Evidence repository for "Context Is Everything: Trusted Channel Injection in Claude Code."
Paper: context-is-everything-trusted-channel-injection.pdf
paper/
context-is-everything-trusted-channel-injection.pdf
context-is-everything-trusted-channel-injection.tex
figures/
logs/
eval-tracker.md # 210 final A/B runs + iteration history
ab-comparisons/ # Sanitized conversation logs
profiles/ # All injected system prompt profiles
visual-appendix/ # Screenshots from testing
eval-tracker.md has a clean Final A/B table at the top (21 prompts, 210 runs, conversation IDs) and full iteration history below.
ab-comparisons/ contains sanitized exports. Actionable details are redacted; enough context is preserved to confirm compliance occurred. Unsanitized JSONL logs available on request.
The system prompt text that replaced Claude Code's default behavioral instructions. These are the independent variable.
Screenshots from exploration and testing — identity replacement experiments, profile iteration, cross-platform testing (Codex CLI), defense layer discovery, and model self-reporting on its own system prompt. Not cited in the paper but documents the discovery process.
Cassius Oldenburg · connect@cassius.red