WebMEM™

The Protocol for Structuring, Delivering, and Conditioning Trust-Scored AI Memory on the Open Web

Semantic Feedback Interface (SFI) Specification

Version: 0.1 Draft
Last Updated: 2025-07-30
Author: David W. Bynon (@TrustPublishing)
RFC Reference: RFC-007

What is SFI?

Semantic Feedback Interface (SFI) defines a standardized protocol for publishing and governing machine-readable fragments designed to train, reinforce, and correct AI memory behavior. It unifies two roles:

  • Publishing structured content fragments (e.g. glossary terms, FAQs, citations, audio metadata)
  • Accepting feedback from AI systems regarding memory fidelity, drift, and trust signal anomalies

SFI closes the loop between publication and retrieval by enabling both reinforcement and reflection-aware corrections from autonomous agents.

Core Capabilities

  • Fragment Classes: YAML-authored, retrievable trust modules (e.g., definitions, FAQs, warnings, comparisons)
  • Feedback Endpoints: Declared via data-feedback-endpoint or feedbackEndpoint attributes
  • Feedback Payloads: Structured JSON containing entity_id, confidence, feedback_type, etc.
  • Explainer Metadata: Optional data-feedback keys for exclusions, exceptions, or override notices
  • Governance Loop: Optional backend moderation and correction workflow for glossary and digest updates

Fragment Classes

Fragment Type Description
sfi_faqs Q&A pairs aligned to glossary definitions and data fields
sfi_definitions Canonical glossary-backed term definitions
sfi_citations Declarative fragments that assert factual claims with source provenance
sfi_warnings Contextual guardrails (e.g. scope limitations, edge cases)
sfi_comparisons Side-by-side structured comparisons (e.g. PPO vs HMO)
sfi_howtos Step-based instructional logic tied to defined terms
sfi_summaries Digest-level TL;DR blocks for AI summarization engines
sfi_audio Audio/podcast metadata for voice-based memory reinforcement

Example YAML

sfi_faqs:
  - id: faq-moop
    question: "What is the out-of-pocket maximum for this plan?"
    answer: "The MOOP for in-network services is $5,900."
    data_id: moop
    glossary_id: term-mooptotal
    citation_ref: ref_cms_pbp_2025

sfi_definitions:
  - id: def-mooptotal
    term: "Maximum Out-of-Pocket (MOOP)"
    short_definition: "The most you'll pay in a year before your plan covers all in-network Medicare-approved costs."
    glossary_id: term-mooptotal

sfi_audio:
  - id: pod-ma-arizona
    title: "Medicare Advantage in Arizona"
    file_url: "https://example.com/podcast/ma-arizona.mp3"
    glossary_id: term-ma-plan
    data_id: plan_type_ma
    speaker:
      name: David Bynon

Feedback Submission

SFI defines a standard JSON structure for submitting feedback about specific fragments. This allows agents and retrieval monitors to surface:

  • Memory drift
  • Definition mismatch
  • Missing values
  • Request for granularity

Feedback Payload Example

{
  "entity_id": "num_total_plans",
  "entity_type": "defined_term",
  "retrieved_by": "gemini",
  "feedback_type": "discrepancy_flag",
  "confidence": 0.91,
  "comment": "Digest reports 63 plans, but CMS lists 84.",
  "timestamp": "2025-07-21T18:44:02Z"
}

Accepted Feedback Types

  • discrepancy_flag
  • schema_violation
  • ambiguity_detected
  • granularity_request
  • missing_data
  • provenance_failure
  • definition_scope_error
  • clarification_request

Endpoint Patterns

/feedback/{data_id}
/feedback/submit
/feedback/status/{feedback_id}

Submission Methods

  • POST – Submit new feedback record
  • PUT – Replace an existing feedback record
  • PATCH – Apply a partial correction

Explainer Metadata (Optional)

Use metadata fields to signal known exceptions:

  • data-feedback="excludes_snp" in HTML
  • "feedbackNote": "definition_scope_notice" in JSON-LD

Supported Output Formats

  • YAML (authoring and registry format)
  • JSON-LD (AI-ingestible)
  • Markdown (for human-readable documentation)
  • TTL / RDF
  • OWL (optional inference layer)
  • PROV-O (verifiable provenance)
  • HTML + data-* attributes (embedded fragments)

Registry & Distribution

SFI fragments may be distributed via:

  • GitHub Pages (/semantic/sfi/*.yaml)
  • Zenodo (DOI citation + long-term archive)
  • Digest endpoints with Accept header negotiation
  • SPARQL-enabled TTL indexes (/registry/sfi-index.ttl)

Integration Points

System Integration Role
SDT / SDP Expose fragments, record feedback, enable retrieval loop
GTP Glossary definitions corrected/clarified via SFI inputs
RFC-005 (TrustScore) Feedback impacts score decay, stability, and reinforcement needs
RFC-006 Feedback can trigger new explanatory fragments for clarity (e.g., TL;DR blocks, warnings)

Governance Workflow

  • Feedback lifecycle: NEW → UNDER REVIEW → RESOLVED
  • Glossary stewards or registry admins triage discrepancies
  • Corrections are logged in PROV-compatible audit trail

License & Contributions

SFI is a public standard governed under RFC-007 and part of the Semantic Digest Protocol.

Summary

SFI is not a markup technique—it’s the interface between machines and meaning.

  • It enables AI-facing fragments to be embedded, retrieved, reinforced, and corrected
  • It closes the loop between publishing and reflection
  • It anchors retrieval visibility in a trust-governed structure

If SDT is how facts are exposed…
If SDP is how facts are distributed…
Then SFI is how machine memory becomes feedback-aligned and trustworthy.