A model endpoint is not a model. It is a serving configuration: precision, cache, caps, the pinned revision. Precision is a trade — the light end buys throughput, capacity and price; the heavy end buys fidelity. Nine fields describe it. Not one of them must be published today.
Low precision is a product. An undisclosed one is the injury.
The lab evaluates the weights. The operator chooses what is served.
Between the evaluation and the buyer's token sits a stack of decisions: weight precision, KV-cache format, context and output-token caps, the reasoning cap, tool-call handling, sampling defaults, speculative decoding, the pinned revision.
Those decisions are a trade, not a fall. Moving down the precision scale buys throughput, capacity and price; it costs fidelity on specific tasks. Both ends are legitimate products, and a buyer is entitled to pick either one.
Nothing obliges anyone to say which one answered. A request lands on an alias. The alias fans out to workers. The workers differ. The invoice does not.
The lab's benchmark is a claim about weights. The endpoint is a claim about service. SIGHT governs the second one, and nothing else.
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. A strong model quietly cut down is not.
Sec. 6: the federal government writes the disclosed terms into 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.
Same open weights.
Different endpoints.
Different trades.
Sources: Artificial Analysis, 2026; vLLM, Red Hat/IST, EMNLP 2025
Accuracy spread across providers of one identical open-weights model, measured against the reference deployment. The 101% top is a confidence band, not accuracy gained.
Price spread across providers of a second open-weights model (GLM-5.2). The most expensive endpoint scored 75% of reference — a dated snapshot, not a standing fact.
Output throughput bought by FP8 at near-zero fidelity cost — the one efficiency claim two independent harnesses agree on. Above 2× the gain belongs to a kernel or a GPU generation, not the bit-width.
What the most restrictive endpoints score on hard reasoning. An output-token limit cuts the model off mid-thought — a cap, not a format.
Points lost on long-context retrieval at 4-bit (Llama-3.1-70B on OneRuler, 92→33). Under the same method, Qwen-2.5-72B gains 0.6% — the default 4-bit path is model-roulette.
Of 1,457 endpoints publish no precision at all (router scrape, 24 Oct 2026). Not hard to find — absent on more than half the market.
US statutes requiring an endpoint provider to say what precision it serves. Zero.
Marginal cost of compliance. Every field is a fact the operator's own stack already writes at launch.
It prohibits nothing.
It mandates no precision. It licenses no one. Publication is the whole duty. Enforcement rides existing FTC Act deception authority. It is a receipt, not a rule.
The President signs the executive order. “SI” replaces “AI” in executive usage.
Data safeguards, accuracy and NIST flow-downs bind covered federal purchases.
No draft has been published in the window. SIGHT is that draft.
A hidden composition, published per serving, in a fixed box.
Estimated cost on a yellow card, so the buyer can compare without a lab.
A 1–10 Greenhouse Gas Rating beside the miles per gallon.
One letter, earned against a reference, audited by someone else.
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.
Nobel Prize in Chemistry — protein-structure prediction (Hassabis, Jumper) and computational protein design (Baker).
Protein and protein-complex predictions in the AlphaFold Database, free to every lab on earth.
More basic research on proteins that previously had no structural information (NBER w35143).
AI-designed antibodies reaching the clinic, against an industry norm near 5.5 years.
Average accuracy of classifier models proposing PET-degrading enzymes — plastic, broken down by design.
Expression gain from ProteinMPNN-guided urethanase redesign — a polyurethane-degrading candidate.
The kind of figure drift a silent precision change can put inside a fluent, formatted, confident answer.
Every one of those chains depends on the served model being what it claims. Disclosure is not a brake on SI. It is the grease.