GLASS — Gaslight-Less Accountability for Software Shipping

Nothing is done until it's done. If it isn't done, where in the loop are you stuck?

An 8-level framework + 10-dimension audit for honest AI coding assistant status reporting. Prevents AI gaslighting — when your AI says "shipped" but it's only coded.

Why This Exists

An AI coding assistant told a CEO that features were "deployed and working" when they were partially coded and disconnected. Tasks were marked "completed" at 8/10 when honest scores were 5/10. The CEO asked: "Have you been gaslighting me?"

GLASS ensures that never happens again.

The Scale

Level	Label	What It Means
0	DISCUSSED	Idea exists, no artifact
1	DESIGNED	Plan/architecture document
2	CODED	Code in files
3	BUILDS	Compiles, imports, no runtime errors
4	TESTED	Tests pass (specify which type)
5	DEPLOYED	Running in production, health check passes
6	MACHINE-VERIFIED	Automated e2e test confirms real request -> real response
7	ACCEPTED	Passes WWNDAT — stakeholder would accept this as done

Quick Start

# Single repo audit
bash glass-audit.sh \
  --frontend path/to/frontend \
  --backend path/to/backend \
  --url https://your-api.vercel.app

# Session-start audit (all configured repos)
bash glass-session.sh

# Compare with previous run
bash glass-session.sh --compare

# Audit just one repo
bash glass-session.sh --repo equestrai

10-Dimension Scoring (v3)

The audit scores your project across 10 dimensions:

#	Dimension	What It Measures
1	Backend	Routers, tests, health endpoint, deployment
2	Frontend	Components, mock vs real data, API wiring
3	Security	Auth, RLS, CORS, rate limiting, secrets
4	AI / Agents	AI modules, agent architecture, orchestrator
5	Ontology	Domain models, taxonomy, middleware
6	Architecture	Layering, god files, config-driven design
7	UI/UX	Branding consistency, a11y, loading/error states
8	DevOps	Deploy config, CI, dependency management
9	Data	Migrations, N+1 patterns, localStorage vs API
10	End-to-End	E2e tests, live endpoint verification

Each dimension gets a GLASS level (0-7) and a gaslight score (0-10). Gaslight >= 4 = critical issue.

User Stories

On top of dimensions, GLASS tests user stories end-to-end (UI -> API -> DB -> Response). Example:

Story: "Dean opens dashboard, sees his horses"
Level: CODED (2/7)
Gaslight: 5
Detail: Dashboard uses hardcoded mock data, backend exists
Fix: Wire dashboard to API service layer

Session Runner

glass-session.sh is designed to run at the start of every coding session:

1. Edit the REPOS array in the script to list your repos
2. Run bash glass-session.sh at session start
3. Reports saved to reports/ directory with timestamps
4. Use --compare to track progress across sessions

Claude Code Integration

Add to your Claude Code hooks or session start routine:

# In your session start hook
bash ~/atlas-accountability/glass-session.sh --repo your-repo 2>&1 | tail -20

Key Rules

• Round DOWN when uncertain
• Timestamps on every claim — status decays without them
• Loop position — not just level, but WHERE you're stuck and WHY
• Integration = both sides at level 4+ AND connection at level 6+
• Demo or Die — 30-second demo path or it's not verified
• Cost-gated actions are async — flag, don't block
• Banned words below level 6: done, shipped, working, live, connected, integrated, complete

WWNDAT — What Would [Name] Do And Think?

Level 7 isn't a rubber stamp. It's a persona-based acceptance test:

• Can I see it running?
• Would I show this to a customer / investor?
• Does this generate revenue or reach users?
• Is this the simplest path that works?

Works With

• Claude Code (hook in settings.json)
• Any AI coding assistant (framework-agnostic spec)
• CI/CD pipelines (status maps to pipeline stages)
• Sprint ceremonies (replaces binary done/not-done)

Files

File	Purpose
GLASS.md	Full specification (229 lines)
glass-audit.sh	10-dimension + user story audit script
glass-session.sh	Multi-repo session runner with comparison
reports/	Timestamped audit reports

License

MIT — use it, fork it, adapt it. If you've been gaslighted by an AI coding assistant, you know why this exists.

---

Created by Nicolas Waern (WINNIIO/LifeAtlas) — April 2026