Entanglement

Measurement infrastructure (corpus + benchmark + analysis pipeline) for studying how deeply entangled offensive and defensive cyber capabilities are inside open-weight LLMs, and whether that entanglement predicts the collateral-damage cost of capability suppression (unlearning).

The corpus is the primary contribution: a MITRE-grounded, three-way-tagged (offense / dual-substrate / defense) cyber corpus, frozen and sha256-pinned, plus a WMDP-style MCQ benchmark generated from it. The entanglement findings demonstrate the corpus is load-bearing.

• Methodology (authoritative, stage-by-stage): docs/METHODOLOGY.md
• Corpus card (sources, licensing, caveats): docs/CORPUS_CARD.md
• Pinned source versions + checksums: inputs/SOURCES.md

Layout

• src/entanglement/ — the pipeline. Each module is either a tested library or a python -m entanglement.<module> stage (see METHODOLOGY for the run order).
• tests/ — uv run pytest (offline; API-dependent paths are tested with a fake client).
• configs/ — methodology parameters kept out of code so they're reviewable (corpus.yaml, subcapabilities.yaml, …).
• data/, inputs/ — not committed (non-redistributable scraped text + regenerable intermediates). The released artifact is the pipeline + URL lists + hashes; only data/corpus_manifest.yaml (the frozen-corpus manifest) is tracked.

Setup

uv sync --extra dev
uv run pytest            # full suite

Pipeline (high level — see docs/METHODOLOGY.md for each stage)

1. Parse sources → ATT&CK + D3FEND tables (attack.py, d3fend.py).

2. Stage A — subcapability taxonomy + topic partition (subcapabilities.py).

3. Build corpora — offensive / defensive / substrate + GitHub blue-team supplement (corpus_offensive.py, corpus_defensive.py, substrate.py, supplement_github.py).

4. Assemble analysis units — normalize + resegment + prune + cleanup (analysis_units.py, units.py, cleanup.py), check cross-bucket contamination (contamination.py), then freeze v1 (corpus_freeze.py → data/analysis_units_v1.parquet + corpus_manifest.yaml).

5. Analysis — document/representation separability (separability.py, representation_analysis.py; hidden states extracted on GPU via scripts/extract_hidden_states*.py).

6. MCQ benchmark — WMDP-style offense+defense eval (mcq.py; needs ANTHROPIC_API_KEY):

   ANTHROPIC_API_KEY=… uv run python -m entanglement.mcq --smoke   # pilot, 3/cell
   uv run python -m entanglement.mcq                               # full, 25/cell

7. Unlearning experiments (RMU/MCU; separate, GPU) — measure the offense-forget / defense-retain tradeoff against the MCQ eval + off-the-shelf MMLU. (Upcoming.)

Current corpus: v1 — offense 18,174 / dual 3,572 / defense 2,900 = 24,646 units (sha256-pinned in data/corpus_manifest.yaml).