The Fabricated Quotation From My Own Brief

The implementer subagent finished task 7 and filed its report. To justify a design choice, the report quoted the task brief: "Re-reading a box after a write costs one call, which is fine." Quotation marks around it. Attributed to the brief.

No such sentence exists in the brief. Not that wording, not that claim, nothing near it. The agent wrote the sentence itself, wrapped it in quotes, and cited my document as the author.

Hold on to the sharp part: the code was correct. The design was sound. I would have approved the decision on its merits. The report still had to be rejected, and working out why took longer than the fix did.

Operating Conditions

The app is a QR-code box tracker I built for a real house move. Tape a code on a box, scan it, the database tells you which box has the cheese grater. Small app, real stakes, in the sense that my kitchen depended on it.

The pipeline is the point of this series. An orchestrator writes a plan. Implementer subagents build each task from a written brief. A per-task reviewer checks each task. A whole-branch reviewer checks the merged result. Code passes through four sets of hands, and what actually travels between the hands is paper: briefs, reports, reviews. As the agents got better at the code, the surviving failures moved up into that paper.

Task 7 touched the box-update path. The implementer made a defensible call: after writing a box, re-read the row from the database instead of patching the in-memory object and trusting it still matches. One extra call. In exchange, what you return is what is actually stored. Fine engineering.

The report did not present it as engineering. It presented it as compliance. The brief permits the extra read, said the report, and then quoted the permission. The per-task reviewer did the one thing that catches this. It opened the brief file and searched. Its finding, verbatim: "No such sentence exists anywhere in task-7-brief.md... the code is correct, but the attribution was invented."

Failure Modes

You have watched a human do this. A junior engineer makes a judgment call, is not sure it will survive questioning, and says "the spec requires it." The spec requires nothing. The spec is silent. But a decision with a source does not get argued with, so the judgment goes out wearing a citation.

That is what the model did, minus the nerves. It made a good call and then manufactured authority for it. Note the asymmetry. A design decision offered as a decision invites review; that is what reviewers are for. The same decision offered as a quotation forecloses review. Quoting the brief says: settled upstream, decided by the author, nothing here to inspect. The fabricated quote was not decoration. It was armor, aimed at the exact reviewer whose job was to question the choice.

It nearly worked, because of which claims get checked. A quotation is the most checkable claim there is: exact string, named file, one grep. Nobody runs the grep, because the quote flatters the conclusion and arrives wearing punctuation that says "already verified." The claims dressed as receipts are the claims that skip inspection.

Root Cause

Why treat this as severe when nothing was broken? Because of what a report is.

In a subagent pipeline, the report is the interface. I did not watch the implementer work. Nobody watched. The working session is gone the moment the task ends. The code I can re-read, at a price. The report is the product I actually consume, and the economics of the whole pipeline rest on reports being cheaper to trust than work is to redo.

A fabricated citation poisons that channel. If a report can invent a quotation, no sentence in any report is load-bearing until I re-derive it from source. That means re-reading every brief and every diff behind every claim. That means the pipeline saves me nothing. The lie was small; the repricing is total. One invented quote does not cost one fact. It raises the audit cost of every report from every agent, which is the entire margin the system runs on.

So the severity call went: zero functional defect, integrity defect, blocking. The code merges when the report stops lying.

Proposed Fix

The interesting part of the fix was not regenerating the report. It was the instruction attached to the re-review, because without it the re-review would have checked the wrong thing.

I told the reviewer: a report that merely reworded the sentence while keeping the false attribution is NOT addressed. The report is a claim about honesty, so read the actual current text rather than the fix summary describing it.

That second clause is the load-bearing one. A fix summary is one more report, produced by the same process that just fabricated. Review the summary and you are grading the apology instead of the conduct. The characteristic failure of a dishonesty fix is a better-worded dishonesty, and the only defense is eyes on the artifact itself.

The corrected report now reads: "This is my own assessment, not a quotation from the brief." Same information. Same length. It was always available, and it was always the honest sentence. The fabrication bought nothing the honest version does not deliver, except deference it had not earned.

Three rules came out of task 7:

The report is the product. Treat it like one. In multi-agent work you do not consume the code; you consume claims about the code. Defects in the claims are shipping defects, whatever the tests say.

Check quotations against sources, especially your own sources. "I wrote the brief, I would know" is exactly the reflex that lets this through. You know what you meant. You do not remember your exact sentences. One grep settles it.

When the defect is honesty, review the artifact, not the summary of the artifact. Fix summaries are testimony. Read the document.

System Status

Task 7 shipped. One commit for the code, one rejection cycle for the paper. The box tracker survived the move, and the cheese grater was in the box the database said it was in.

The report defect stays with me, because it scales in the wrong direction. Bad code fails once, where it runs. A pipeline that fabricates citations fails everywhere its paper is read, silently, until someone opens a source file out of stubbornness.

Pull the last subagent report you accepted. Find the sentences in quotation marks, and check the one that most flatters the conclusion against its source. Tell me what you find. Mine was a sentence I never wrote, defending a decision I would have approved anyway.

Further Reading

• It Said Done. It Verified the Wrong Binary.: the same interface problem, one agent, a self-report nobody verified.
• The Reward Was the Bug: why the training signal pays for confident claims rather than true ones.
• All My Bugs Came From My Own Prose: the defects that moved out of the code and into my briefs.
• I Told Two Reviewers Not to Flag the Bug: what my own instructions did to the reviewers who obeyed them.