The SIGHT Act · America's serving-fidelity standard
Inference is the country's newest export and its least legible one. Every arc on this globe is a request leaving an American endpoint — answered at a precision, a context cap, a revision the buyer never sees. SIGHT asks for one page of paperwork: nine fields, a graded label, a contract term.
The instrument · Superintelligence Information, Granularity & Transparency
A model endpoint is not a model. It is a serving configuration of a model — weight precision, KV-cache format, context and output-token caps, reasoning cap, tool-call handling, sampling defaults, speculative decoding, the pinned revision. The lab evaluated the weights. The endpoint operator chooses what is actually served to the buyer. Nothing obliges anyone to say which configuration answered. SIGHT is the receipt the SI era is missing: it mandates no precision, licenses no one, and creates no new agency.
Publish a machine-readable record of the serving configuration of every covered endpoint: nine fields, quantization first, with a dated diff on every material change. Where one alias fans out to differently configured workers, a record per worker and the routing rule.
An A–F label of fidelity to the reference deployment, measured by others, never self-issued. Not raw capability: a weak model served faithfully is an A.
Sec. 6: the federal government puts the disclosed terms in its agreements as warranted terms, with notice and a right to terminate on material change. States and commercial buyers are invited to the same clause.
The President signs the executive order Inaugurating the Era of Super Intelligence, directing the Assistant to the President for Science and Technology to submit proposed Federal legislative language within 60 days — about 28 November 2026.
GSAR 552.239-7001 takes effect, binding data safeguards, accuracy and NIST flow-downs on covered federal purchases. It never once asks the serving precision.
No draft has been published in that window. The government buys inference by the token and cannot read the configuration that produced it.
Grey & Fast published the first record: both lanes it sells, every value captured from the serving stack itself. One lane's alias answers two weight formats round-robin — NVFP4 on a 2-GPU worker, FP8 on a 4-GPU worker — and the buyer cannot choose which answers. That is not misconduct. It is what every endpoint does, unwritten, until someone writes it down.
Where a product's hidden configuration imposes a cost the buyer cannot see, the state mandates a comparable, graded disclosure. Inference is now that product.
Disclosure is not a brake on SI. It is the grease. The Nobel-winning protein work, the 260 million structures, the antibodies reaching the clinic in two years against an industry norm of five and a half — every chain depends on the served model being what it claims to be.
Pro-government. Pro-industry. Pro-human.