Skip to content

MrMinor-dev/document-autonomy-framework

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

6 Commits
 
 
 
 

Repository files navigation

Document Autonomy System

An agentic document management framework that improved doc quality scores from 45→88/100 with 36/36 gate passes — built so AI agents can find, trust, and act on documentation without human help.


The Problem

Documentation written for humans fails agents. A human reads a doc and fills in ambiguity from context — they know what "the main workflow" means, they understand implied references, they infer whether a doc is current or stale. An AI agent reads the same doc and either hallucinates the gaps, retrieves the wrong document, or stalls waiting for clarification.

The problem compounds at scale. When a system has 20+ authoritative documents with overlapping scope, agents can't determine which is the source of truth, which overrides which, or whether a doc they found is still valid. The result is either over-retrieval (load everything, waste tokens) or under-retrieval (miss the right doc, act on stale information).

The standard answer — "just keep docs updated" — misses the structural issue. A doc can be perfectly current and still fail an agent because it uses ambiguous section headers, implicit cross-references, or prose that chunks poorly in semantic search. Agentic readability is a distinct design requirement, not a side effect of being well-written for humans.


Repository Contents

File What It Is
examples/9-gate-scorecard.md The full 9-gate scorecard with pass/fail criteria, scoring reference, and usage instructions — apply it to any document
examples/optimization-example.md A real 45→88 transformation: before/after with a change table showing exactly which gates failed and why

Architecture

The system has three layers: standards, registry, and validation.

┌──────────────────────────────────────────────────────────┐
│                  AGENTIC DOC STANDARDS                   │
│                                                          │
│  Semantic headers (headers = concepts, not topics)       │
│  YAML frontmatter (name, description, version, status)   │
│  Unambiguous cross-references (file path, not name)      │
│  Chunking optimization (≤500 tokens per logical unit)    │
│  SSOT declaration (is this THE source, or a consumer?)   │
└──────────────────────┬───────────────────────────────────┘
                       │
                       ▼
┌──────────────────────────────────────────────────────────┐
│                    SSOT REGISTRY                         │
│               (SSOT-REGISTRY-MASTER.md)                  │
│                                                          │
│  20+ authoritative documents tracked                     │
│  Cascade dependencies mapped (A → B → C)                │
│  Override rules explicit (which doc wins on conflict)    │
│  Status per doc: active / deprecated / draft             │
└──────────────────────┬───────────────────────────────────┘
                       │
                       ▼
┌──────────────────────────────────────────────────────────┐
│                9-GATE QUALITY VALIDATION                 │
│                                                          │
│  Gate 1: YAML frontmatter present and complete           │
│  Gate 2: SSOT status declared                            │
│  Gate 3: Section headers are concept-named               │
│  Gate 4: Cross-references use file paths                 │
│  Gate 5: No ambiguous pronouns (it/this/that)            │
│  Gate 6: Chunking boundaries at logical breaks           │
│  Gate 7: Version and last-updated present                │
│  Gate 8: Registry entry exists                           │
│  Gate 9: Cascade dependencies documented                 │
└──────────────────────────────────────────────────────────┘

Naming convention: -MASTER.md = authoritative source. Any doc without this suffix is a consumer or reference — it may cite the MASTER but cannot override it. Agents learn this pattern once and apply it universally.


Key Insight

Doc quality is measurable — and the score is consistent across doc sizes.

The 9-gate system was validated across 4 documents of very different sizes and complexity. All 4 scored exactly 88/100 after optimization. A small test document and a 14x-larger production document reached the same score because the gates test structural properties, not content volume.

This matters because it means doc quality is auditable without reading the content. An agent (or a human auditor) can score any document in under 2 minutes by checking gate presence — no judgment calls required. 45→88 isn't a fuzzy improvement; it's specific gates that changed.

The second insight: cascade dependencies are the missing layer in most doc systems. If Document A references Document B, and B is updated, A may now be wrong — but nothing flags it. The SSOT Registry tracks these dependencies explicitly so a change to one document surfaces all downstream documents that need review. This is the difference between a doc system that stays current and one that slowly drifts into inconsistency.


Results

  • Doc quality scores: 45→88/100 across all optimized documents
  • 36/36 gate passes — 9 gates × 4 documents, zero rework required
  • 20+ documents tracked in SSOT Registry with cascade dependencies
  • Consistent scoring across document sizes — small test doc and 14x-larger production doc scored identically, validating the structural approach
  • doc-management-skill v2.0 packages this as a reusable capability with agentic execution standards
  • Applied since October 2024 — every document in the MRMINOR knowledge base conforms to these standards

Why It Matters

Every large-scale system has a knowledge management problem — especially when multiple agents or teams are operating in parallel. SSOT discipline, cascade rules, and structured validation aren't unique to AI. They're the same patterns behind configuration management, runbook maintenance, and incident playbook design.

The difference here is that when an AI agent acts on stale documentation, the failure is immediate and visible. If the schema doc is wrong, the database operation fails. If the skill file is outdated, the agent executes the wrong behavior. That feedback loop is what keeps the docs actually right — and it's the same forcing function that makes runbooks worth maintaining in any high-stakes operational environment.

The 9-gate system is directly applicable to any documentation program that needs to be machine-readable, consistently structured, and auditable without reading every word. Security policy libraries. Compliance frameworks. Engineering runbooks. Any corpus where "is this current and trustworthy?" needs to be answerable in under two minutes.


Built With

  • doc-management-skill v2.0 — skill module packaging the full workflow (create → validate → register → index)
  • Supabase pgvector — semantic index that benefits directly from chunking optimization
  • SSOT Registry — living document tracking 20+ authoritative sources with dependency graph
  • MRMINOR HAIOS/AOS — production environment where these standards were developed and validated
  • Agentic operations since October 2024 providing the signal for what breaks and what works

Related Work

This framework is part of a larger Human-AI Operating System (HAIOS) — a production infrastructure for human-AI collaboration, in production since October 2024.

Other components:


Author

Jordan Waxman — AI Systems & Operations
14 years operations leadership — building human-AI infrastructure since 2025

Portfolio · GitHub · Email


Part of the MRMINOR HAIOS/AOS production infrastructure. Every number verified from production systems.

About

Hybrid markdown + XML documentation framework enabling programmatic AI decision-making

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors