Methodology change notice — the normalisation ladder: rungs 1–3 published as reference, rung 4 declined
Notice date: 2026-08-20 (UTC)
Class: methodology publication (reference surface). Not a correction — no published value is
restated, and no new printed series is created by this notice.
Instrument: none — the ladder is a reference surface. The existing series it touches
(gap_h100_spot, the token_pricing collection behind RWI) are unchanged in method, class and
print discipline.
Effective: 2026-08-20 — on publication. Publish-before-effect (decision of record
2026-07-24 §2) is satisfied trivially: the ladder governs no print, sets no level, and feeds no
downstream series; it can therefore take effect the day it is published.
1. What is adopted
The Grid publishes the normalisation ladder — the four denominations in which the market quotes the price of compute, each rung stated with its own honesty furniture, as a directional reference surface:
$/MW ──▶ $/GPU-hour ──▶ $/M-token ──▶ $/task
(1) (2) (3) (4)
The ladder is not an assessment and not a set of assessments. No rung creates a
derived_series row, no rung prints a daily value, and no conversion between rungs is published
as a factor with a level. Each arrow between rungs hides a conversion factor (power density and
PUE; throughput for a named model under named conditions; tokens per task, which varies on the
order of 2× between models) that nobody publishes under controlled conditions today — the
surface states each factor's status honestly rather than collapsing the chain into one number.
2. The four rungs and their dispositions
| Rung | Unit | Disposition |
|---|---|---|
| 1 | US$ per MW | Reference / explainer. The Grid collects no capacity-cost-per-megawatt series. The rung page states the conversion arithmetic with cited public inputs and marks the empty rung plainly — an empty rung is stated, never faked. Worked conversions from Grid-collected data are labelled worked conversions, never prints. |
| 2 | US$ per GPU-hour | The existing GAP surface. The rung page links to the GAP-H100 specification (/standards/instruments/gap-h100) and republishes no number: one definition stays one definition. |
| 3 | US$ per million tokens | Reference table rendered from the stored token_pricing rows (OpenRouter + DeepInfra documented public API endpoints; source ratified by founder ruling 2026-08-20; one request per endpoint per daily cycle). Published list prices as recorded — a record of what providers publish, not an assessment: no adjustment, no judgment, no panel. The table carries source, method, observation count and date. |
| 4 | US$ per task | DECLINED. See §3. The declination is published as its own page — declining in public is part of the method. |
3. The rung-4 declination
The Grid declines to publish a $/task figure, for stated reasons:
- Task definitions are not standardised. There is no governed, market-accepted definition of "a task" to price. A price for an undefined deliverable is not a price.
- Task benchmarks are gameable. Where task suites exist, they are optimisation targets; a $/task print built on them would inherit their drift without inheriting any governance.
- A $/task print would be false precision. Tokens-per-task varies on the order of 2× between models on the same work; any single $/task number silently embeds a model choice and presents it as a property of the market.
The rung is declined until the market defines the unit: a standardised, independently governed task definition — submission, peer validation, publication under rules — would reopen the question by a further notice. The declination follows the same reasoning as the existing refusal to score capability (the y-axis saturates; capability is not a ratio scale and cannot be divided by).
4. Benchmark-source compliance
Two third-party benchmark sources sit adjacent to the ladder. Their terms govern what renders:
- MLPerf® (MLCommons). Named on the ladder only as the anchor of record for the GPU-hour → token conversion (audited, controlled-conditions, hardware-specific results). The Grid republishes no MLPerf result figures and makes no performance-per-watt claims (official MLPerf Power results are the only permissible basis for any future efficiency claim, and none is made). Every rendered reference uses the ® mark and carries the MLCommons trademark attribution footnote. Where a citation is in doubt, the ladder cites less.
- Artificial Analysis. Its terms of service bar scraping; its Data API free tier is
licensed for internal use only. No AA-derived value renders on any public surface until a
commercial-tier licence is purchased: every cell a licensed AA feed would fill renders
[gap: licence pending]— the gap is displayed, not hidden. The collection path exists only as an environment-gated stub (scrapers/artificial_analysis.py: requiresAA_API_KEYand an explicitAA_LICENCE_TIER; the publication path refuses to run unless the stated tier is Pro/Commercial). The stub is not registered in the daily pipeline.
No new scraped source enters with this notice (R-SRC-001 applies): OpenRouter and DeepInfra are the existing, ratified collection; MLPerf and Artificial Analysis are cited-or-gated, not ingested.
5. Record
- Decisions of record: founder rulings 2026-08-20 — (a) the ladder adopted as a directional reference surface with rung 4 declined; (b) MLPerf strict messaging compliance and Artificial Analysis publication gated on a paid licence.
- Surfaces:
/normalisation(the ladder),/normalisation/mw,/normalisation/gpu-hour,/normalisation/token,/normalisation/task; masthead navigation entry. - Rendering definitions:
dashboard/lib/standards/ladder.ts— renders this notice's content; the notice, not the file, is the document of record. - Rule pack: unchanged at v1.1.0. The two compliance rules in §4 are queued in
data/governance/QUEUE.mdas candidates for the next founder-ruled version bump. - The
token_pricingcollection, the RWI paths built on it, and every existing instrument specification are unchanged.