A versioned framework for deciding whether today's cheapest GPU listing is historically stable enough for a budget assumption.
No supported range is published. The July 15, 2026 read-only preflight found no rights-cleared observations, only 41 eligible live H100 days, and at most two providers on any observed day. Missing history is never fabricated or interpolated.
Every report is keyed by canonical GPU variant and billing mode. On-demand, spot, interruptible, reservation/capacity block, serverless, cluster, and unknown observations are never compared across modes. Ambiguous variants and unknown billing remain excluded until reviewed.
All windows also require at least three independent providers on 80% of observed days, consistent GPU and billing classifications, and cleared rights and provenance. Archive observations are excluded from public calculations and cannot satisfy continuity.
The engine reduces eligible observations to one provider median per UTC day, then calculates continuous p25, p50, and p75 bands. It also records calendar coverage, explicit missing intervals, provider entry and exit effects, the p10 daily-floor reference, observed-day floor persistence, and verified revisions. Failed gates publish only insufficient_evidence with null supported bounds.
The p25–p75 interval is a historically supported published-rate range, not a forecast, guaranteed budget, transaction price, availability claim, performance benchmark, or total workload-cost estimate.
Download the JSON Schema contract for versioned analysis records. It defines the output shape only; it is not a data export or evidence of a supported range.
Collection and public redistribution are deny-by-default and field-scoped. Wayback-derived observations are excluded from public output. Source rows and earlier versions are retained; verified corrections append a superseding version rather than rewriting history.