Task environments with hidden verification. Built because the interesting question is not whether an agent looks right, it is whether your scorer is harder to fool than your agent.
A medical-claims auditing task environment where the labels are verified by an independent engine before any record reaches disk, and the same agent is scored across three orchestration frameworks.
I did not trust labeled data I had not generated, so I built the scorer first and made the agents swappable arms behind it.
Deterministic rule engine pinned at F1 1.000 in CI. Extended thinking cut LangGraph recall from 1.000 to 0.600 and raised Vercel precision from 0.714 to 0.938.
A controlled ablation of stateless RAG, knowledge graph, and hybrid retrieval on one frozen corpus behind one shared typed interface.
Every retrieval comparison I could find changed the corpus along with the architecture, which makes the result unreadable.
Graph-only retrieval hit 98.7% Recall@5 and refused 100% of adversarial distractors where hybrid refused 10%. Two of four pre-registered hypotheses refuted.
The evaluation harness and evidence discipline behind a 14-run pilot of a multi-agent system on a contractor's live operations data.
The methodology is the part worth publishing. The capability number means nothing without the holdout design that produced it.
Two-database holdout, pre-registered thresholds, Cohen's kappa 0.95, 0 fabrications across 51 held-out decisions, one metric marked instrument-invalidated.
A customer-success pipeline combining classical ML, LLM labeling, causal inference, and an approval-gated agent, running fully offline on synthetic data with a known churn process.
A confounder planted on purpose is the only way to check whether your causal correction actually corrects anything.
DoWhy adjustment moves adoption's effect from a naive 30.4 pp to the true 11.7 pp, removing about 88% of the confounding bias. 0.806 ROC-AUC, 55 pytest tests.