Skip to content

Latest commit

 

History

History
277 lines (205 loc) · 11.9 KB

File metadata and controls

277 lines (205 loc) · 11.9 KB

UPIF Conformance Specification

Version: 1.0
Date: April 18, 2025
Author: Roshan George Thomas, XWHYZ
DOI: 10.5281/zenodo.15242693
Status: Active


Overview

This document defines conformance requirements for implementations of the Unified Prompt Intelligence Framework (UPIF). It specifies what it means for a system, platform, or tool to claim UPIF compliance across three conformance tiers.

UPIF is a modular, seven-layer protocol. Conformance is layer-scoped — implementations may adopt individual layers independently without implementing the full stack, provided they meet the requirements defined for each adopted layer.


Conformance Tiers

Tier 1 — Minimal Conformance

A system achieves Minimal Conformance by implementing at least one of the seven UPIF layers in full accordance with the layer requirements defined in this document.

Minimal Conformance is appropriate for:

  • Single-purpose tools (e.g. a compliance firewall plugin)
  • Experimental or proof-of-concept implementations
  • SDK modules targeting a single layer

A Minimal Conformant implementation must:

  • Implement all MUST-level requirements for the adopted layer(s)
  • Correctly handle prompt metadata according to the UPIF JSON schema (Appendix A of the whitepaper)
  • Not claim full or partial UPIF conformance for unadopted layers

Tier 2 — Partial Conformance

A system achieves Partial Conformance by implementing three or more UPIF layers, including L4 (Attribution Ledger) as a mandatory component.

L4 is mandatory at this tier because authorship tracking is the minimum auditable signal required to distinguish UPIF-governed prompts from ungoverned ones.

A Partial Conformant implementation must:

  • Meet all MUST-level requirements for each adopted layer
  • Implement L4 in full
  • Produce UPIF-compliant prompt metadata objects for all governed prompts
  • Expose a mechanism for downstream systems to verify prompt provenance

Tier 3 — Full Conformance

A system achieves Full Conformance by implementing all seven UPIF layers in accordance with this specification.

Full Conformance is the target for:

  • Enterprise AI governance platforms
  • Multi-modal content generation pipelines
  • Regulated-industry AI deployments (healthcare, legal, government)
  • Platforms seeking UPIF certification or licensing

A Full Conformant implementation must:

  • Meet all MUST-level and SHOULD-level requirements for all seven layers
  • Produce, store, and expose complete prompt metadata objects
  • Support revision history and contributor attribution across the full prompt lifecycle
  • Pass all test cases defined in the UPIF conformance test suite (forthcoming in v1.1)

Layer-by-Layer Requirements

L1 — Co-Prompting Interface

Purpose: Enable multi-user, real-time collaborative prompt authoring with version control and role-based access.

Requirement Level Description
Multi-user sessions MUST Support two or more simultaneous contributors to a single prompt
Role differentiation MUST Distinguish at minimum between author and reviewer roles
Version control MUST Maintain a revision history with contributor UID and timestamp per revision
Conflict resolution SHOULD Provide a defined strategy for concurrent edits
Access control SHOULD Support permission scoping per contributor role
Real-time sync MAY Provide live synchronization of prompt state across contributors

L2 — Cross-Modal Router

Purpose: Route prompts across modalities — text, image, audio, video, and code — with appropriate transformation metadata.

Requirement Level Description
Modality declaration MUST Every routed prompt must declare its source and target modality
Routing metadata MUST Attach routing decision and target modality to prompt metadata object
Text support MUST Support text-to-text routing at minimum
Multi-modal support SHOULD Support at least two of: image, audio, video, code
Fallback handling MUST Define behavior when target modality is unavailable
Transformation log SHOULD Record modality transitions in the prompt revision history

L3 — Personalization Engine

Purpose: Adapt prompt structure and content dynamically based on user context, profile, and interaction history.

Requirement Level Description
Context signals MUST Accept at minimum user role and session context as personalization inputs
Policy alignment MUST Personalization must not override compliance or safety constraints from L5
Signal transparency MUST Declare which signals influenced prompt adaptation in metadata
Sensitive data exclusion MUST Must not use PII, biometric, or health data as personalization signals without explicit consent
Reinforcement input SHOULD Support feedback signals from L7 to improve personalization over time
Profile portability MAY Support import/export of user profiles in a standard format

L4 — Attribution Ledger

Purpose: Track authorship, versioning, and revision history for every governed prompt using unique identifiers and hash functions.

Note: L4 is mandatory for Tier 2 and Tier 3 conformance.

Requirement Level Description
Unique prompt ID MUST Assign a globally unique identifier (UID) to every prompt
SHA-256 hash MUST Generate and store a SHA-256 hash of prompt content at each revision
Contributor attribution MUST Record the UID of every contributor per revision
Timestamp MUST Record an ISO 8601 timestamp for each revision
Revision chain MUST Maintain an ordered, append-only revision history
Immutability MUST Revision records must not be modifiable after creation
Blockchain anchoring MAY Optionally anchor hashes to a distributed ledger for legal-grade timestamping

Minimum metadata object per revision:

{
  "prompt_id": "upif-{uid}",
  "revision": 1,
  "contributor_uid": "user-{uid}",
  "timestamp": "2025-04-18T00:00:00Z",
  "content_hash": "sha256:{hash}",
  "modality": "text",
  "parent_revision": null
}

L5 — Compliance Firewall

Purpose: Screen and evaluate prompts against ethical, legal, and organizational policies before execution.

Requirement Level Description
Pre-inference evaluation MUST All policy checks must occur before the prompt reaches a model API
Policy rule format MUST Rules must be declarative and inspectable (not opaque ML classifiers alone)
Block/allow decision MUST Every prompt must receive an explicit pass or block decision
Block reason logging MUST Blocked prompts must record the triggering rule and reason
GDPR flag support SHOULD Support evaluation against GDPR-relevant content signals
COPPA flag support SHOULD Support evaluation against age-appropriate content signals
HIPAA flag support SHOULD Support evaluation against health information signals
Custom rule packs SHOULD Allow organizations to define and load custom policy rule sets
Override mechanism MAY Allow authorized roles to override a block with logged justification

L6 — Tone / Brand Governor

Purpose: Enforce stylistic, tonal, and brand-aligned communication consistency within the prompt structure.

Requirement Level Description
Tone profile definition MUST Support at least one named tone profile (e.g. formal, friendly, academic)
Pre-inference enforcement MUST Tone evaluation must occur before prompt execution, not post-output
Profile attribution MUST Record which tone profile was applied in prompt metadata
Conflict handling MUST Define behavior when prompt content conflicts with active tone profile
Multi-profile support SHOULD Support multiple simultaneous tone profiles with priority ordering
Brand vocabulary SHOULD Support allowlists and blocklists of brand-specific terminology
Profile inheritance MAY Allow tone profiles to inherit from parent profiles

L7 — Feedback Loop

Purpose: Capture and integrate feedback from users, evaluators, and system metrics to refine prompt structures over time.

Requirement Level Description
Feedback capture MUST Accept explicit feedback signals (e.g. accept, reject, revise) per prompt
Feedback attribution MUST Attribute each feedback signal to a contributor UID and timestamp
Feedback linkage MUST Link feedback records to the specific prompt UID and revision
Loop closure SHOULD Feed aggregated feedback signals back to L3 (Personalization Engine)
Metric tracking SHOULD Track quantitative metrics (e.g. acceptance rate, revision count) per prompt type
Audit trail MUST Feedback records must be append-only and non-modifiable
Aggregation API MAY Expose an API for downstream analytics over feedback data

Metadata Schema

All UPIF-conformant implementations must produce prompt metadata objects compatible with the following base schema. Additional fields may be added; defined fields must not be renamed or removed.

{
  "prompt_id": "string — globally unique identifier",
  "author_uid": "string — original author identifier",
  "timestamp": "string — ISO 8601 creation timestamp",
  "modality": "string — one of: text, image, audio, video, code",
  "intent_tags": ["array of strings — optional semantic tags"],
  "tone_profile": "string — active tone profile name",
  "compliance_flags": {
    "gdpr_safe": "boolean",
    "coppa_compliant": "boolean",
    "hipaa_safe": "boolean"
  },
  "revision_history": [
    {
      "rev": "integer — revision number",
      "contributor_uid": "string",
      "timestamp": "string — ISO 8601",
      "content_hash": "string — SHA-256 hash of prompt content at this revision"
    }
  ],
  "routing": {
    "source_modality": "string",
    "target_modality": "string",
    "router_decision": "string — pass or block"
  },
  "compliance_result": {
    "decision": "string — pass or block",
    "triggered_rule": "string or null",
    "evaluated_at": "string — ISO 8601"
  }
}

Claiming Conformance

Implementations claiming UPIF conformance must:

  1. Specify the conformance tier (Minimal, Partial, or Full)
  2. List the specific layers implemented
  3. Reference this document and version: UPIF Conformance Specification v1.0
  4. Link to the canonical specification: https://doi.org/10.5281/zenodo.15242693

Example conformance statement:

This implementation is Partial UPIF Conformant (v1.0), implementing layers L1, L4, and L5 in accordance with the UPIF Conformance Specification v1.0 (https://doi.org/10.5281/zenodo.15242693).


Terminology

Term Definition
MUST Absolute requirement. Non-conformance disqualifies the conformance claim for that layer.
SHOULD Strongly recommended. Non-conformance must be documented with justification.
MAY Optional. Implementations may include or omit at their discretion.
Prompt A structured input artifact governed by the UPIF lifecycle.
UID Unique identifier — any format that guarantees global uniqueness (UUID v4 recommended).
Governed prompt A prompt that has passed through at least one UPIF layer and carries UPIF metadata.

Versioning

This conformance specification follows semantic versioning. Breaking changes to MUST-level requirements increment the major version. Additive changes increment the minor version.

Version Date Notes
1.0 April 18, 2025 Initial release

License

This specification is released under the same license as the UPIF repository. See LICENSE for details.


UPIF is developed and maintained by Roshan George Thomas / XWHYZ. For questions or conformance inquiries, open an issue at https://github.com/XwhyZ-WHYLD/upif-specs