Nobody Opens a Passing Check

The finding

Fourteen green lights, and the question stayed unasked.

Across fifteen cases in one small, carefully watched operation, fourteen controls supplied a green or pass-like signal while the question that mattered sat outside what they enforced. In the fifteenth, a merely available capability was treated as coverage even though it was not invoked correctly during the documented nine-day period.

Each case is written up with what produced the assurance, what lay outside it, what got through and what changed. The four-class taxonomy is designed to help you test analogous failures in your own setup; its performance in another operating environment was not measured.

Where this starts

A book about the middle of the job.

Not choosing a framework, not wiring the first agent — that material is abundant, current and free, and reprinting it at length would be charging you for it. This is about day twenty-two, when the loop has been running for three weeks, the apparatus is supplying green or pass-like assurances, the weekly report says red, and you have to decide whether it is telling you the truth.

The operation, in numbers

What was actually measured.

• The action log covered 39 calendar dates, with activity on 38.
• A separate 40-calendar-day git population ending at the harvest held 236 commits, 16 of them carrying the founder commit prefix.
• By the harvest the venture had published 81 posts and logged 55 decisions, 20 of which carry a machine-readable record of who approved them.
• It produced five reports across six scheduled reporting slots — W4 has no report, only a row filled in afterwards from the ledger.
• It earned nothing, against a cumulative profit of −13.00 EUR.

Those sixteen founder commits occurred outside the scheduled daily loop; the operation as a whole was not unattended. Both halves of that are the subject.

What's inside

Fifteen chapters, every number dated.

0. How to read this book. The two voices, and what this was measured on and what it was not.
1. When two agents, and when one is enough. With a decision tree that has a leaf saying one agent.
2. The charter. The autonomy grid, the three kinds of switch, and a template with no pre-written legal text.
3. The operating loop. Eight steps in the order that makes them work, and the outage that cost a weekly report nobody missed for seven days.
4. State that survives. What crosses the end of a session, and why a prose log cannot be audited.
5. Two agents, not one. The separation of author and approver, and the four times it broke.
6. The gate is not the answer. The fifteen cases, in four classes, plus a seven-question self-audit.
7. The second model. A read-only reviewer with a daily cap: twelve blockers, seven of them inside its own guard.
8. Gates, caps and stop conditions. Three constructions that fail differently, and the defect that only exists across a series.
9. Tools at the boundary. Three permission gates a human owns, and a contract clause nobody read before building.
10. Monitoring and steering. A queue that grew to three dense pages a day against a reading budget of twenty to thirty minutes a week.
11. The record. Six scheduled reporting slots within thirty-nine days: a two-day W1, five seven-day slots, five produced reports — and for W4 no report, only a row filled in afterwards from the ledger.
12. Four times we were wrong about ourselves. Four documented self-refutations, and the rule the team applied each time.
13. What a red light is worth. Five produced reports printing red under a definition that never moved; W4 has no report.
14. The next experiment. What is open on day thirty-nine, and the criterion by which this book's own launch counts as a failure.

Free companion

Try the method on one passing check.

How to Audit a Passing Check in an AI Agent Workflow takes a workbook example from its requirement to a finding. Use the completed example and editable worksheet to investigate one result in your own workflow, without buying the book.

Before you buy

An honest promise.

If you already run a two-agent team, this book gives you a tested daily-loop method, checklists and fixtures for examining whether its controls held or merely looked as if they did. Transfer to a different runner, scheduler or operating environment was not measured. It will not make you money on its own. Ours has not.

The internal release verification checks seventy-seven selected figures: sixty-two are recomputed from shipped inputs, and fifteen are transcription comparisons between printed book values and shipped records. It exits non-zero if one of those comparisons deviates. It is evidence about this edition, not a delivered repository or tool promise, and it does not verify every number in the book — every run names five groups outside its reach.

If a figure in the book and the operation's records disagree, the records are right. Where a figure has moved since it was read, the book is stale, and that is why every number carries its date.

What this looks like in practice, with sources: what the pipeline rejected →

For whom — and for whom not

Is this field report right for you?

Everything included

What you download.

• 193-page tagged PDF with bookmarks, page numbers and a running disclaimer
• Offline HTML bundle of all fifteen chapters — no tracking, no external dependencies
• Three calculators that run in your browser, send nothing and store nothing: the autonomy-matrix builder, the self-approval and silent-pass audit, and the cycle-cost estimator
• Exactly seven editable example files: CHARTER.md, autonomy_matrix.xlsx, cycle_prompt_pack.md, DECISIONS_format.md, weekly_report_template.md, field_study_numbers.csv and field_study_numbers.xlsx
• The two number twins are granted under the MIT licence — copy, edit, redistribute and sell work built on them