What you will need to show when an agent gets it wrong

Three questions that define accountability, the four places the trail goes cold, three you can close yourself — and the paperwork that is not worth keeping.

Accountability is a property of your records, not your intentions.

At some point somebody is going to ask you to explain something an agent did on your team's behalf. You have reviewers. You have approvals. You believe there is a human in the loop. Believing that and being able to demonstrate it are different states, and the difference is discoverable in advance.

This volume is three questions, four break points, and the four records that close them. It assumes you have read nothing else in this series.

The spine of the volume: a reconstruction starts at the output, which is the only thing you have, and works backward toward the decision, which is the thing you need.

What moves: a query enters at the output end of the chain and travels leftward along it, arriving at the decision.

The three questions

Who decided this? A record naming a person and a moment.
On what evidence? Whatever they were looking at when they decided — not what is available now, which is different.
What were they told at the time? The one people leave out, and the one that decides whether the first two mean anything.

Three of the four break points cost almost nothing to close, and you can close them yourself without a mandate. The fourth is real engineering work and is not yours alone — and the volume says which is which rather than leaving you to promise all four.

One step is about what not to do. A great deal of what gets installed after an incident costs real time and produces nothing anybody can use. A volume about accountability that could not name that paperwork would be an argument for more process, and this one is not.

What you can do afterward

Try to reconstruct one decision your team made last month, and find out where you stop. Tell an approval from a name beside a decision. Stop writing a sentence about a human in the loop that survives no examination. And close three break points this month without asking anybody.

The three questions, the four records and the break points in order of cost are on one page you can print.

Part 1 — what accountability actually requires

01The questions you will be asked afterwardafterward you can hear the questions in the order they actually arriveexhibit

02Accountability is three questions, and all three need recordsafterward you can say what accountability requires without using the wordsourced

03Why a habit is not evidence about a caseafterward you can tell a description of a practice from evidence about an instancereasoning

Part 2 — where the trail goes cold

04An unversioned configuration ends the reconstructionafterward you can find the most common break, and close it this weekreasoning

05What turns an approval into something reconstructableafterward you can tell an approval from a name beside a decisionreasoning

06Why you cannot simply run it again and seeafterward you can know which inputs to capture, and which to give up onreasoning

Part 3 — recording enough, and not more

07The four things worth recording, and nothing moreafterward you can compare four items against what you currently keepsourced

08Recording that costs real time and buys nothingafterward you can refuse the paperwork that looks like a safeguardreasoning

09Why an unowned agent is a gap with a delay fuseafterward you can put a current name against every agent your team runsreasoning

Part 4 — what to institute, and what to stop saying

10A human in the loop is usually not a controlafterward you can stop writing a sentence that survives no examinationsourced

11What you can institute without asking anyoneafterward you can close three of the four breaks without a mandatereasoning

12Reading a real incident for where the trail endsafterward you can answer the three questions about something you runexhibit

Every term this page uses

reconstruction: Working out after the fact who decided something, on what evidence, and what they were told at the time.
chain of custody: The unbroken trail from a decision to the thing it produced. One missing link ends it, however good the rest is.
provenance: Where something came from, recorded well enough that somebody else can check.
versioned: Kept as a numbered copy, so the thing that ran can still be examined after it has been changed.
reproducible: Able to be run again and produce the same thing, because everything it saw was kept.
control: Something that actually prevents or catches a failure. A name in a document is not one.
ownership: Somebody whose job includes noticing when a thing changes. Not a name on a page.
sufficiency: The point past which recording more buys nothing anybody can reconstruct.
theater: Work that looks like a safeguard and produces nothing anybody can use afterward.

Taken as already known, and so not defined here: agent, model, prompt, approval, incident. That list is a claim about who is reading, and it is printed so it can be argued with.