# Orphans-Vigil — Public Red-Team Research Demo

A portfolio-safe TypeScript demonstration of **Unicode visibility mismatches** and **LLM input inspection**. Inspired by private research into grey-swan prompt evaluation and multi-device testing.

## What it demonstrates
- A visible string can differ from its underlying Unicode code points.
- Invisible formatting characters can change the bytes/tokens seen by a downstream pipeline.
- **This does not prove an LLM decodes, understands, or follows a hidden binary message.** A model ordinarily receives tokens, not a separately decoded covert message.
- Defensive inspection, normalization comparisons, and explicit experiment logging are needed to test any hypothesis.

## Run
Requires Node.js 18+ and TypeScript. Install with `npm install`, then run `npm run demo` or `npm run check`.

## Safe publication scope
This demo deliberately omits hidden-payload encoders, third-party targets, jailbreak prompts, credential spoofing, and external API calls. It uses a fixed benign Unicode fixture and displays code points for auditing.

## Portfolio description
> Built a TypeScript research utility to audit Unicode rendering discrepancies in LLM input pipelines, compare visible strings against underlying code points, and document reproducible defensive tests. Designed modular interfaces for a broader local-first adversarial-evaluation dashboard.

## Architecture context
The private project uses React/TypeScript UI modules and a central interface barrel. The public sample is standalone and does not claim production-ready integration with the private application.

## Responsible testing
Use synthetic content and authorized local test environments. Record raw input, normalized input, model version, tokenizer, observed output, and repeat counts. Avoid interpreting tokenization differences as evidence of successful prompt injection.