AI systems should be engineered as operational systems.
This repository distinguishes between hypotheses, implementations, and evidence. Not every idea here is mature. Every claim is versioned, evidence-graded, and open to revision as new data emerges.
The Human Stack is a living engineering reference for deploying, operating, and evaluating AI systems in production. It is not documentation for a single product. It is not a blog. It is a public research lab: a continuously revised body of engineering knowledge where claims become stronger as evidence accumulates.
- VISION.md — What this project is and why it exists.
- CONSTITUTION.md — The editorial rules that govern every page.
- EVIDENCE.md — How we grade claims and what each level means.
- manual/START-HERE.md — Guided path through the chapters.
| Directory | Purpose |
|---|---|
manual/ |
The living engineering reference — evidence-graded, versioned, continuously revised. |
templates/ |
Standard formats for chapters, reviews, benchmarks, and postmortems. |
references/ |
Links to reference implementations (Animus, drift-monitor, memboot, etc.). |
EVIDENCE.md |
The evidence framework that governs every claim in the manual. |
MATURITY.md |
The maturity model for ideas and capabilities. |
ROADMAP.md |
Planned work and evidence targets. |
- Engineers building AI systems: Start with manual/30-operations/.
- Architects making tradeoffs: Start with manual/40-architecture/.
- Hiring managers evaluating senior engineers: Start with manual/70-engineering-reviews/.
- Researchers: Start with manual/80-research/.
| Metric | Current | Target |
|---|---|---|
| Hypotheses (E0–E1) | 3 | ≤ 20 |
| Benchmarked claims (E3) | 4 | Increasing |
| Production case studies (E4) | 3 | Increasing |
| Independent reproductions (E5) | 0 | Increasing |
| Retired or superseded claims | 0 | Increasing |
We optimize for validated knowledge, not volume.
Evidence is graded on three axes:
- Evidence Level (E0–E5): How much empirical support exists?
- Confidence (Low/Medium/High): How strongly do we currently believe it?
- Scope (Narrow/Moderate/Broad): How broadly do we think it applies?
Claims can be downgraded as well as upgraded. Public downgrades are a feature, not a bug.
The Human Stack v0.8 — July 2026
See ROADMAP.md for planned revisions.
Last updated: 2026-07-07