Docs
Epistemic boundaries & auditability
What we promise
- Pre-provision ranking on control-plane evidence + workload-adjusted TRAC (modeled comparison, not calibrated expected loss)
- Probabilistic SIR / PIS with visible signal coverage
- Cost composition where fee data exists, with fee lineage labels
- Panel indices from the same observatory tape, with public descriptions of index intent
- Scope: deployment reliability and listing economics, not FLOPS leaderboards
What we don’t promise
- Certainty that an ephemeral listing will survive your specific job
- That PIS or SIR are calibrated interruption probabilities
- That TRAC risk premium equals realized economic loss
- That execution-risk flags mean hardware does not exist
- Complete discovery of undisclosed path-dependent fees
- That a ranked listing remains available at click
- Multi-node / NCCL outcomes from a single listing row
- GPU performance parity across similar PIS
- Fair cross-lane ranking without lane context
- That broker API latency equals host training-network health
- Open access to the raw tape or exact engine coefficients
Audit trail for buyers
| Artifact | Role |
|---|---|
| This methodology page | Public interpretation; index intent (not internal weights) |
| Quote honesty layer (Listed · TRAC · Δ, lineage, coverage) | Per-decision evidence |
| Methodology version stamp | v1; bump on material index / mark-family changes |
| Design-partner pilot | Ranked marks / file or API, not a database export |
Observational marks only. Not legal, investment, or settlement advice.