Mark Compute

Docs

Observatory architecture

One tape (surface view)

Public listings (sample | live | hybrid)
        → Normalize into a shared schema
        → Identity · control-plane probes · temporal history · execution-risk enrichment
        → Quote marks (SIR · PIS · TRAC) + Panel indices (CireX family)
        → Quote / Panel surfaces

Five stages (buyer-facing)

StageFunctionWhat is public
1 GatewayIngest & normalizeMulti-provider listings → unified schema
2 EnrichmentIdentity & evidenceDeterministic listing identity, broker-level responsiveness probes, append-only history, execution-risk enrichment
3–4 AnalyticsMarks & indicesSIR, PIS, TRAC, stability, coverage, CireX / CIX / Basis
5 PresentationSurfacesBasis Quote, Panel, documentation

Data we use / don’t

We useWe don’t
Public listing metadata, verification flags, availabilityHost agents or root access
Accumulated continuity / downtime from our tapeGPU throughput benchmarks
Broker API responsiveness (control-plane proxy)Equating broker RTT with host training-network health

Proprietary (not on this page)

  • Exact scoring weights, penalty schedules, friction dollar defaults, CireX blend weights
  • Anti-gaming / Sybil heuristics and fingerprint edge cases
  • Raw snapshot databases, probe schedules beyond TTL class, partner outcome telemetry
  • Infrastructure secrets and adapter credentials

Buyers can audit why a mark exists and how to read it. Competitors cannot reconstruct the engine or the tape from this document alone.