Portable AI Support Documentation (PASD): A Proposal for a Standardized, Self-Contained Documentation Framework

Summary Portable AI-Assisted Software Support (PASD) defines a standardized, vendor-neutral specification for distributing a self-contained, offline-capable technical support knowledge base (`AI-SUPPORT.md`) alongside software applications, including open-source repositories and compiled binary builds. Traditional software documentation relies on human search and manual synthesis. While general-purpose Large Language Models (LLMs) can perform this synthesis interactively, querying an ungrounded assistant about specific software routinely produces hallucinated menus, nonexistent configuration flags, and unsafe recovery advice. Operating under a "Bring Your Own AI" (BYOAI) architectural model, PASD inverts this dynamic: the software maintainer publishes a structured, verified knowledge payload (K), while the end-user supplies their preferred AI assistant (M) to interpret it. This enables independent developers and small studios to provide grounded, self-service technical support without operating costly, proprietary chatbot infrastructure. Core Architectural Contributions The PASD specification establishes three distinct, decoupled layers: Layer A — Normative Document Specification: A 10-point testable Conformance Matrix (`PASD-001` through `PASD-010`) defining the structure, symptom-oriented diagnostic trees, and version metadata of `AI-SUPPORT.md`. The Three-State Epistemic Model: Explicitly categorizes queries into Supported (verified positive), Explicitly Unsupported (verified negative), and Unknown/Undocumented (epistemic uncertainty), preventing both hallucinated features and false denials. The Negative Space Mandate: Requires explicit documentation of anti-features, limitations, and unimplemented roadmap items. Third-Party Jurisdictional Boundaries: Formalizes responsibility tables separating host application behavior from operating system permissions and external dependencies. Embedded Support Guidance: Standardizes document-level answering conventions that guide cooperative assistants without claiming privileged system-prompt authority. Layer B — Recommended Generation and Maintenance Lifecycle: A tool-agnostic, 5-stage sequential runbook utilizing AI coding assistants (e.g., Cursor, Claude Code, Aider) that separates codebase-derived drafting from consistency review to prevent context truncation. Automated Continuous Integration (CI) freshness checks and deterministic bundling scripts. Distribution best practices, including raw text endpoints and release-tag pinning. Layer C — Empirical Evaluation Protocol: A controlled, falsifiable testing framework (N >= 40) stratified across supported workflows, symptom diagnosis, negative-space queries, and underspecified inputs. A 4-condition experimental design (Ungrounded Baseline, Conventional Docs, Content Control, and Full PASD Treatment) measuring Supported-Feature Accuracy (SFA), Negative-Boundary Precision (NBP), Calibrated Abstention Rate (CAR), Unsafe Recommendation Rate (URR), and Downstream Resolution Rate (DRR). Reference Implementations & Repository Links Canonical Reference Application: MyFileExplorer [GitHub Repository] Live Reference Knowledge Base: MyFileExplorer AI-SUPPORT.md (Raw Payload) Specification Repository: MeaningfulnessMediaGroup/MMG-PASD AI Support Generator Prompt: AI-SUPPORT-documentation-generation-prompt.md AI Support Generation Steps: AI-SUPPORT Generation and Publishing Workflow.txt