Checking the artifact instead of the status report: provenance, validation, receipts and evidence before claims.
One small task, followed through a crash, a corrected input and a retried side effect: which records let someone else carry on without losing progress, repeating what cannot be undone, letting a presumed-dead worker write, or trusting a result built from the wrong input.
A valid model answer became a chatbot failure because its response format had changed. The repair shows how a small interface assumption can spread through a system—and how a work graph helps carry the change.
Taking a modular instrument from the browser to C++ exposed a tiny difference in when an oscillator stopped. Finding it required several investigations to meet at one explanation.
Choosing tools is not the end of design. After a graph-shaped way of working is available, something still has to hold the drawing, advance the pieces that are ready, and decide whether what came back was acceptable.
Why “I turned that off” reports a belief rather than a verified state—and how practitioners can inspect the real surface of privacy, synchronisation and retention controls.
When a finished production returned the wrong article, the rejection became part of the work: recursive judgment, not a discarded miss.
A teardown of our own generated-content corpus — spin versus real. We counted both stores, drew a seeded sample of fifty, and read them. The number in the title was wrong.
We already know the shape of an accurate report that nobody acts on. This is not that. Here the monitoring is not failing.
In March 2026 we wrote a draft about agent damage and did not publish it. The number in that draft was thirteen.
We run enough automated, long-running work that we built layers whose whole job is to tell us what state everything is in. That was supposed to solve the problem. It didn't.
An operator console that combines current system health with durable work state, so exceptions can be investigated without switching between separate service consoles.
Research ingestion that follows selected sources, extracts usable material and keeps provenance attached, with explicit outcomes when evidence is missing or a run fails.
An evidence-first research workspace that keeps source provenance attached and makes missing or invalid material visible before interpretation.
A governance control plane that keeps operational evidence durable and routes sensitive transitions through policy and human review.
A bounded orchestrator-worker runtime keeps image-analysis results linked to their source records and validates them before durable save.