Pillar 4 · Actor-Critic Drafting
Adversarial IDE
“Do not submit a brief until it has survived the Red Team.”
The Causel Adversarial IDE operates on a cyclical actor-critic workflow. As an associate drafts a motion, a critic agent continuously queries the knowledge graph for contradictions, a deterministic judge profile lookup retrieves the assigned judge’s vulnerability vector, and a hedge generator proposes strategic revisions — all before the document leaves the firm.
How It Works
The associate drafts a paragraph in the integrated editing environment.
The Critic node queries the knowledge graph for evidence contradicting the substantive claims of the draft text.
The Judge Profile Lookup node (deterministic, no LLM) retrieves the assigned judge’s posterior doctrinal vector and computes a vulnerability score for the argument pattern.
The Severity Aggregator combines the contradiction set and the judicial vulnerability score into a single risk value.
If severity exceeds the threshold, the Hedge Generator proposes a revision that preserves the legal argument while reducing the identified weakness.
The cycle repeats until severity drops below threshold or the three-iteration cap is reached.
Technical Detail
severity = max(contradiction_score, judge_vulnerability) // Must drop below threshold or cap at 3 iterations
DRAFT: Motion for Summary Judgment · Apex Corp v. Horizon LLC
"Mr. Hayes was unaware of any systemic thermal defect..." — Flagged: 98% contradiction confidence (PLAT-SLK-00942). Judge Rakoff denies SJ on this argument pattern 92% of the time. Hedge suggested: reframe around "reasonable inquiry" standard.
98%
Contradiction confidence on flagged paragraph
92%
Probability Rakoff denies SJ on this pattern
3
Maximum critic-hedge iterations per paragraph