What is the difference between a human override and a human confirmation?

An override occurs when a human reviewer changes an AI decision — the AI recommended approval, the reviewer rejected it. A confirmation occurs when a reviewer reviews an AI decision and approves it unchanged. Both are valuable training signals: overrides identify where the AI is wrong; confirmations identify hard cases where the AI is correct. Tenet captures both with tenet.record_override() and tenet.record_confirmation(), and both appear in the audit trail as EU AI Act Article 14 compliance evidence.

Does EU AI Act Article 14 require capturing human overrides?

Yes. Article 14 requires high-risk AI systems to enable human oversight and to document when humans intervene in or override AI outputs. Article 14(4) requires the ability for humans to decide not to use AI system outputs. Article 14(5) requires that human oversight measures be appropriate to risk. For documentation purposes, each override must record the actor, timestamp, original AI decision, override decision, and reason — exactly the schema that tenet.record_override() captures.

How many override records are needed for fine-tuning?

It depends on the fine-tuning objective. DPO fine-tuning has produced measurable improvements with 500-2000 high-quality preference pairs. The quality threshold matters more than volume: pairs where the reviewer confidence is high (≥0.7) and the reason category is not REVIEWER_ERROR are more valuable than larger volumes with lower confidence. For domain adaptation, 200-500 targeted examples on specific failure modes can be effective.

What is DPO and why is it relevant for AI agent fine-tuning from overrides?

DPO (Direct Preference Optimization) is a fine-tuning technique that optimizes a language model's policy directly from preference pairs — (prompt, chosen_completion, rejected_completion) tuples — without requiring a separate reward model. Override records map directly to DPO pairs: the original AI decision is the rejected completion, the human override is the chosen completion, and the context snapshot is the prompt. DPO is more stable than RLHF for small datasets and requires less infrastructure.

How do I prevent bias from entering fine-tuning data through override records?

Four practices: (1) exclude records with reason_category=REVIEWER_ERROR — these are cases where the AI was correct but the reviewer was wrong; (2) require a minimum confidence threshold (≥0.7) from reviewers before including pairs; (3) audit override rates by protected class characteristics if the AI operates in a fair lending or employment context; (4) sample from multiple reviewers rather than a single expert to reduce individual bias propagation. Tenet's export API supports all four filters.

How do override frequency patterns detect semantic drift?

Override rate by decision category is a leading indicator of semantic drift. If your loan agent's override rate on DTI-borderline cases increases from 3% to 8% over 30 days, the agent's decision boundary has shifted — before any eval regression is visible. Track: override rate by decision type (rising = AI diverging from correct behavior), policy exception rate rising (= AI misapplying policy), confirmation rate on hard cases decreasing (= AI degrading on difficult examples), and actor disagreement rate (= consistent AI failure mode with multiple reviewers). These signals typically appear 2-4 weeks before eval regressions.

Can override records be used for both RLHF and DPO?

Yes. For RLHF, override records serve as negative reward signal (AI decision that was overridden) and confirmations serve as positive reward signal (AI decision that was approved unchanged). For DPO, each override maps directly to a preference pair without needing a reward model. DPO is generally preferred for smaller datasets (under 50K pairs) because it requires less infrastructure. RLHF becomes competitive at larger scale and when the reward model generalizes beyond the training distribution.

Does capturing human overrides violate HIPAA or GDPR?

Capturing override records requires care in regulated contexts. For HIPAA: override records should use a pseudonymized actor_id rather than the reviewer's name or credentials, and the context snapshot should tokenize or redact PHI fields before storage. For GDPR: the reviewer is typically a data controller's employee, not a data subject — Article 22 rights apply to the patient or customer whose data was processed, not the reviewer. Use Tenet's on-premise VPC deployment if override records must stay within your infrastructure for data residency compliance.

How to Capture Human Overrides of AI Agent Decisions for Fine-Tuning

Q: Why are human override records better training data than synthetic preference pairs?

Human override records capture four properties that synthetic data cannot replicate: production context (the real input state that caused the error), expert demonstration (the correct output from someone accountable for the outcome), failure mode diversity (overrides cluster around the AI's actual blind spots, not hypothetical scenarios), and regulatory anchoring (in regulated industries, overrides often reflect policy constraints the AI misapplied). Synthetic preference pairs rate hypothetical completions; override records demonstrate production decisions.

Human override records are the highest-signal training data available for AI agents. When a loan officer rejects an AI approval, a clinical reviewer modifies a prior auth recommendation, or a compliance analyst reverses an automated flag, that override encodes production context, correct behavior, and edge case handling that RLHF and synthetic datasets cannot replicate. This guide shows how to capture these overrides with tenet.record_override(), structure them as DPO preference pairs, and use override frequency patterns as a semantic drift detection signal — while satisfying EU AI Act Article 14 documentation requirements.

Why Human Overrides Are Your Best Training Data

Standard training pipelines use preference data from human labelers rating synthetic completions. Override records are different: they capture a subject-matter expert making a real production decision, with real consequences, on a case the AI got wrong. Override data has four properties synthetic data lacks: production context (the real input state that triggered the wrong decision), expert demonstration (the correct output from someone accountable for the outcome), failure mode diversity (overrides cluster around the AI's actual blind spots), and regulatory anchoring (in regulated industries, overrides often reflect policy constraints the AI misapplied).

EU AI Act Article 14: Capture Is Required, Not Optional

EU AI Act Article 14 requires high-risk AI systems to enable human oversight and to capture when humans intervene in or override AI decisions. Specifically, Article 14(4) requires systems to allow humans to decide not to use an AI system output, and Article 14(5) requires logging when human oversight is applied. This means: for in-scope systems, override capture is a compliance obligation — not a training optimization. A record must include the actor ID, timestamp, original AI decision, override decision, and reason for the change.

Override Record Schema

A complete override record contains nine fields: session_id (links the override to the originating AI decision), actor_id (pseudonymized identifier of the human reviewer), timestamp (ISO 8601 with timezone), original_decision (the AI output being overridden), override_decision (the human's replacement decision), reason_category (enum: POLICY_EXCEPTION, FACTUAL_ERROR, EDGE_CASE, REVIEWER_ERROR, NEW_INFORMATION), reason_text (free-text justification, optional), confidence (reviewer's stated certainty 0.0-1.0), and outcome (downstream result if tracked). This schema maps directly to a DPO preference pair: the AI decision is the rejected completion, the override decision is the chosen completion.

Implementation: record_override() and record_confirmation()

Install pip install tenet-ai-sdk. Initialize TenetClient with your API key. For override capture, call tenet.record_override() with the session_id from the original AI decision record, actor details, original and override decisions, and a reason from the OverrideReason enum. For confirmation (human approves the AI decision unchanged), call tenet.record_confirmation() — these records are equally valuable: they identify the cases where the AI was right on hard examples. Both calls use the same fire-and-forget Ghost SDK architecture — under 0.3ms blocking overhead.

Exporting DPO Training Data

DPO (Direct Preference Optimization) fine-tuning requires preference pairs: chosen completion vs rejected completion, each with the original prompt context. Use Tenet's export API to retrieve override records filtered by confidence threshold (≥0.7 recommended), reason category (exclude REVIEWER_ERROR), and date range. Each record maps to a DPO pair: the context snapshot is the prompt, the override_decision is chosen, the original_decision is rejected. For RLHF, the same records can be used as reward signal: overrides are negative examples, confirmations are positive.

Override Patterns as Drift Detection Signal

Override frequency by decision category is a leading indicator of semantic drift. If your loan agent's override rate on DTI borderline cases increases from 3% to 8% over 30 days, the agent's decision boundary has shifted — before any eval regression is visible. Track: override rate by decision type, override rate by reason category (policy exception rate rising = agent is misapplying policy), confirmation rate on edge cases (decreasing = agent is degrading on hard examples), and actor disagreement rate (multiple reviewers overriding the same decision = consistent AI failure mode).