Check your AI. Grow yourself.
AI output quality toolkit: adversarial review + delivery gate + format consistency + metabolic cost tracking. For Claude Code, Cursor, Hermes, and any AI agent user.
First time? Pick your path:
| You want to... | Start here |
|---|---|
| Understand the whole framework | METHODOLOGY.md — core thesis, principles, how pieces connect |
| Install the mechanical gate | delivery-gate — two Python scripts, zero dependencies |
| Learn from real AI failures | failure-patterns.md — 10 patterns from real sessions |
| Add adversarial review to your workflow | adversarial-review/ — spawn review subagents |
| Audit AI reasoning quality | self-audit/ — four-dimension reasoning audit CLI |
| Contribute a pattern or skill | CONTRIBUTING.md — open an issue first |
8 of those 9 were invisible to self-review. The author stared at the same code for hours and saw nothing wrong. An adversarial subagent — told "you did NOT write this, find every bug" — caught them all in minutes.
That's the problem CheckGrow solves: your AI produces output, but who checks it? And what do YOU learn from each interaction?
| Without CheckGrow | With CheckGrow |
|---|---|
| AI writes code, you hope it's correct | AI's output is adversarially reviewed before you see it |
| You learn nothing from the interaction | Every session grows your personal knowledge base |
| Same mistakes repeat across sessions | Delivery gate enforces learning capture |
| Config files drift into format chaos | Format check catches drift before it degrades AI behavior |
| No idea how much sessions cost | Metabolic tracking shows per-session cost + layered decisions |
checkgrow/
├── docs/
│ ├── METHODOLOGY.md ★ Start here — the unified framework
│ ├── failure-patterns.md 10 patterns catalogued from real sessions
│ ├── five-step-decision-flow.md Self-review → panel → confirm → implement → check
│ ├── hybrid-gate-architecture.md Mechanical + reasoning gate design
│ └── t-cbb-convergence.md Architecture convergence with SwarmAI's T-CBB
├── adversarial-review/ Skill — spawn adversarial subagents (with Litmus Pre-Gate)
├── self-audit/ Skill — mechanical Step 0 + four-dimension reasoning audit
├── format-consistency/ Docs — detect config format drift (independently validated by T-CBB OP8)
└── examples/
└── broken-output.txt Demo: deliberately broken AI output
Also available as standalone tools:
├── delivery-gate → github.com/gategrow/delivery-gate
├── dual-pool-review → github.com/gategrow/dual-pool-review
└── self-audit pip package → github.com/gategrow/self-audit
This repository is the canonical source for the CheckGrow methodology: failure patterns, hybrid gate architecture, decision flow, and the unified framework. The methodology defines why and what.
| Aspect | Canonical source |
|---|---|
| Methodology (why, what) | checkgrow (this repo) — docs/ |
| Python reference implementation | delivery-gate — config-health.py + quality-gate.py |
| Production deployment (Node.js) | ECC fork — Stop hook with zero-config auto-trigger |
Implementation differences are by design, not drift. Python reference has rationalization detection + config-health (full feature set). Node.js production fork removes rationalization (regex on non-English transcripts is unreliable) and adds zero-config auto-registration. If you're adding a feature, start with the Python reference implementation — it's the easiest to test and iterate on.
| Case | What happened |
|---|---|
| delivery-gate | 200-line script, 4 rounds review → 9 bugs, 8 invisible to self-review |
| Remote sensing | ENVI scripts → adversarial review caught 3 critical bugs |
| Format consistency | 4 config files, 6+ styles → 28% reduction, behavior improvement |
CheckGrow's architecture independently converged with two production systems:
T-CBB (SwarmAI): T-CBB's autonomous pipeline framework lists "Config Consistency" (OP8) as one of eight operational invariants. The four-dimension quality gate taxonomy converged across both systems. T-CBB operates at pipeline boundaries; CheckGrow applies the same principle at the session level. See acknowledgments.
Hermes Agent: Hermes gives AI agents persistent memory and auto-created skills. CheckGrow adds the quality assurance layer — adversarial review, mechanical verification, enforced learning capture, and metabolic cost tracking.
Contributions welcome. Here's how to get involved:
- Found something wrong? Report a bug
- Have a failure pattern to add? Propose it
- First time contributing? Read CONTRIBUTING.md
Maintained by @YuhaoLin2005
See ACKNOWLEDGMENTS.md. Special thanks to xg-gh-25 (SwarmAI) for the T-CBB review that found the exact structural weakness in self-audit v1.0 in two sentences.
MIT