Researchers tested six verification layers around an open-weight language model and found that accuracy gains came only from withholding answers, not correcting them. On clinical-style forum posts, the base extractor answered correctly in about one case in three while reporting full confidence on two-thirds of answers. The layers included deterministic checks, retries, majority voting, a second-model judge, and human review. The result matters for business because stacking checks increases review workload without producing more correct outputs.
How six layers were tested on CADEC records
The test used CADEC, a corpus of 1,250 forum posts about two drugs with every adverse reaction marked as a text span and SNOMED CT code. The model never generated codes directly but quoted each reaction, after which a retriever searched 227,554 keyword-to-code rows from the SNOMED CT Australian edition of 31 July 2026. Each quote was embedded with granite-embedding:30m into 384 dimensions, and the model selected from 20 nearest concepts presented as a menu. Codes were recorded in bar notation such as 271782001 |Drowsy|. The team did not build a tuned CADEC system and used the corpus only as a graded instrument.
The pipeline was fixed before layer comparison, with three alternatives measured on 40 development documents over three cold runs. Recalling SNOMED codes from memory proved broken, with up to 38 percent null answers and 13 to 18 percent of committed codes absent from any SNOMED release. Naming the concept for lookup performed better, while picking from the retrieved menu performed best. The chosen extractor used gpt-oss:20b, which delivered 6.5 extra F1 points over llama3.1:8b. Four other model families produced byte-identical output across runs, while only the mixture-of-experts model varied between runs.
Run-to-run variation set the noise floor for all later claims, with three identical runs at temperature zero differing by four F1 points. Two runs were byte-identical, while the third found 98 correct mentions against 87 for the pair, from 233 mentions grouped by span. Where all three runs found the same span, code agreement reached 84 percent, which suggests instability in span detection rather than coding. The authors kept the varying model for its higher score and reported every later change as a range across three runs. A supervised CONORM system fine-tuned on 875 of 1,250 files reached 0.72 F1 under lenient matching, against 0.47 to 0.50 for the zero-shot extractor.
What this means for business use of LLM checks
For companies processing tickets, reports, or clinical notes, the finding changes the economics of quality control rather than model selection. The free vocabulary check endorsed only 32 percent of mentions with wording matching the vocabulary, removing 128 errors at the cost of 49 correct answers and 177 records sent for human review. On the held-out set of 60 documents, five in six answers marked ACCEPT were correct, against one in five among the rest. That sorting is useful for triage, but a small team receives a large manual queue while a large operation pays for review capacity without expanding throughput.
Paid layers added cost without repair, which affects vendor evaluation and budgeting. Self-correction almost never triggered, majority voting across samples did not improve results, and the second-model judge produced useful verdicts that no downstream step acted upon. Removing the three paid layers changed refusal decisions by one record on one run and zero on two others. The menu-shown judge rule reached F1 of 0.420 against 0.397 for shipping everything, the only rule to beat the baseline. Buyers should therefore ask what a proposed layer changes in the output, not only what it flags.
The marker to watch is whether vendors report yield alongside accuracy, meaning correct answers as a share of all records. If future reliability products show higher precision without disclosing withheld correct answers and human-review volume, the pattern from this test is repeating. A genuine advance would increase correct shipped answers on a fixed corpus such as CADEC while holding review load steady. Until that metric moves, layered verification remains a sorting mechanism.
