Make “done” something we can check
This site’s film passed its file checks but stayed frozen in the browser. It had frames and a duration; the page did not move through them. The follow-up had to show the frames advancing as someone scrolled. That is the distinction I ask agents to make before starting: what result is needed, what must stay unchanged, and what evidence would establish that the result works. I inspect the artifact and keep the check’s limits beside its result.
Question my assumptions, too
In Receiver Reliance, the sample changed after collection began: 1,800 scheduled entries plus five remaining controls, with 619 assignments deferred. The paper gives that amendment and its reason beside the completion cost. Directing the study does not exempt my own claims from examination. If a premise is wrong, I want the evidence, what it changes and a better next step. My confidence, an agent’s agreement, the effort already spent and my permission to proceed do not prove it.
Give disagreement a fair test
Before the study was published, an internal reviewer independently recomputed all six estimates and bounds and examined twelve predetermined traces. The record identifies this as an internal check, not external peer review. I give a fresh reviewer the work and the criteria so they can challenge the approach. I then examine their findings, including disagreements, before deciding whether more work is warranted.
Give agents the context they need
The offline timeclock repair depended on an existing primary-key index. Knowing it was there let me choose a device-generated ID without adding a migration. I kept that choice and the rejected alternative in the project record. Agents can search those decisions and corrections, but I ask them to check the current source before acting. A handoff carries the immediate decisions, open questions and evidence; reusable guidance has a separate home.
Make corrections last
I recorded the frozen-film correction in the verification note: what failed, the page-level check that worked and what it could establish. That gives the next agent a procedure to use, with its limits, instead of a conversation to repeat. I also combine repeated rules and remove instructions that no longer apply.
Make the effort earn its place
I retired Catchwright in July 2026. I did not see a reason to expand it into a full subcontractor-management platform or compete for a narrow position between established vendors and a lower-cost entrant. I kept the engineering privately; the workflow survives as a demonstration. For new work, I set a useful outcome and a stopping point. Extra agents and checks cost tokens, time and attention, so I judge them by the result they help establish. Someone’s data or a live product warrants more scrutiny than a reversible experiment.