How the checking ran
The full cycle ran on each file: scoping, terminology, AI translation with automatic checks, separate checking agents, the devil’s advocate, the judge, then the quality gate and the report. The dialect-consistency check ran on the two Arabic files. The checking ran in three rounds.
- Round 1, the cycle. The judge confirmed 1 defect in the Gulf Arabic file, 4 in the MSA file, and 3 in the French file. Each was fixed, and each file passed the quality gate.
- Round 2, a second checking round after delivery. An AI check of the delivered files (the reports call it the delivery review) found wording that the judge had ruled a preference because the finding cited no written reference. We corrected the judge’s rule: a native-reader finding now counts when it names a concrete defect in the target text, even if its only reference is the source segment or the grammar of the language. A finding that only says another wording would be nicer stays a preference. With that rule, the checks, the devil’s advocate, and the judge ran again on the delivered text. They fixed 6 more issues in the Gulf Arabic file, 4 in the MSA file, and 6 in the French file.
- Round 3, a separate re-read. A separate native-reader agent that saw only the target text re-read each file. A separate devil’s advocate challenged each of its findings, and a separate judge ruled. This reader raised 13 findings on the Gulf Arabic file, 9 on the MSA file, and 6 on the French file. One was ruled a defect, in the MSA file, and fixed. The other two files did not change.
Across the three rounds the judge confirmed 25 defects: 7, 9, and 9. All 25 are Minor: points of meaning, wording, grammar, and punctuation. Each one is listed in the projects above, with the check that found it, the devil’s advocate’s challenge, and the judge’s ruling.
What a client job includes: every file goes through the full cycle as in round 1, and the judge now applies the rule we corrected in round 2. Rounds 2 and 3 were extra checking we ran on these demonstrations to test the cycle itself; a client job does not include them unless your agreement adds them.
The reports state the limits of round 2 themselves: the independence of its checks from one another was not assured, and in the two Arabic files they were not blind to the delivery review. Round 3 supplied a separate blind read. The delivery review was itself an AI review, not a human review. Each report says: “No human reviewed or edited the file.”
Checking on all three projects was completed on 11 November 2026, on the cycle as it stood then. Each report carries its own dates.
What these projects show, and what they do not
- They show the cycle at work on content we wrote: what it flagged, how the devil’s advocate challenged each finding, how the judge ruled, and what changed in the file.
- They are not client work. We wrote the source text, the glossaries, and the style guides ourselves. These files say nothing about any client’s content, and we draw no figure from them about yours.
- No linguist read the files. Three rounds of AI checking are not a human review, and the reports do not present them as one. The round-3 judge states its own limits in the reports. In the Gulf Arabic project, some of its rulings “turn partly on whether a Saudi wording sounds natural, which it could not verify against usage evidence.”
- The cycle checked its own output. This is not an outside audit.
- A passed gate covers mechanical points. It does not settle the open questions or the proposed terms. Each report lists them for the client to confirm, and here there is no client to answer them.
- You get counts here, not scores. The page gives no quality score and no timing.
- “Nothing Critical or Major” is what the cycle found. It does not say that nothing else is there.
- The published copies leave out the internal run records. Under your NDA, we tell you which AI services process your content.
Human review is an optional add-on: + Human check, + Human edit, or + Full review, by a native linguist. None was used in these projects.