When evaluating identical candidates against similar job descriptions, you may occasionally observe minor variances in candidate scoring explanations (Reasoning API). This behaviour is expected and stems from how underlying AI models process data and cache results.
The HrFlow.ai Reasoning API relies on Large Reasoning Models (LRMs) which are inherently probabilistic. Whether the generated explanations remain deterministic (identical) depends entirely on the Profile Keys and Job Keys passed in your API requests.
How Results Are Processed
Same Profile Key + Same Job Key (Deterministic)
Results are 100% reproducible. HrFlow.ai implements a key-based caching mechanism. When a request matches an existing Profile Key and Job Key combination, the API retrieves the stored explanation from the cache rather than querying the model again.
Different Keys or New Entries (Probabilistic)
If a job description or profile is assigned a new key - even if the underlying text content is identical or differs by only a single character (e.g., "Manager" vs. "manager") - the request bypasses the cache. Because LRMs are probabilistic, generating a new evaluation for a new key pair will naturally yield slight variations in the output.
Best Practice for Consistency
Slight score variations (typically a minor percentage difference) across different job IDs or slight text modifications fall within normal operating tolerances for AI-driven scoring.
To guarantee identical reasoning explanations across your workflows, always reuse existing Profile Keys and Job Keys when evaluating the same candidate against the same role. Creating duplicate job entries with distinct keys will trigger fresh model inferences and introduce minor output variations.