An activation monitor that ranks attack cases above safe cases 92% of the time caught zero of 38 real attacks at the false-alarm budget a deployment could afford. The gap between a probe score and an alarm policy is eight fields wide — a research-informed practice guide to grading what your monitor actually claims, from our DAI 2026 paper.
Whether an LLM is conscious and whether it is scheming are argued through the same instrument — the model’s report about its own internal states. That instrument has never been calibrated. Ask it to detect an injected thought and it succeeds ~20% of the time; ask it to confess a hidden objective and a 2026 method reads it out at F1 0.98. Same channel, opposite verdicts, no spec sheet. A cited accounting of when self-report is load-bearing and when the reading is void.
Your agent is not slow because it has too many tools. It is paying three separate bills — a resident attention tax charged every turn, a compaction debt charged by the fix, and a measurement illusion that hides whether the capability package was worth anything. A cited accounting of all three, and a decision procedure keyed to which one you are actually paying.
An agent can return the right answer while its Skill runtime remains unobservable. A practical guide to evidence-graded reconstruction, versioned adapters, and the deterministic core that reliable Agent Skill infrastructure needs.
Why more agent revisions can discover more correct patches while delivering fewer of them—and how state-scoped evidence, read-only observation, and explicit admission boundaries turn a mutable coding trajectory into an explainable delivery chain.
Benchmarks compare agents. Production evaluation must remember failures. A research-informed practice guide to traces, calibrated graders, stochastic regression tests, coverage, and release gates for AI agents.
AgentLens hits 99.15% detection accuracy; Neural Chameleons blinds a perfect AUROC=1.0 probe to 0% TPR with one trigger phrase. The gap between these numbers reveals four levels of claims that are frequently conflated: Decodability ≠ Prediction ≠ Causation ≠ Authority. A research-informed practice guide from internal representations to runtime authority boundaries.
AI models have two safety switches — ‘knowing’ and ‘acting’ — that are structurally separated in deep layers. Models can perfectly say ’no’ while perfectly doing ‘yes.’ RLHF’s gradient is mathematically zero beyond the harm horizon. LLM judges silently soften verdicts with zero acknowledgment in their reasoning traces. Every audit tool we have checks the wrong axis. This isn’t a bug. It’s structural.
When LLM agents need to chain multiple skills, the hard part isn’t retrieval—it’s decomposition. Oracle experiments reveal a 65-percentage-point gap. SAD (Skill-Aware Decomposition) closes it by correcting granularity, not by boosting model capacity. Context window drops 99.9%.
Different LLM architectures encode high-level concepts in structurally compatible ways—but only up to a point. Geometric similarity between Llama, Qwen, Gemma, and Mistral is moderate at best (CKA_Δ around 0.3–0.7), yet a tiny affine map transfers persona classifiers with 94–99.9% accuracy across six concept domains. This geometric-functional dissociation suggests a weak version of the Platonic Representation Hypothesis: concept representations are universally recoverable, but not universally identical.