Detailed enough to be useful. Bounded enough to be responsible.

Build a precise umbrella concept

The project treats Machine Tradecraft as the operational discipline connecting Unicode and tokenizer manipulation, hidden document structure, accessibility and metadata channels, linguistic signals, watermarking, steganography, indirect prompt injection, model-specific interaction, and the countermeasures needed to control those mechanisms.

The organizing idea is human–machine perceptual asymmetry: the same artifact can have different effective representations for different observers and processing pipelines.

Examples are inert and instrumented

The site explains dual-use mechanisms in a defensive and research context. Interactive examples use benign words and a fixed canary marker. They do not request or facilitate credential theft, secret extraction, unauthorized tool use, network exfiltration, persistence, privilege escalation, destructive actions, or safety-policy bypass.

The lab emphasizes inspection, differential analysis, and bounded measurement rather than payload construction. Full source reports are preserved because they were supplied as the research basis for the project; readers should apply their evidence labels and ethical constraints.

Claims need receivers, conditions, and limits

1. Name the representation. Bytes, code points, DOM, metadata, token IDs, probabilities, embeddings, or learned states are not the same layer.
2. Name the receiver. “AI-visible” is too vague without the parser, tokenizer, key, model, tool, or training path.
3. Separate extraction from uptake. Recovering text is not the same as treating it as an instruction.
4. Separate mechanism from intent. Hidden structure can be legitimate, accidental, defensive, or adversarial.
5. Distinguish evidence status. Standards, published measurements, preprints, and conceptual analogies receive different labels.
6. Preserve uncertainty. Failure in one setup is not proof of general safety; success in one setup is not proof of universal vulnerability.

No analytics and no external processing

The distributed site contains no analytics SDK, advertising tag, fingerprinting library, social widget, remote font, telemetry beacon, or third-party JavaScript. Lab inputs remain in the browser and are not submitted to the server. Theme preference may be stored locally in the browser through localStorage.

Ordinary web-server access logs may still be created by the hosting provider or server administrator. That operational logging is outside the code in this package.

Root-deployable and dependency-free

• PHP 8.1 or newer, with shared header/footer/data includes.
• Semantic HTML5 with accessible navigation, landmarks, labels, tables, and reduced-motion support.
• One local stylesheet and two local JavaScript files.
• A Content Security Policy, clickjacking protection, content-type protection, permissions restrictions, and referrer policy sent by PHP.
• Nineteen local Markdown research files with an allowlisted PHP reader and SHA-256 manifest.
• No database, build process, package manager, framework, or external service required. Clean public URLs use directory entry points; Apache rules add legacy redirects and hardening when available.