DEV Community

correctover
correctover

Posted on

Correctover: The Verified Failover SDK for LLM APIs — Complete Product Guide

#python #ai #llm #opensource

Correctover可瑞沃 is the first verified failover SDK for LLM APIs. Unlike traditional gateways that switch providers when HTTP 200 comes back, Correctover validates every response across 6 dimensions before accepting a failover.

pip install correctover
Enter fullscreen mode Exit fullscreen mode

What Makes Correctover Different

Every AI gateway offers failover. But "failover" in 2026 means: provider A returned a non-200, so route to provider B. That's transport-level failover. It tells you nothing about whether provider B's response is actually correct.

Correctover = Verified Failover. Each response from a backup provider is validated against a configurable contract before it reaches your application. If the contract fails — truncation, schema mismatch, cost overrun, latency breach — Correctover rolls back and tries the next provider.

The 6-Dimension Contract Validation (CANON)

Dimension What It Checks
Schema JSON structure matches expectations
Latency Response within SLA bounds
Cost Token usage not anomalous
Completeness No truncation or missing fields
Identity Correct provider identified
Integrity Response not corrupted

Benchmark: P50 = 22µs, P99 = 99µs per validation pass. This is 3-4 orders of magnitude faster than the LLM call itself.

Key Features

Verified Failover

Traditional: "HTTP 200 = success" → switch providers
Correctover: switch → validate 6-dimension contract → confirm or rollback

MAPE-K Self-Healing Loop

Monitor → Analyze → Plan → Execute → Knowledge. 87 self-healing rules that grow automatically via the knowledge loop.

BYOK Direct Connect

Zero markup, zero relay, zero lock-in. Your keys connect directly to providers. Correctover is an embedded SDK, not a proxy gateway.

Drift Detection

Real-time monitoring across all 6 dimensions with automatic alerting and verified failover when drift exceeds thresholds.

Checkpoint Recovery

Long-chain agent tasks can resume from interruption points instead of restarting.

Benchmarks

Metric Value
Contract validation P50 22µs
Contract validation P99 99µs
MAPE-K decision loop 50-100µs
L3 Failover E2E (with validation) 949ms
Overhead per LLM call <0.01%
Self-healing rules 87 (growing)
Supported providers 8+ (OpenAI, Anthropic, Google, Groq, DashScope, Azure, DeepSeek, etc.)
SDK size ~375 KB
Dependencies 1 (httpx)

Quick Start 

from correctover import NeuralReliabilityEngine

engine = NeuralReliabilityEngine()

# One call, multi-provider, auto-failover with verification
response = engine.chat_completion(
    messages=[{"role": "user", "content": "Explain verified failover"}],
    providers=["openai", "anthropic", "google"],
    contract={
        "max_latency_ms": 5000,
        "max_cost_tokens": 2000,
        "require_json": True
    }
)
Enter fullscreen mode Exit fullscreen mode

Installation

Python:

pip install correctover
Enter fullscreen mode Exit fullscreen mode

Node.js:

npm install correctover
Enter fullscreen mode Exit fullscreen mode

Why Not a Gateway?

Traditional AI gateways operate as reverse proxies — your API calls go through their infrastructure. This means:

  • You pay markup on every token
  • Your data passes through third-party servers
  • You're locked into one vendor's routing logic
  • Failover is transport-level (HTTP 200 only)

Correctover runs in your process. Your keys stay local. Your data never leaves your infrastructure. Failover includes semantic validation, not just HTTP status codes.

Failover switches. Correctover verifies.

Website: correctover.com | PyPI: pip install correctover | npm: npm install correctover | GitHub

Top comments (0)