Skip to content

RFC-026: User Feedback Signal — Reference Register

RFC-026: User Feedback Signal — Reference Register

© 2024-2026 Kanjani AI Research & Causum. All rights reserved.


Reinforcement Learning from Human Feedback (RLHF)

# Reference Relevance to RFC
1 Kaufmann, T. et al. “Reinforcement Learning from Human Feedback: A Statistical Perspective.” arXiv:2604.02507, 2026. Comprehensive statistical survey of RLHF — provides theoretical grounding for why feedback signals (like AIGP’s FEEDBACK message) are necessary for calibrating AI governance decisions against real user experience.
2 Casper, S. et al. “Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback.” arXiv:2307.15217, 2023. Systematizes RLHF flaws including noisy/sparse human feedback — validates AIGP’s design choice to capture structured metadata (NLP analysis, specificity, expectation_gap) alongside binary thumbs_up/down to improve signal quality.
3 Li, S. et al. “A Survey of Reinforcement Learning from Human Feedback.” arXiv:2312.14925, 2024. Surveys RLHF as a variant of RL learning from human feedback rather than engineered reward functions — contextualizes AIGP’s FEEDBACK message as a governance-layer feedback collection mechanism distinct from model training feedback.
4 Gao, H. et al. “AI Alignment through Reinforcement Learning from Human Feedback? Contradictions and Limitations.” arXiv:2406.18346, 2024. Critically evaluates attempts to align AI with human values through feedback methods — validates why AIGP separates governance feedback (policy calibration) from model training feedback (RLHF), avoiding conflation of governance signals with training signals.
5 Zhong, H. et al. “Objective Mismatch in Reinforcement Learning from Human Feedback.” arXiv:2311.00168, 2024. Demonstrates that optimizing for human preference feedback doesn’t necessarily optimize for downstream evaluation metrics — validates AIGP’s approach of using feedback for governance calibration (guardrail tuning) rather than direct model optimization.

User Feedback Measurement and Design

# Reference Relevance to RFC
6 Microsoft Data Science. “Beyond Thumbs Up and Thumbs Down: A Human-Centered Approach to Evaluation Design for LLM Products.” 2025. Argues simplistic thumbs up/down mechanisms fail to capture nuances for meaningful improvements — directly motivates AIGP’s NLP analysis object (sentiment, specificity, expectation_gap, topics) augmenting binary ratings.
7 Tianpan.co. “Behavioral Signals That Actually Measure User Satisfaction in AI Products.” 2026. Demonstrates that feedback widget placement and timing dramatically affect signal quality — validates AIGP’s design where feedback is captured while users are “still evaluating” rather than through delayed surveys.
8 Tianpan.co. “Why Your AI Metrics Are Lying to You: The Feedback Signal Timing Problem.” 2026. Shows that immediate post-interaction feedback measures momentary state rather than actual utility — contextualizes AIGP’s session-level feedback aggregation (via session_id) for longitudinal signal quality.
9 Kalloori, S. et al. “Scalable User Feedback via Implicit Sentiment in Developer Prompts.” arXiv:2509.18361, 2025. Addresses the challenge of evaluating AI satisfaction at scale when explicit ratings are sparse — validates AIGP’s dual approach: explicit ratings (thumbs_up/down) plus implicit NLP-derived signals from comment text.
10 Holmes, G. et al. “The Chatbot Usability Scale: Design and Pilot of a Usability Scale for Interaction with AI-Based Conversational Agents (BUS-15).” 2021. Develops a validated instrument for measuring chatbot interaction quality — provides psychometric context for AIGP’s feedback measurement approach and validates the need for structured metrics beyond binary satisfaction.

Privacy-Preserving User Analytics

# Reference Relevance to RFC
11 Pilipiszyn, A. et al. “A General Pseudonymization Framework for Cloud-Based LLMs.” arXiv:2502.15233, 2025. Proposes pseudonymization for LLM interactions to protect user privacy — directly validates AIGP’s pseudonym_id design (HMAC-derived, non-reversible, app-scoped) for feedback collection without user identification.
12 European Court of Justice. “Pseudonymised Data Classification under GDPR.” Case T-557/20 (SRB v EDPS), September 2025. Landmark ruling that pseudonymised data is not automatically personal data for all parties — provides legal basis for AIGP’s pseudonym_id design where the governance server cannot re-identify users.
13 UK Information Commissioner’s Office. “Pseudonymisation Guidance.” ICO, 2025. https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-sharing/anonymisation/pseudonymisation/ Regulatory guidance on pseudonymisation as a risk reduction and security improvement technique — contextualizes AIGP’s design as a GDPR-compatible approach to behavioral analysis without identification.
14 Worklytics. “Building a GDPR-Compliant AI Usage Analytics Program Without Storing Personally Identifiable Data.” 2025. Practical guidance on AI usage analytics that don’t create additional privacy risks — validates AIGP’s approach where governance server stores metadata only (no comment content, no prompt/response content).

NIST AI RMF Alignment

# Reference Relevance to RFC
15 NIST. “AI 100-1: Artificial Intelligence Risk Management Framework.” National Institute of Standards and Technology, 2023. Defines MAP 3.2 (scientifically valid and reliable AI system evaluation) and MEASURE 2.6 (structured evaluation through human subject interaction) — these are the specific NIST requirements that RFC-026’s FEEDBACK message addresses.
16 NIST. “MEASURE-2.2: Human Subject Evaluations.” AI RMF Playbook, 2023. Specifies that evaluations involving human subjects must be representative and meet applicable requirements — validates AIGP’s pseudonymized, consent-compatible feedback collection design.

Parent-Child Evidence Relationships

# Reference Relevance to RFC
17 W3C. “PROV-O: The PROV Ontology.” W3C Recommendation, 2013. https://www.w3.org/TR/prov-o/ W3C standard for representing provenance information — AIGP’s parent-child relationship model (FEEDBACK_ON with parent_id/parent_ref) follows established provenance patterns for derivation and attribution.
18 Moreau, L. et al. “The Open Provenance Model Core Specification (v1.1).” Future Generation Computer Systems, 2011. Foundational work on provenance graphs with causal relationships — validates AIGP’s three-level parent-child discovery (S3 path convention, bottom-up parent_id, top-down has_children) for evidence chain construction.

Sentiment Analysis and NLP Heuristics

# Reference Relevance to RFC
19 Liu, B. Sentiment Analysis: Mining Opinions, Sentiments, and Emotions. 2nd ed. Cambridge University Press, 2020. Comprehensive treatment of sentiment analysis including intensity scoring — provides the theoretical basis for AIGP’s NLP analysis object (sentiment float -1.0 to 1.0, severity_language classification).
20 Devlin, J. et al. “BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding.” NAACL-HLT, 2019. Foundation model enabling lightweight topic extraction and sentiment classification — contextualizes AIGP’s requirement for heuristic NLP analysis “without LLM call required” by referencing lightweight transformer-based approaches.

Last updated: June 2026

Additional Sources (provided by author)

# Reference Relevance to RFC
21 Xu, Y. et al. “Towards Boosting the Open-Domain Chatbot with Human Feedback.” Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL), pp. 4060–4078, Toronto, 2023. https://aclanthology.org/2023.acl-long.224/ Demonstrates how structured human feedback (beyond binary ratings) improves chatbot quality through iterative refinement — validates AIGP’s approach of capturing comment text with NLP analysis alongside thumbs_up/down for actionable governance signal.
22 Lindström, E. “Human Feedback Mechanisms for AI Systems.” Master’s Thesis, Linköping University, 2023. https://www.diva-portal.org/smash/get/diva2:1782678/FULLTEXT01.pdf (local copy: specification/papers/FULLTEXT01.pdf) Systematic analysis of human feedback mechanisms for AI systems — provides the empirical and design foundation for AIGP’s FEEDBACK message type, including discussion of feedback timing, modality (explicit vs. implicit), and privacy considerations in feedback collection.