Runtime, workflow, recovery, tools, skills, observability and governance mechanisms.
Ninety-six internal controlled cases expose how the released scorer blends semantic judgment and call-count rewards. The failed rank-reversal prediction is retained, without inferring real model behavior.
An Orca run treated an Agent message as human permission. We tested six authorization-source and request-binding controls at CodeFlowMu's preparation interface.
A tool says the ticket was escalated just as the user asks about an order. In a bounded AG2 before/after comparison, one extra model call replaces the handoff result with a shipping answer. We compare that turn boundary with CodeFlowMu's delivery evidence.
Agents plan steps, use tools, and delegate work. Their plans can also expand beyond the task they were given. BoundaryXP uses 7+1 to organize requirements and mathematical methods to check budgets, permissions, ordering, and relationships across tasks.
The business system committed an order while the AI worker saw a timeout. We killed a process at three boundaries and used a blind retry without an idempotency key as a control.
A report has the right number of items, the current evidence version, and valid authorization, yet says B-23 where the task asked for B-22. Seven controlled cases show what rules and two models can catch, and what they cannot establish.
Nonempty placeholder chains and nonexistent endpoints can activate a causal label in a pinned attribution module. Twelve synthetic calls expose its local acceptance predicate and functioning negative controls, without measuring real-run misattribution rates.
Strict labels combine target votes with private stance. Mixed signals are excluded and missing stance is nonpositive; score interpretation needs evidence completeness and the retained population.
A controlled submission failed while the task stayed in the active directory, yet its metadata already said review. We compared retry and lost-receipt paths to ask how an agent team proves a transition completed.
Six internal substituted-environment probes of pinned original methods show that reward evaluation can append reference actions to the action record, including through a terminal step. Mutation can affect downstream provenance attribution but does not itself establish an incorrect reward; historical prevalence, repair effects and current-version behavior remain unverified.