Builds agentic systems. Evaluates them honestly.

Quorum fans out a finder agent per source file, then routes each finding to K skeptic agents that independently try to disprove it. Only findings that survive all K challenges ship. The result: a sharp precision improvement at a moderate recall trade-off, on a benchmark that includes adversarial prompt- injection traps to catch sycophantic verification.

The trace UI is a real deployed product at quorum.thomaspeng.ca, browsable below. Model routing is structured as a cost-first 4-tier ladder: DeepSeek for extraction, Haiku for mechanical checks, Sonnet for judgment, Opus only when convergence requires it.

Aegis pits an adaptive attacker agent against two harmless proxies: canary-string extraction and prompt-injection sentinel. Scoring is deterministic (exact-match, no LLM judge). The study measures whether reasoning models resist attacks better, and whether layered defenses close any gap.

The adaptation lift result (24% to 29.9%) was a null at small sample size, then became significant after scaling. The write-up frames this explicitly: "scaling is the legitimate power lever, not p-hacking." Vendors Quorum core/.

FieldAgent reads a real commercial contract and flags risk-bearing clauses with span, severity, and plain-English risk descriptions. Graded span-IoU against CUAD gold labels on 20 held-out contracts. No LLM judge in the success path.

The initial "+0.45 agentic lift" result was compelling but wrong: DeepSeek was hitting its output token limit and returning truncated results, making the baseline look worse. A fair rerun with adequate token budgets collapses the lift to +0.07 with overlapping confidence intervals. Reported exactly as measured. Vendors Quorum core/.