Skip to content

Latest commit

 

History

History
129 lines (67 loc) · 6.7 KB

File metadata and controls

129 lines (67 loc) · 6.7 KB

Glossary

Document Version: 2.3.0 | Last Updated: 2026-07-01


This glossary defines key terms, concepts, and abbreviations used throughout the ASR Feedback Intelligence Platform documentation.


A

AGI (Artificial General Intelligence): An AI system with the ability to understand, learn, and apply knowledge across a wide range of tasks at a human-equivalent level. ASR Feedback evaluates AGI-class systems alongside narrower AI tools.

Annotation: The act of attaching structured feedback to a specific AI response. In ASR's methodology, annotations are organized across four pillars: Good, Bad, Learned, and Remember.

Audit Trail: An immutable, chronological record of all actions taken on a feedback entry — from creation through review and any modifications. See compliance/audit-trail.md.

B

Bad Signal: Pillar 2 of the ASR 4-Pillar Methodology. Captures what the AI response did poorly — errors, failures, and quality gaps requiring correction. See METHODOLOGY.md.

C

Calibration: The process of ensuring evaluators apply consistent standards when scoring and annotating. New evaluators undergo a multi-phase calibration process, and ongoing calibration is maintained through monthly exercises.

Calibration Score: A numeric score (0.0–1.0) representing an evaluator's agreement with gold-standard annotations. Minimum threshold for independent evaluation: 0.80.

Category Code: A standardized identifier for a specific type of feedback (e.g., GOOD_REASONING, BAD_HALLUCINATION, LEARN_EDGE_CASE). Defined in the taxonomy.

Cohen's Kappa (κ): A statistical measure of inter-rater reliability that accounts for chance agreement. ASR targets κ ≥ 0.80 across all evaluators.

Composite Score: The overall quality score (0–100) assigned to an AI response, calculated as a weighted combination of Good Signal Score, Bad Signal Penalty, Severity Impact, Completeness Score, and Consistency Score.

D

Domain Tag: A classification label indicating the industry or subject area of an AI response (e.g., healthcare, finance, education). Defined in taxonomy/domain-tags.yaml.

Dual Evaluation: The process of having two independent evaluators assess the same AI response, used for 10% of all entries to measure inter-rater reliability.

E

Entry ID: The unique identifier for a single feedback entry. Format: FB-YYYY-NNNN-NNN (e.g., FB-2026-0547-001).

Evaluator: A trained, calibrated human expert who performs 4-pillar analysis on AI responses. All evaluators must maintain a calibration score ≥ 0.80.

Evidence: Specific quotes, line numbers, or references from the AI response that support a feedback annotation. All annotations must include evidence.

F

Feedback Entry: The atomic unit of the ASR Feedback system. A single, schema-validated annotation capturing observations across all four pillars for one AI response. See FEEDBACK_SCHEMA.md.

Four Pillars: The core framework of ASR's methodology: ✅ Good Signal, ❌ Bad Signal, 📘 Learning Capture, 🧠 Memory Persistence.

G

Gold Standard: A set of pre-evaluated AI responses with "correct" annotations, used for calibration and inter-rater reliability measurement.

Good Signal: Pillar 1 of the ASR 4-Pillar Methodology. Captures what the AI response did well — behaviors, reasoning, and qualities to reinforce.

H

Hallucination: An AI-generated claim that is factually incorrect, fabricated, or unsupported by any source. Classified as BAD_HALLUCINATION with typical severity P0–P1.

I

Inter-Rater Reliability (IRR): The degree of agreement among independent evaluators. Measured using Cohen's Kappa (κ) and Krippendorff's Alpha.

K

Krippendorff's Alpha: A reliability coefficient for measuring agreement among multiple raters, accommodating ordinal, nominal, and interval data. ASR targets α ≥ 0.80.

L

Learning Capture: Pillar 3 of the ASR 4-Pillar Methodology. Records novel insights, edge cases, and discoveries that emerge during evaluation.

LLM (Large Language Model): A neural network trained on large text corpora to generate human-like text. Examples: GPT-4o, Claude, Gemini, Llama.

M

Memory Persistence: Pillar 4 of the ASR 4-Pillar Methodology. Creates durable institutional memory — persistent rules, client preferences, and critical guardrails.

Memory Rule: A persistent evaluation rule stored in the Memory Store. Types include Client Preference, Safety Rule, Domain Rule, Model Behavior, Correction Pattern, and Quality Standard.

N

NPS (Net Promoter Score): A client satisfaction metric calculated as % Promoters - % Detractors. Scale: -100 to +100.

P

P0 / P1 / P2 / P3 / P4: Severity levels for Bad Signal classifications. P0 = Critical (safety risk), P4 = Trivial (cosmetic). See taxonomy/severity-levels.yaml.

Pillar: One of the four dimensions of ASR's feedback methodology. Each pillar captures a distinct type of signal from the evaluation.

Q

Quality Score: See Composite Score.

R

Reinforcement Value: A numeric score (0.0–1.0) assigned to Good Signal entries, indicating how strongly the positive behavior should be reinforced in training.

Remediation: A concrete recommendation for fixing an issue identified in a Bad Signal annotation.

Root Cause: The underlying reason why an AI response exhibited a particular bad behavior (e.g., "Training data may not include asymptomatic presentations").

S

Schema Validation: The process of checking that a data payload conforms to its formal JSON Schema definition. All feedback entries must pass schema validation before storage.

Session: A bounded period during which an evaluator reviews multiple AI responses. Sessions have a start time, end time, and are associated with a specific client and model.

Severity Level: A P0–P4 classification indicating the impact of a Bad Signal. See METHODOLOGY.md.

SLA (Service Level Agreement): Contractual commitments for feedback turnaround time, quality standards, and availability. See compliance/sla-definitions.md.

T

Taxonomy: The hierarchical classification system for feedback categories, severity levels, model profiles, and domain tags. See taxonomy/.

Turnaround Time: The elapsed time between an AI response being submitted for evaluation and the completed feedback being delivered to the client.


See Also: