Twenty-one verifiers read every draft with fresh context. One caught a promise that existed nowhere.
IN PRODUCTION
This runs in client systems today. The sample below was regenerated for a fictional company.
A fleet of small verifiers, each looking for one class of problem, reading with no memory of the conversation that produced the draft. Over seven drafts they raised about forty findings. The one that matters is a service-level promise stated in a document that appeared in no source, and which a person had already approved.
TRY IT / RUNS IN YOUR BROWSER
Use it here.
The mechanism is the production one. The company, the catalogue, and every figure are fictional.
Loading the run.
If this message stays on the page, the verifier data at /assets/systems/qa-fleet/data/ did not load.
THE OUTPUT / GENERATED FOR APEX INSTRUMENTS
What comes out.
JSON
The fleet, 21 verifiersData · Each carries an id, a name, a family, exactly what it looks for, the severity it raises, what it reads, what it cannot see, the document kinds it runs on, and whether it needs source retrieval. 52 declared blind spots in total, which is the field that makes the rest of the entry believable. Applicability is what produces the 17 not applicable cells in the grid.Open the file ↗
JSON
Seven drafts and the thirteen sources behind themData · A partner letter, a price sheet, a service levels sheet, a quick start guide, a proposal, a launch note and a product one pager. Each carries its kind, audience, author role, length in words and pages, the sources it was built from, and the human review it had already passed. The thirteen sources carry an owner, a review date and a review interval, which is what lets one verifier report a document that is faithful to something nobody has checked.Open the file ↗
JSON
The run, 42 findingsData · 10 blockers, 24 major, 8 minor. Every finding names the draft, the verifier, the location, the exact quoted text at fault, and either the source line it contradicts or the list of sources that were searched and did not contain it, plus a resolution with an owner and a date. One is marked spotlight: the invented four hour response promise, with the five sources searched, the closest thing any source does say, a five step walkthrough of how it was caught, and five reasons the human review missed it. Two are marked withdrawn and are left on the run rather than deleted.Open the file ↗
JSON
The two tiers of the pre-delivery scannerData · The draft tier runs 9 verifiers in about 41 seconds on every save. The final tier runs all 21 in about 6 minutes before delivery. Each tier states why it is shaped that way, what it catches, what it misses, and the honest part. The stored expectations are checked against a recomputation: the draft tier catches 18 of the 42 findings and none of the 10 blockers, and it does not run the verifier that caught the invented promise.Open the file ↗
JSON
Twelve findings put through independent reproductionData · Every blocker plus two majors. Each entry records what the verifier claimed, what the independent reader was asked to do, what that reader found, and whether the finding survived. Ten survived. Two did not: one where the verifier compared a partner price against the list column, and one where the claim was true and sat in an appendix the source index had never pulled. Both name who was wrong and what changed as a result.Open the file ↗
Each verifier starts without the conversation that produced the draft.
02Look for one thing
A verifier that checks everything checks nothing well.
03Say what it cannot see
Every verifier declares its own blind spot.
04Tier the run
A draft tier is fast and deliberately weaker. The final tier is not.
05Reproduce before trusting
A finding is not acted on until someone reaches it independently from the source.
WHAT IT READS
The drafts and the sources they were built from
One verifier per class of problem
A tier definition for draft and final
An independent reproduction step
WHAT IT CHECKS BEFORE IT SHIPS
Every finding quotes the text at fault
A finding with no source contradiction is still reported as unsourced
Two findings failed reproduction and were dropped
No verifier claims to see what it cannot
In production: 21 fresh-context verifiers over 7 drafts, which caught an invented service-level promise that had already passed a human review.
The mechanism runs inside client systems. Everything on this page was regenerated for Apex Instruments, a company that does not exist. Client names, source files, and internal wording stay private.
