The intelligence desk.
Follow the system beyond the specification. Inspect what was measured, who confirmed it, what changed and which terms remain undisclosed.
Datasets & training data ↗
Embodiments, collection methods, modalities, scale and dataset-specific reuse gaps. A metadata listing never grants rights to training artifacts.
26 reviewed recordsStandards & regulations ↗
Public scope, editions, jurisdiction and revision status. A relationship to a standard is not a product certification.
93 reviewed recordsCommercial offers & access ↗
Dated terms, public access statements and customer-specific agreements. Prices, deposits, billing periods and gaps remain separate.
14 reviewed recordsFunding & ownership ↗
Event amounts, commitments, round totals and valuations. Extensions share a round group so cumulative announcements are never added together.
32 reviewed recordsMarket & version history ↗
Actual dated announcements and changes, with source dates and operational dates kept distinct from editorial reviews.
12 reviewed recordsDeployment evidence ↗
Named customer, site, generation, task and operational metrics. Customer reports, lab trials and vendor demonstrations stay distinct.
528 reviewed recordsModel & product versions ↗
Exact release and paper-variant identities. An unspecified checkpoint remains unknown; a newer generation never inherits an older one’s results.
557 reviewed recordsBenchmark result intelligence ↗
Paper-reported results with protocol, original table, upstream revision and comparability limits. No score combines different benchmarks.
Depth is more than a record count.
1,442 existing disclosed claims and 2,926 additional disclosed record fields are tracked separately from 3,366 benchmark score cells. Model-variant metadata is included in record fields; it is not a claim of deeper coverage for every company.
90 access records reorganize existing disclosures. That migration adds no new facts. Imported paper results passed context checks; most original papers have not been manually rechecked, and no experimental reproduction is claimed.
Unknown fields distinguish research pending from absence in the sources already reviewed. A published dataset profile does not authorize downloading or training on the dataset. Sponsor payments cannot alter evidence labels or outcomes.