Because failover switches. Correctover verifies.
Circuit breaker + health scoring + L3 failover in <1s. Your LLM calls don't fail — they self-heal.
BYOK direct connect — zero markup, zero token resale. 6-dimension contract validation catches silent overcharges.
MAPE-K loop decides in 50–100μs. Drift detection + contract verification = every silent failure is visible.
Health scoring + drift detection spots anomalies in real-time
6-dimension contract validation: schema, latency, cost, format, semantic, compliance
87 self-healing rules auto-remediate — retry, re-route, re-prompt, or escalate
Loop closes: result verified, contract met, SLA maintained
| Capability | Failover | Correctover |
|---|---|---|
| Switches provider on error | ✓ Yes | ✓ Yes |
| Validates response contract | ✗ No | ✓ Yes |
| Detects silent model swap | ✗ No | ✓ Yes |
| Catches schema drift | ✗ No | ✓ Yes |
| Prevents cost overruns | ✗ No | ✓ Yes |
| Self-heals autonomically | ✗ No | ✓ Yes |
| Runs embedded (no gateway) | ✗ No | ✓ Yes |
| BYOK zero markup | ✗ No | ✓ Yes |
Correctover is an embedded LLM API reliability runtime that validates response contracts before accepting them. Unlike traditional failover that only checks if a provider responded, Correctover verifies 6 dimensions: structure, schema, latency, cost, identity, and integrity. It runs in-process as a Python or JavaScript SDK with zero external dependencies.
Failover switches providers when one fails. Correctover verifies that the response is contractually correct before accepting it. Failover checks "did Provider B respond?" while Correctover checks "is the response correct?" — catching silent model substitution, schema drift, cost overruns, and semantic quality degradation that failover cannot detect.
The 6-dimension CANON validation engine runs in 22μs P50 (99μs P99). The MAPE-K self-healing loop operates in 50-100μs per decision cycle. L3 failover completes in 949ms E2E including full contract validation on the new provider's response. The overhead is less than 0.01% of a typical LLM API call.
Correctover is Apache-2.0 licensed with a commercial restriction. It is free for development and non-commercial use. A commercial license is required for production deployments. The core validation engine is distributed as compiled bytecode for performance.
Install via pip: pip install correctover. Or via npm: npm install correctover. The Python package requires Python 3.8+ and has one dependency (httpx).
No. Correctover is an embedded SDK that runs inside your application process. It is not a gateway, proxy, or SaaS. This means zero network overhead, zero external dependencies, zero data leakage (BYOK direct connect), and zero markup on token costs.
MAPE-K stands for Monitor, Analyze, Plan, Execute, Knowledge. Correctover monitors LLM API responses in real-time, analyzes failures using 87 classification rules, plans remediation strategies (L1 retry, L2 semantic downgrade, L3 failover, L4 flywheel learning), executes the plan automatically, and accumulates knowledge for future incidents.
Correctover validates 6 dimensions on every LLM response: Structure (response format matches expected schema), Schema (required fields present and correctly typed), Latency (response time within SLA bounds), Cost (token usage matches expected range), Identity (model field matches what was requested, detecting silent substitution), and Integrity (output meets semantic quality threshold).