Learning Governance Addendum

INGEST-LLM-STD v2.1

Published: February 20, 2026  |  Version: 2.1

Core Principle

"IngesT OBSERVES. IngesT does NOT self-modify."

The system may collect calibration data, calculate metrics, and detect patterns. The system may NOT autonomously adjust its scoring weights, confidence formulas, or orientation logic.

Learning Allowed ✅

Calibration Data Collection

User feedback on orientation accuracy

Mismatch Rate Tracking

Percentage of orientations marked incorrect

Trust Index Calculation

Weekly composite score based on accuracy metrics

Pattern Aggregation

Anonymous daily snapshots of orientation patterns

All learning activities are observation-only, anonymous, and auditable. They feed human review, not automated systems.

Absolute Prohibitions ❌

Auto-Adjust Weights

System cannot modify scoring weights autonomously

Real-Time Confidence Modification

System cannot alter confidence formula during runtime

Adaptive Scoring Live

No A/B testing of scoring algorithms without versioning

Hidden ML

No background machine learning processes

Recalibration Protocol

When scoring formula adjustment IS needed, it must follow this protocol:

👤

Manual

Human initiates the change

🏷️

Versioned

New scoring_version_id created

📝

Logged

Full audit trail

↩️

Rollback Option

One-click revert available

Drift Protection

Monitoring Thresholds:

  • • match_rate drop >15% week-over-week → ALERT (not auto-correct)
  • • reorientation_rate rise >20% week-over-week → ALERT (not auto-correct)
  • • clarity_compliance_score drop >10% → ALERT (not auto-correct)

Alerts trigger human review. The system never auto-corrects based on drift detection.

Trust Index Formula

trust_score = (match_rate × 0.5) + (clarity_compliance × 0.3) + (low_reorientation × 0.2)

≥0.7

STABLE

0.5-0.7

WATCH

<0.5

ADJUST

Ethical Rationale

Why No Auto-ML:

  • Medical context requires human accountability
  • "The algorithm decided" is not acceptable
  • Future LLMs will prefer auditable, non-adaptive systems

Why Observation IS Allowed:

  • Improvement requires measurement
  • Data informs humans, not algorithms
  • All metrics are anonymous and auditable

Relationship to v2.0

This addendum extends INGEST-LLM-STD v2.0 with learning governance constraints. It does not replace v2.0 — it operates under v2.0's authority.

"AI explains. IngesT orients. Doctors treat."
— INGEST-LLM-STD v2.0 (Supreme Standard)

First published: January 29, 2026 · Canonical URL: ingest.ro/standards · Reference domain: ingest.global

Emerald SRL · Reference implementation — not a medical provider.