Benchmark methodology preparation
A benchmark is a versioned measurement method, not a screen price. This guide defines the evidence to prepare before a Megatron Markets benchmark is published or supplied to an external product partner.
Measurement scope
Freeze the economic identity before selecting observations:
- model and provider endpoint;
- service tier and any declared performance threshold;
- exact Prompt, Cache read, Cache write, and Output bundle weights;
- spot or forward term, including tenor and delivery window where applicable;
- quote unit, quantity unit, currency, and observation time.
Changing any one of those fields changes the measured product. A generic GPU-hour series and a model-token bundle therefore remain separate measurements unless a published method defines and validates a conversion.
Observation classes
Keep each source class distinct:
| Observation class | Megatron Markets source | What it demonstrates |
|---|---|---|
| Provider endpoint rates | GET /models/{slug}/price-history | Published Prompt and Output rates plus available endpoint-quality observations |
| Venue snapshot | GET /market/prices and symbol books | Current quoted interest at a stated time |
| Venue executions | GET /instruments/{symbol}/trades | Executed prices and quantities for one instrument |
| Term structure | GET /curve | Available spot and forward points for one model and leg |
| External reference product | Provider-owned data access | A separately specified third-party market |
Provider rates are not venue executions. Quotes are not trades. A third-party contract is not a Megatron Markets bundle. Do not combine those classes without an explicit inclusion rule and source label.
Methodology checklist
A candidate method should state, in advance:
- The eligible product universe and the reason each constituent belongs.
- Data rights, contributor identity, timestamps, and a maximum observation age.
- Whether the input is a provider rate, bid, ask, midpoint, or executed trade.
- The observation window, time zone, minimum contributor count, and concentration limits.
- Missing-leg, stale-data, outlier, provider-outage, and model-retirement rules.
- Exact bundle arithmetic, currency and unit conversions, precision, and rounding.
- A deterministic fallback that can return “no value” rather than inventing a price.
- A versioned correction procedure with an effective time and retained prior value.
The four bundle weights must be non-negative, at least two must be positive, and their exact decimal sum must equal one. See Formulas for the bundle calculation.
Publication evidence
Prepare one reproducible package for every published value:
- methodology version and effective time;
- eligible universe and observation count by model and provider;
- included and excluded observations with reason codes;
- source timestamps, input checksum, calculation output, and rounding step;
- contributor-concentration and stale-data results;
- original and corrected values when a correction occurs;
- named owner and publication time.
This is a readiness checklist. Completing it does not turn catalogue data, a market snapshot, or a demand signal into a Megatron Markets benchmark. The external exchange reference guide covers comparison and integration boundaries.
Publication API
GET /api/v1/benchmarks returns only methodology identities marked PUBLIC that also carry an append-only, attributed methodology promotion. Before an identity is returned, its definition is reconstructed through the typed methodology rules and its code, version, effective time, and checksum must match the stored identity. Each item carries code, positive version, definition_checksum, and effective_from. An empty items page is valid and means there is no public methodology to display at that request time.
GET /api/v1/benchmarks/{code}/publications returns only artifacts that belong to a public methodology and have an explicit public promotion. Each item contains the checksum and its checksum-verified canonical_payload. A valid public methodology can have an empty publication page; an unknown or non-public code returns 404.
Both benchmark endpoints accept limit=1 only—the default and maximum are both one—because an individual methodology definition or publication artifact can be as large as 16 MiB. Requesting another limit returns 400. Consumers should retain the checksum with the canonical payload and verify it independently before using the artifact.