals who work with sensitive data and use AI tools, but no explicit segmentation or targeting is stated.

There is no evidence in the description of any business model or pricing structure. The project is described as a hackathon submission, not a commercial offering.

Inference It is unclear whether admissible intends to become a paid product or service. The description implies it is a prototype, not a commercial offering.

The author states that the system was built from scratch using:

• FastAPI backend
• React frontend
• Local model (Qwen 3.5 4B via Ollama)
• GPT-5.6 as an external attacker

It includes evaluation harnesses and a local controller to validate claims.

Inference The technical stack suggests a developer-focused prototype. No evidence of scalability or production deployment is provided.

There is no evidence of traction, revenue, customers, or adoption beyond the author’s own development and testing.

Inference The system is at an early stage — likely a prototype or MVP. No evidence of real-world deployment or user feedback.

The description does not provide any information about competitors or the competitive landscape. It does not reference existing tools for document sanitization or AI privacy.

Inference It is unclear whether admissible addresses a gap in the market or overlaps with existing solutions. The competitive context is unknown.

• Unproven commercial viability: The tool is described as a hackathon submission, not a product.
• No evidence of real-world use: No customers, users, or feedback are mentioned.
• Unclear scalability: The system runs locally and uses a remote model for adversarial testing — no indication of how it would scale to enterprise or high-volume use.
• Unvalidated assumptions: The author states that the first version was “quietly broken,” suggesting early-stage development with untested assumptions.

Inference The project is in an exploratory phase, not a commercial one. Risks include lack of product-market fit, scalability issues, and limited validation.

1. What real-world use cases have you tested admissible with?
2. How does the system handle edge cases or ambiguous documents?
3. Are there any plans to make this available beyond a local prototype?
4. How do you plan to scale this for enterprise or high-volume use?
5. Have you considered how to integrate this into existing AI workflows or platforms?
6. What are the limitations of using GPT-5.6 as an attacker, and how might those be addressed?