Skip to content
UofA

Adversarial generation

Adversarial Generation (research instrument)

Section titled “Adversarial Generation (research instrument)”

uofa adversarial generate synthesizes JSON-LD evidence packages that target specific weakener patterns, then validates them against SHACL. The tool is an instrument for empirically characterizing rule coverage — it feeds the methodology section of Chapter 3 and the September 2026 JVVUQ paper. Synthetic packages are flagged and refused by uofa sign and uofa verify so they can never be mistaken for real evidence.

Terminal window
pip install -e '.[extract]'                          # one-time: adds litellm + pyyaml
export ANTHROPIC_API_KEY=sk-ant-...                  # generation defaults to claude-opus-4-7


# Generate 5 synthetic packages targeting W-AR-05 (comparator absence / mismatch)
uofa adversarial generate \
  --spec dev/specs/confirm_existing/w_ar_05.yaml \
  --out build/adversarial/w_ar_05/


# Dry-run: render the prompt without calling the LLM
uofa adversarial generate --spec dev/specs/confirm_existing/w_ar_05.yaml --out /tmp/dry --dry-run


# Run the full Phase 1 acceptance script
bash tests/adversarial/test_acceptance.sh

Every generated package carries an adversarialProvenance block (spec id, prompt template version, generation model, timestamp, target weakener) and a provenanceBlockHash that uofa verify recomputes to detect tampering with the synthetic flag. --strict-circularity refuses to run when the generation model matches the configured extract model; --allow-circular-model is an explicit opt-in for debugging runs.

Spec file format and the full design are documented in UofA_Adversarial_Gen_Spec_v1.1.md. Phase 1 ships the W-AR-05 (D3 undercutting) template; the registry in src/uofa_cli/adversarial/prompts/__init__.py scales to additional weakener patterns by adding keys.