Measurement contract
How InfraBench measures and governs RPC paths
InfraBench publishes deterministic route evidence. READY means every configured gate passed. ABSTAIN means at least one blocking gate prevented endpoint handoff; it is a governed result, not an application error.
Collection
Regional probes execute across Washington, D.C. (iad1), San Francisco (sfo1), Frankfurt (fra1), Singapore (sin1), Tokyo (hnd1), Mumbai (bom1), Sydney (syd1), Sao Paulo (gru1), Cape Town (cpt1). Scheduled cycles measure getSlot, getLatestBlockhash, getBalance, getTokenAccountsByOwner and record WebSocket connection evidence for MEV-sensitive paths.
Providers are measured sequentially with no retries or parallel dispatch. Interactive routing uses three bounded getSlot samples. Scheduled JSON-RPC calls use a 3.5-second timeout and WebSocket evidence uses a four-second timeout. The requested cadence is every 15 minutes.
Raw events include round-trip samples, returned slot height, success rate, method, region, timestamp, and normalized error context. Raw observations, credentials, and runner internals remain in private storage and are never copied into public artifacts.
Scoring and governance
The composite score is a workload-specific weighted execution cost. Lower values indicate a stronger eligible path. A score ranks candidates only after evidence exists; it never overrides governance gates.
Thresholds are governed by the versioned registry thresholds-2026-06-16, effective 2026-06-16T00:00:00.000Z. The current policy and change history are published at /api/v1/thresholds.
score = (p95_RTT x p95_weight) + (slot_drift x drift_weight) + (WebSocket_connect_ms x WebSocket_weight) + ((100 - success_rate) x failure_weight)
| Workload | P95 RTT | Slot drift | WebSocket | Failure penalty | RTT budget |
|---|---|---|---|---|---|
| trading | 0.45 | 0.35 | 0.1 | 0.1 | ≤ 45 ms |
| reads | 0.5 | 0.3 | 0.05 | 0.15 | ≤ 150 ms |
| indexing | 0.35 | 0.45 | 0.05 | 0.15 | ≤ 150 ms |
| mev | 0.3 | 0.4 | 0.2 | 0.1 | ≤ 100 ms |
Slot drift is the regional maximum observed slot minus the provider slot. READY requires at least 95% request success, available p95 RTT and slot evidence, RTT inside the workload budget, and slot drift no greater than 3. MEV also requires WebSocket timing and at least 95% WebSocket success.
Confidence
Composite confidence is a deterministic weighted result from four 0-100 sub-scores, then multiplied by measurement integrity. Regional and coverage each contribute 25%, freshness contributes 20%, and measurement contributes 30% before the multiplier is applied.
confidence = weighted_subscores x (measurement / 100)
- Regional
- Measured regions divided by the nine configured probe regions.
- Coverage
- Observation count relative to the evidence sufficiency target.
- Freshness
- Recency of the latest provider observation, decaying as evidence ages.
- Measurement
- Average successful response rate for the provider evidence rows.
A Measurement sub-score of zero means the current evidence did not contain a usable successful measurement. It forces composite confidence into the no-usable-evidence band instead of allowing regional, coverage, or freshness context to appear moderately confident on its own.
Known limitations
- 1. Probes are read-only. They measure RPC response behavior and do not submit transactions or estimate landing probability.
- 2. JSON-RPC coverage spans four methods, but method-specific workload models remain intentionally limited. gRPC is not measured.
- 3. WebSocket connection and first-message timing are collected; sustained throughput and message fanout are not measured.
- 4. Scores describe point-in-time rolling snapshots. A provider may change between probe cycles.
- 5. Every region is evaluated independently. Missing regions remain explicit gaps and are never inferred from nearby observations.
- 6. Providers are not notified of probe timing. Rate limiting or edge-IP treatment remains part of observed behavior rather than being filtered out.