SIGHT.
Superintelligence Information, Granularity & Transparency

Say what you serve.

SIGHT: the nutrition label for SI.

The same model answers 7.6% better at 3am than at 3pm — not because the model changed, but because the settings around it did. When the service gets busy, providers change those settings to trade quality for capacity to keep up. Nine fields would describe that setup, and none of them has to be published today.

SIGHT endpoint label

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.

Superintelligence is the greatest technological advance in the history of humanity, and that disclosure is a pro-SI instrument: it makes the served artifact trustworthy at the point of use, so a buyer receives the thing they paid for. Nothing here sets a precision floor or licenses a provider.

0
Statutes requiring an endpoint to say its precision1
1.4–1.5×
Throughput bought by FP8, at near-zero fidelity cost3
59
Points lost on long-context retrieval, 4-bit, worst case4
58%
Endpoints publishing no precision at all, of 1,457 surveyed5

A model endpoint is not a model. It is a configuration.

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, sampling defaults, 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.

Three duties. That is the whole statute.

Duty 01

Disclose

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.

Duty 02

Grade

An A–F label of fidelity to the reference deployment. Computed, not judged: the metric and the band cut-points are printed in the statute, and the grade is the measurer’s — the provider prints it, attributes it, and never issues it.

Not raw capability. A weak model served faithfully is an A. A strong model quietly cut down is not.

Duty 03

Contract

Sec. 6: the federal government writes the disclosed terms into its agreements as warranted terms — notice before a material change, and termination for cause, at no termination liability, if one comes.

States and commercial buyers are invited to the same clause.

The trade is already measured.

Same open weights.
Different endpoints.
Different trades.

70–101%

Accuracy spread across providers of gpt-oss-120b, measured against the reference deployment (index v1.0, 4 Aug 2026). The 101% top is a confidence band, not accuracy gained.2

6.8×

Vendor price spread for GLM-5.2: $0.37 to $2.52 per 1M tokens across providers. The most expensive endpoint scored 75% of reference — a dated snapshot, not a standing fact.2

1.4–1.5×

Output throughput bought by FP8 at near-zero fidelity cost, measured on vLLM (Red Hat/IST, EMNLP 2025), batched serving at 8B–405B. Above 2× the gain belongs to a kernel or a GPU generation, not the bit-width.3

<½

What the most restrictive endpoints score on hard reasoning. An output-token limit cuts the model off mid-thought — a cap, not a format.2

59

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.4

58%

Of 1,457 routers surveyed on 24 Oct 2026, 58% publish no precision at all.5

0

US statutes requiring an endpoint provider to say what precision it serves — a sweep of 8 Oct 2026. Zero.1

nil

New fields to collect. Every field is a value the stack already holds — a few engineering days to publish the first time, a few hours per release after; an estimate, not an audit, and no provider’s publishing cost has been measured.6

The negation

It bans no configuration.

No Filing
No Fee
No Veto
No New agency
No Rulemaking

It mandates no precision. It licenses no one. Publication is the whole duty — a provider that says nothing is the violation; a provider that says something false was already reachable under the FTC Act. It is a receipt, with a date, and a price for a false one.

This is not a new instrument.

Since 1994

Nutrition Facts

A hidden composition, published per serving, in a fixed box.9

FTC

EnergyGuide

Estimated cost on a yellow card, so the buyer can compare without a lab.10

EPA / DOT

Fuel Economy label

A 1–10 Greenhouse Gas Rating beside the miles per gallon.11

European Union

A–G energy label

One letter, earned against a reference, audited by someone else.12

Discovery runs on the served model.

2024

Nobel Prize in Chemistry — protein-structure prediction (Hassabis, Jumper) and computational protein design (Baker).13

260M+

Protein and protein-complex predictions in the AlphaFold Database, free to every lab on earth.14

15–40%

More basic research on proteins that previously had no structural information.15

~2 yr

AI-designed antibodies reaching the clinic, against an industry norm near 5.5 years.16

89%

Average accuracy of classifier models proposing PET-degrading enzymes — plastic, broken down by design.17

22.6%

Expression gain from ProteinMPNN-guided urethanase redesign — a polyurethane-degrading candidate.17

41.7→47.1

The kind of figure drift a silent precision change can put inside a fluent, formatted, confident answer.18

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.

A foundation on the obligated side.

SIGHT is a proposal from One Frontier Foundation, a Delaware nonprofit corporation. It publishes its own captured record for both — so the regime it argues for is one it already meets.

That is the point. A rule that costs its author nothing is easy to propose; this one costs the foundation a disclosed record of everything it sells, published before anyone obliged it to. It never grades its own lanes — the record is captured from the serving stack and printed as the label prints, so the first SIGHT record is its own evidence, not its own award.

The rule the proposal is held to: a public claim must match the report behind it — the receipts print the limit beside each one.

Claims current as of 17 November 2026

The SIGHT Act · Sec. 6 · The record is the remedy

Say what you serve.

Sources

  1. 1Statute search — US Code and the bills of the 119th Congress, seven terms, swept 8 Oct 2026: no enacted statute requires serving-precision disclosure. congress.gov’s own search door returned 403 at capture, so the record came from web-grounded search of congress.gov and govinfo. own run
  2. 2Artificial Analysis, Endpoint Accuracy Index v1.0 — 4 Aug 2026, artificialanalysis.ai: the 70–101% spread of one open-weights model, the 6.8× price band whose dearest endpoint scored 75%, the “half the reference or less” cap. A private, point-in-time index over open-weights models; it reaches no closed model. measured index
  3. 3FP8 serving measurements — vLLM, Red Hat / IST, EMNLP 2025: 1.4–1.5× output throughput against BF16 at near-zero fidelity cost. Above 2× the gain belongs to a kernel or a GPU generation, not the bit-width. measured
  4. 4Long-context evaluation of quantized LLMs — arXiv 2505.20276: 4-bit loses up to 59 points on OneRuler (Llama-3.1-70B, 92→33); Qwen-2.5-72B gains 0.6% on the same recipe. preprint
  5. 5Router scrape — 1,457 endpoints, 24 Oct 2026, 58% publishing no precision. Our own capture of routers, not of first-party APIs. own capture
  6. 6The nine serving fields — the draft, v0.2, Sec. 4. The zero-collection claim is an argument about them: each is a value the stack already holds; no one has audited a provider’s publishing cost. argument
  7. 7Executive order, “Inaugurating the Era of Super Intelligence” — 29 Sep 2026, whitehouse.gov: the SI rename (Sec. 2(a)) and the 60-day directive for legislative language (Sec. 3(b)) — the window closing about 28 Nov 2026. primary document
  8. 8GSAR 552.239-7001 — 48 CFR, proposed 17 Jun 2026 (91 FR 36559): data safeguards, accuracy and NIST flow-downs on covered purchases, and never the serving precision. GSAM Change 200 (19 Oct 2026) still lists Parts 539 and 552.239 as reserved, so the clause is a proposal, not a rule. proposed rule
  9. 9FDA, Nutrition Facts — standardized 1994 under the NLEA; the 2016 redesign made calories and serving size prominent. fda.gov statute
  10. 10EnergyGuide — the FTC attaches the label, the DOE takes the measurement the label prints: two agencies, one card. consumer.ftc.gov regulation
  11. 11EPA / DOT fuel economy label — the 1–10 Greenhouse Gas Rating beside the miles per gallon: 40 CFR 600 Subpart D and 49 CFR 575.401. epa.gov/fueleconomy regulation
  12. 12EU energy label — the rescaled A–G bands in force from March 2021, with a QR link into EPREL. energy-efficient-products.ec.europa.eu regulation
  13. 13Nobel Prize in Chemistry 2024 — Hassabis and Jumper for protein-structure prediction, Baker for computational protein design. nobelprize.org award
  14. 14AlphaFold Database — more than 260 million protein and protein-complex predictions, September 2026. the September count measured
  15. 15NBER w35143 — 15–40% more basic research on proteins that previously lacked structural information; no statistically significant rise in early-stage drug development on those targets. nber.org/papers/w35143 measured
  16. 16AI-designed antibodies — ABS-101 reached the clinic after about two years against an industry norm near 5.5; no AI-designed antibody has been approved. pharmaceutical-journal.com reported
  17. 17Designed enzymes — PET-degrading classifiers at 89% average accuracy (bioRxiv 2025); 22.6% expression gain from a ProteinMPNN-guided urethanase (Vanderbilt). Candidates, not products. preprint