What happened
QUASA reports that an August 2026 proof of concept used a reserved subset of MLCommons AILuminate safety prompts, OpenMined secure computation and a containerized Google DeepMind model. The arrangement prevented the model developer from seeing the evaluation prompts and prevented the benchmark provider and auditor from seeing proprietary model weights. QUASA says the design reduces some contamination and intellectual-property risks, but does not eliminate earlier exposure to benchmark material or prove that the test has construct validity.
QUASA reports that benchmark contamination can arise when evaluation questions, answers or close variants enter pretraining, fine-tuning data or repeated optimization. The outlet says exact duplication is not necessary: partial answers, distinctive concepts and related examples may make familiar tasks easier.
According to QUASA, the DeepMind–MLCommons pilot used AVERI to run reserved MLCommons AILuminate safety prompts through OpenMined secure computation and a containerized Google DeepMind Gemini Flash Lite model. The model owner could not inspect the reserved questions, while MLCommons and AVERI could not inspect the proprietary weights. These details are reported by QUASA from the cited implementation account and are not independently confirmed here.
The outlet distinguishes execution integrity from construct validity. The first concerns whether the intended unseen items and declared system were used under controlled conditions. The second concerns whether the tasks and metrics support the capability, safety or reliability claim attached to the score.
QUASA identifies prompt leakage, model-weight exposure, evaluator access and future-training contamination as separate risks. It says controls such as access records, retention limits, output restrictions, incident procedures and rules for retiring exposed items must complement cryptographic protections.
Why it matters
AI benchmark scores increasingly influence procurement, safety claims and comparisons between systems. QUASA’s reporting highlights that confidentiality addresses only one part of evaluation integrity: keeping protected prompts and model weights apart during a run. A benchmark can remain unrepresentative, poorly scored or vulnerable to future contamination even when neither side sees the other’s confidential material. That makes unseen-item provenance, configuration disclosure, uncertainty estimates and failure analysis practically important for anyone relying on AI test results.
A confidential evaluation can make it harder for a model developer to optimize directly against reserved questions, while also limiting disclosure of proprietary weights. That is useful evidence about the conditions of a test, but it is narrower than evidence that a model is safe, reliable or generally capable.
QUASA cites a NeurIPS review of 445 LLM benchmarks as finding recurring weaknesses in the phenomena, tasks and scoring metrics used to support benchmark claims. The source does not provide the review’s authors, full methodology or independent assessment of the pilot’s results, so those details remain unverified in this evaluation.
The practical implication is that organizations should treat a double-blind score as conditional evidence. They need the model version, configuration, system prompt, tools, sampling settings, runtime environment, sample size, uncertainty estimates and category-level failures before using the result in a deployment or purchasing decision.
What to watch next
Watch whether future evaluations document how reserved items were protected, how prior exposure was investigated and what happens to prompts, outputs, logs and derived data after testing. Procurement teams should also ask whether the tested model configuration matches the one being deployed and whether the benchmark reflects the specific risks of the intended use. QUASA does not establish through independent testing that the reported pilot improved real-world validity or that its controls prevent every leakage path.
Future reports should clarify who can access prompts, weights, responses, scoring rules, logs and attestation evidence before, during and after a run.
Benchmark maintainers should explain how items are reserved, how prior exposure is investigated, how compromised questions are replaced and how benchmark versions are refreshed.
Organizations should compare the tested configuration with the system actually deployed and check whether the tasks represent their domain, including rare but consequential failures.
QUASA does not report a price, public availability, access process or measured performance improvement for the pilot. It also does not independently verify that the described controls prevent all leakage or downstream training contamination.