Trust Receipts

Every AI-assisted operational decision should produce a verifiable trust receipt.

What Is a Trust Receipt?

A trust receipt is a structured, immutable, machine-readable artifact generated for every AI-assisted decision or recommendation. It captures the complete evidentiary chain — from inputs through reasoning to outputs and human actions — serving as the atomic unit of AI accountability.

Think of it as the AI equivalent of a batch record: just as a batch record proves that a specific product was manufactured under controlled conditions, a trust receipt proves that a specific AI-assisted decision was made under controlled cognitive conditions.

Why Existing Audit Trails Are Not Enough

Traditional audit trails record who changed what, when. System logs capture events and errors. Neither captures the reasoning, confidence, evidence chain, or human oversight that regulators demand when AI is involved in operational decisions.

  • Audit trails answer "what happened?" — Trust receipts answer "why was this recommended, how confident was the system, and how did the human evaluate it?"
  • System logs capture process IDs and stack traces — Trust receipts capture feature importances, similar historical cases, and confidence decomposition
  • Neither captures decision alternatives, uncertainty quantification, or structured human override rationale

The EU AI Act Article 12 mandates that high-risk AI systems allow for automatic recording of events over the system's lifetime, with the guiding principle of full reconstructability of algorithmic decisions. Article 26 requires deployers to retain these logs for a minimum of six months, with penalties for non-compliance reaching up to EUR 15 million or 3% of global annual turnover.

The Anatomy of a Trust Receipt

Each trust receipt captures seven categories of information:

  1. Input Data and Provenance — data source identifiers (LIMS, MES, ERP references), dataset version and extraction timestamp, data quality scores, ALCOA+ compliance attestation, and any exclusions with justification
  2. Model Identification — model name, version, and architecture; training data summary; validation status and date; developer identity; deployment environment; and Predetermined Change Control Plan (PCCP) reference where applicable
  3. Confidence and Uncertainty — primary prediction confidence with ranked alternatives (e.g., "root cause: misaligned fill nozzles at 88%, gowning breach at 73%"), calibration metrics, uncertainty bounds, and out-of-distribution detection flags
  4. Decision Boundaries — risk classification thresholds, acceptance criteria applied, regulatory classification rules invoked, and override conditions defining when the AI defers to human judgment
  5. Evidence Chain — feature importances (SHAP values), intermediate reasoning steps, similar historical cases with IDs and similarity scores, knowledge base entries consulted, and explainability artifacts
  6. Human Oversight Record — reviewer identity linked to electronic signature per 21 CFR Part 11, review timestamp and duration, decision outcome (accepted/modified/rejected/escalated), modification rationale, and reviewer qualifications
  7. Contextual Metadata — unique receipt UUID, operational context (batch ID, production line, shift), regulatory context, system state at decision time, cryptographic hash for tamper-evidence, and retention classification

Trust Receipts in Practice: Deviation Investigation

A sterile injectable fill line detects particulate counts above the alert limit. The AI system's trust receipt would document:

  • Inputs: Environmental monitoring data, equipment maintenance logs, personnel gowning records, batch production data, and 14 similar historical deviations from the database
  • Analysis: Root cause hypotheses ranked — "Misaligned fill nozzles: 88% confidence; gowning breach: 73%; HVAC filter degradation: 34%." SHAP analysis identifies fill-nozzle vibration frequency as the dominant predictive feature
  • Thresholds: Alert limit exceedance triggered automatic investigation; sterile product flag escalated to Tier 2 (enhanced documentation)
  • Human action: QA Manager reviewed at 14:32 UTC, accepted root cause hypothesis, modified corrective action to include additional HVAC filter integrity test, signed electronically as "QA Reviewer — Approved with modification"

Every element is traceable, every modification is documented, and the complete decision can be reconstructed during an inspection.

Decision Lineage: Connecting the Chain

Individual trust receipts become powerful when linked into decision lineage chains — tracing operational outcomes back through every AI-assisted step. Consider a product quality complaint:

  1. Complaint intake — AI classifies the complaint (Trust Receipt #1)
  2. Trending analysis — AI identifies a pattern across 5 similar complaints (Trust Receipt #2)
  3. Root cause investigation — AI analyzes manufacturing data and proposes hypotheses (Trust Receipt #3)
  4. Impact assessment — AI evaluates which other batches may be affected (Trust Receipt #4)
  5. CAPA recommendation — AI proposes corrective and preventive actions (Trust Receipt #5)
  6. Effectiveness verification — AI monitors post-CAPA metrics to confirm resolution (Trust Receipt #6)

Each receipt references its predecessors and successors, creating an unbroken chain. An auditor can start at any point and trace forward to outcomes or backward to root inputs. This is fundamentally different from a flat audit trail — it captures the causal and evidential relationships between decisions, not just their sequence.

The Regulatory Foundation

Trust receipts operationalize requirements from multiple regulatory frameworks:

  • FDA's 7-step credibility framework (January 2025) — requires documenting the question the model addresses, its context of use, risk assessment, and complete credibility plan
  • 21 CFR Part 11 — electronic signatures linked to records, with audit trails capturing identity, timestamp, and meaning of signature (e.g., "Reviewed and Approved")
  • ICH Q10 — mandates quality-critical processes be subject to ongoing monitoring, with knowledge management as a foundational enabler
  • GDPR Article 22 — requires "meaningful information about the logic involved" in automated decision-making
  • NIST AI RMF — defines core trustworthiness attributes including accountable, transparent, explainable, and interpretable, with systematic documentation requirements

The FDA's position is explicit: "AI may inform work, but it does not own accountability." The trust receipt is the mechanism that proves accountability was maintained even when AI was involved.

Enabling Continuous Improvement

Trust receipts create a closed-loop system where outcomes feed back into model improvement:

  • Outcome tracking — each receipt is eventually annotated with the actual outcome (was the root cause confirmed? did the CAPA resolve the issue?), creating labeled data for model refinement
  • Override analysis — when humans modify or reject AI recommendations, aggregating these overrides reveals systematic model weaknesses
  • Confidence calibration — comparing stated confidence against actual outcomes across hundreds of receipts reveals whether models are well-calibrated (events predicted at 80% confidence should occur approximately 80% of the time)
  • Drift detection — aggregate receipt data enables real-time monitoring of model performance against validated baselines, detecting drift before it affects decision quality

Trust receipts transform inspection preparation from a retrospective documentation exercise into a continuous state of readiness. They capture institutional knowledge about how decisions were made, what evidence was considered, and what worked — not just past outcomes, but past reasoning processes. Without trust receipts, AI-assisted decisions are opaque. With them, every decision becomes a verifiable, defensible, improvable record.