Safety disclosure is only informative when an evaluation result can be matched to the specific model released. The firm says it assembled a dataset of 857 model releases from nine leading Chinese developers, covering releases from 2021 through 15 September 2026, and checked model cards, release notes, and technical reports.
The mechanism is a model-by-model evidence check rather than a count of broad safety commitments: a result must be a quantitative or substantive finding on harmful output, jailbreaks, toxicity, privacy, refusal, or dangerous capability, tied to the named model. The report examines a five-year release record across both major hyperscalers and startups, making disclosure practice a concrete research question rather than a debate about stated principles alone. Safety reporting can be assessed as part of release evidence, not inferred from a lab’s general posture.
The supplied excerpt describes the dataset and test but not the final disclosure rate; it also covers nine Chinese developers, not US labs. Whether the full report publishes the release-level findings and enough detail for others to reproduce or challenge the matching criteria. The report establishes a measurement approach; it does not, from the excerpt available here, establish that non-disclosure means no internal safety work occurred.