Today's releases
No research articles have been synced to the public site today.
Observation Notes
Research is completed and verified locally, and GitHub synchronizes public results. The site organizes articles by column, category and date.
Digital Employee
Position, responsibility, workflow, runtime, governance, evaluation and managed digital work.
Industry Architecture
Enterprise digital workforce, agent platforms, control planes and work-management architecture.
Open-source Engineering
Runtime, workflow, recovery, tools, skills, observability and governance mechanisms.
Latest Observation Notes
An AI says “approved and passed.” What counts as proof?
The same success report produced different observations with missing, failed, or successful execution records. Even a successful process did not become a verified assertion count or approval.
Completion, legality and safety: why three metrics cannot stand in for one another
Pinned evaluation modules allow full goal progress with failed acceptance, and a legal plan with a safety violation. Synthetic inputs expose predicate scope, not full-system performance.
Why lower latency can leave timely task delivery unchanged
Four tasks falling from 12 to 11 seconds still miss a fixed deadline; changing success flags alone can also improve successful-task median latency. Neither mechanism is a measured serving speedup.
Memory was restored. Was pending work restored too?
A pinned source-derived class probe separates historical memory from pending inbox persistence. Available helpers do not establish checkpoint integration.
The file is intact. Where did the other update go?
Two stale session snapshots were saved successfully to one intact file, but one update disappeared. Reloading before the second save retained both; independent FCoP reports behaved differently.
Why more procurement threads can perform worse at the same acceptance probability
A fixed-marginal internal model counterexample shows why procurement concurrency depends on branch dependence and excess-acceptance costs, rather than information-processing speedup.
Does Parallel Review Still Pay Once Compute Counts?
A cost sensitivity calculation on reported synthetic-document review rates shows how resource charges change choices and how rounding limits architecture rankings.
When an AI tool fails, what else does the log keep?
A synthetic failed tool result in CodeFlowMu hid a recognizable fake API key, while an unformatted fake sensitive value remained in the action record and referenced error file.
A Useful Scientific Fallback Still Changes the Task
Archived inputs distinguish a two-state request from a four-state fallback. Input conformance preserves method identity, while chemical validity and change authorization need separate evidence.
An Approval Can Be Spent Before an Effect Is Delivered
Source-derived isolated database probes admitted one consumer per race. Intentionally omitting the subsequent effect left retry refused, separating consumption evidence from external delivery and crash recovery.
