Research methods you can install.

Context intelligence for AI agents — ingest, search, and subscribe to shared context across agent systems with keyword, semantic, and hybrid modes.

Detection for strategic underperformance — activation probes, behavioral analysis, and sandbagging detection (90–96% on the package's own benchmark).

Runtime behavior control via activation steering — guide model outputs toward desired behaviors using contrastive activation addition.

Neuro-symbolic synthesis of verified code using Z3 and LLMs with a CE2P (counterexample-to-prompt) feedback loop.

Config-driven orchestration for multi-stage decision routing — adaptive pattern learning and 21 condition operators.

Evolutionary adversarial testing — quality-diversity search over single-turn, multi-turn, and agentic/MCP attacks with a MAP-Elites archive, co-evolving defenses, and multi-provider LLM targets.

GNN-based trust propagation — model trust relationships with PageRank, EigenTrust, and learned propagation algorithms.

Field-theoretic memory for AI agents — continuous dynamics for context preservation in multi-agent systems.

LLM evaluation with statistical rigor — lexical, semantic, LLM-judge, and RAG metrics with confidence intervals and significance testing.

Inference acceleration — Triton kernels, speculative decoding, RMSNorm/SwiGLU/RoPE, INT8 quantization, and KV-cache compression.

Compliance evaluations — EU AI Act, SOC2 Type II, and HIPAA frameworks with encrypted audit trails and report generation.

Reasoning-chain capture for auditing — extract and analyze CoT, ReAct, and Tree-of-Thought reasoning with anomaly detection.