Transparency Report

Real numbers from a one-time test run of the Spineforge Gateway: the path every IDE, CLI, custom and cloud-hosted agent takes through us. We don't invent metrics or use unqualified marketing language.

Point-in-time snapshot. This is not a live dashboard. These numbers represent a single test pass executed on September 4, 2026 against our production infrastructure running on the Supabase Free Tier.

Note on Database Limitations: The throughput ceilings shown below are explicitly bound by the Supabase Free Tier limits (maximum 60 concurrent database connections, restricted token buckets). If Spineforge were deployed on a paid or dedicated tier, these throughput thresholds would scale significantly higher.

System Limits & Throughput (Gateway)

Max Sustained QPS (Free Tier Bound)

150 req/sec

Measured via: Constant-arrival-rate test against the OpenAI proxy endpoint. At 150 RPS, p95 latency overhead was ~32ms. Bottleneck remains the Supabase Free Tier connection limits for validating tokens. (As of September 4, 2026)

Proxy Throughput (LLM Streaming)

>2,000 tokens/sec

Measured via: High-speed streaming proxy test bypassing the LLM via mock responder. Streaming chunks are processed with near-zero buffering. (As of September 4, 2026)

Reliability

Uptime Target (SLO)

99.9%

Measured via: Active synthetic monitoring (UptimeRobot) tracking continuous health checks across our global edge endpoints.

Gateway Latency Overhead

The Gateway sits between your agent and the LLM provider. This inherently adds a small amount of latency to each request.

Average Latency Added Per Call

~15 - 35ms

Measured via: Time-to-first-byte (TTFB) difference between a direct OpenAI call and a Spineforge Gateway proxy call.

What affects this latency?

Database Lookups: Every gateway request requires a fast Database lookup to validate the sfgw- token, check agent budget limits, and verify security scopes before proxying.
Geographic Routing: If your agent is in US-East, the Spineforge Gateway is in US-East, and the OpenAI server is in US-West, the network hops will add physical travel time. We deploy our backend globally to minimize this, but the database region plays a major role.
Streaming vs Blocking: Once the initial validation finishes, streaming responses (stream: true) are piped instantly back to the client. The overhead is only incurred on the initial request handshake, not per-token.

Accuracy

Cost Calculation Accuracy

100%

Measured via: Table-driven correctness test comparing computed cost against rate_cards expected values. (As of September 4, 2026)

Audit Log Completeness

100%

Measured via: Spot-checking audit log records against a known action count. (As of September 4, 2026)