Methodology · ComputeLlama

Evidence first

Every imported price or compatibility assertion points to an immutable snapshot, its retrieval metadata, and a SHA-256 digest. The original source remains the authority.

Claims are not observations

Provider documentation creates claims. Independent probes create observations. ComputeLlama never presents a published compatibility statement as if we verified it ourselves.

Freshness is explicit

Each source declares a freshness window. Prices remain in history after that window expires, but default discovery excludes stale evidence.

Normalized, not simplified

Token prices are normalized to exact decimal USD components. Conditional eligibility, tiers, qualifiers, and incomplete workload coverage remain visible instead of being silently flattened.

Cheapest for a workload

The default rank estimates 1 million input plus 1 million output tokens. Change the workload to change the decision; incomplete estimates sort after complete ones.

Weather is slice-specific

Independent measurements compare one exact endpoint route inside one methodology, scenario, region, concurrency, and six-hour window. ComputeLlama does not turn these service observations into universal model or provider rankings.

Small samples remain insufficient

Every aggregate reports completed, eligible, successful, failed, and documented infrastructure-excluded runs. P50 is withheld below three numeric samples and p95 below twenty; missing percentiles are never estimated.

Aggregates are reproducible

Every aggregate revision is append-only and carries the methodology version, code commit, time window, metric dispersion, freshness, and a SHA-256 digest of its contributing runs and observations.

Anomalies need a baseline

Automatic alerts require a complete current window and three complete prior windows for the same exact slice. Incident annotations and statistical alerts remain separate from provider claims and raw probe facts.

Video facts are source-run current

A video route is visible only when a capability, release/availability, or native price fact belongs to that source’s latest successful or partial ingestion run. Historical revisions stay inspectable, but an orphan identity or superseded adapter fact cannot look current.

Video formulas stay native

Video comparison evaluates the complete bounded current candidate set with exact decimal source formulas, tiers, constraints, minimum charges, availability, and ComputeLlama callability separately. It ranks totals only within each native currency or credit group, exposes that group’s cheapest eligible route, and never turns credits or other currencies into a fake USD price.

Price history is append-only

Raw price offers, components, and evidence snapshots are never deleted or deduplicated. The public revision view groups only consecutive identical normalized offer terms for the same source offer, retaining first and last observation times, counts, and evidence lineage. A price that returns later is a new revision.

Reasoning effort is not a price guess

An effort filter is available only when the authoritative source payload explicitly publishes supported values. An exact effort-qualified rate wins; a shared source rate is labeled shared. ComputeLlama never invents an output-token multiplier or turns a generic reasoning capability into a supported effort list.

Latest means a concrete pin

A model family resolves only to a source-backed stable concrete release and its exact observed route. A provider’s moving latest alias is disclosed separately, never substituted for that reproducible pin.

Unknown ordering stays unresolved

We do not lexically sort arbitrary vendor IDs. Vendor release dates, catalog listing times, source freshness, release channels, and route variants remain distinct. OpenRouter’s created timestamp is catalog listed-at, not a vendor release date; mixed or conflicting evidence fails closed.