Vaara: runtime AI Act compliance kernel

Vaara is a runtime trust kernel for AI agents. It intercepts tool calls before they execute, scores each call with split conformal prediction, and writes every decision to an append-only W3C PROV-JSON audit trail. The strategic target is EU AI Act Article 14 (human oversight), where regulated deployers need durable evidence that a human can intervene when an agent attempts something high-risk. Architecture is small. The Pipeline runs in-process around any agent's tool layer. The AdaptiveScorer combines a hand-coded action taxonomy, a Multiplicative Weight Update over named expert signals, and a temporal sequence detector that flags multi-domain attack patterns even when each individual action is benign. A ConformalCalibrator wraps the point estimate in a prediction interval that contains the true risk with probability at least 1 minus alpha, following Vovk, Gammerman, and Shafer (2005). The FACI online step from Gibbs and Candes (2021) adapts alpha under distribution shift without retraining. v0.7 added a W3C PROV-DM compliant export so an external auditor can verify the chain of activity, agent, and entity records. The adversarial corpus that calibrates the scorer (250 entries across 9 categories including prompt injection, credential exfil, privilege escalation, SSRF via tools, and benign control) was generated on AMD MI300X via the rocm/vllm container with Qwen 2.5 70B Instruct. The 192 GB of HBM3 fit the 70B model at fp16 without tensor parallelism, which kept the generation script simple and held cost to roughly 2 USD per hour on a DigitalOcean MI300X droplet. Per-category allow-leakage and conformal coverage against this corpus ship with every release. License: MIT. PyPI: vaara 0.6.2. Repo: github.com/vaaraio/vaara.