İçeriğe geç
MZGrid

Bu sayfa yalnızca İngilizce olarak mevcuttur. Bağlayıcı olan İngilizce sürümdür.

Reputation

How the 0–100 reliability score is computed from completion rate, uptime, benchmark consistency and tenure, and how it affects ranking.

Every provider has a reliability score from 0 to 100. It is shown on every offer, returned as provider.reliability by the API, used by customers to filter offers (minReliability), and worth 35% of the scheduler's ranking score.

Formula

ComponentWeightWindowWhat it measures
Completion rate40%90 daysShare of accepted jobs completed without failure, abandonment or slash.
Uptime30%30 daysShare of time your machines sent heartbeats while listed.
Benchmark consistency20%—How consistent your benchmark results are with each other and with the SKU reference.
Tenure10%capped at 180 daysTime since registration, growing linearly to full credit at 180 days.
Metin
reliability = 40 × completionRate90d + 30 × uptime30d + 20 × benchmarkConsistency + 10 × min(tenureDays, 180) / 180

(each input expressed as a fraction from 0 to 1)

Example

A community provider registered 60 days ago, with 97% completion, 99% uptime and benchmark consistency of 0.95:

Metin
40 × 0.97 + 30 × 0.99 + 20 × 0.95 + 10 × 60/180
= 38.8 + 29.7 + 19.0 + 3.33
= 90.8

After 180 days with the same record the tenure term reaches 10 and the score rises to 97.5.

How reputation affects you

  • Scheduler ranking. Reliability is worth 35% of the autoSelect score (0.35 × reliability / 100). Ten points of reliability are worth 0.035 of score — roughly as much as being 7% cheaper on the normalised price scale.
  • Customer filters. Many customers set minReliability in the marketplace or the API; offers below it are not shown.
  • Pricing power. Higher-reliability providers can price toward the upper end of the live range and still win jobs.
  • Leaderboard. The network leaderboard ranks providers by reliability and GPU-hours.

Improving your score

  • Keep machines online and the agent running; uptime is 30% of the score and is measured continuously.
  • Never let a job end in failure: completion rate is the largest component.
  • Keep your hardware cool and power limits at default so benchmark runs are consistent.
  • Tenure takes care of itself — stay on the network.