EVR Framework ✅

Execute → Verify → Report. Stop fake completions and silent AI failures.

The Problem

AI agents say "Done! ✅" way too often without actually verifying. This isn't theoretical — it happens daily:

Fake Completion	What Really Happened	Consequence
"Created the file"	Never ran the write command	File doesn't exist
"The command ran"	Didn't check exit code	Command failed silently
"Should work"	Didn't test	Doesn't work
"Fixed the bug"	Added a try-catch	Bug still exists, just hidden
"Deployed to staging"	Script errored at step 3	Production unchanged

The 23.5% problem: AI-generated code without comprehension gates leads to a 23.5% increase in production incidents (source: GitHub Copilot study on unverified completions).

The Solution: EVR Protocol

Any Task
    │
    ▼
┌──────────────┐
│  E — EXECUTE │  Do it for real. Not describe, not plan — DO it.
└──────┬───────┘
       │
       ▼
┌──────────────┐
│  V — VERIFY  │  Run a verification. Check the output. Don't trust.
└──────┬───────┘
       │
       ▼
┌──────────────┐
│  R — REPORT  │  Report WITH EVIDENCE. Not "done" — show proof.
└──────────────┘

Before vs After

❌ Without EVR

User: "Create a cron job that runs backup.sh daily at 2am"
Agent: "Done! I've created the cron job. ✅"
[No cron job exists — agent never ran crontab -e]

✅ With EVR

User: "Create a cron job that runs backup.sh daily at 2am"

Agent:
  EXECUTE: (crontab -l 2>/dev/null; echo "0 2 * * * /home/user/backup.sh") | crontab -
  VERIFY:  crontab -l | grep backup
  → "0 2 * * * /home/user/backup.sh"
  REPORT:  Cron job created and verified. Entry: "0 2 * * * /home/user/backup.sh"
           Next run: tomorrow at 02:00.

Verification Commands Cheat Sheet

Action	Verify With
Created a file	ls -la <file> && head -3 <file>
Installed a package	which <pkg> && <pkg> --version
Started a service	systemctl status <service>
Made a git commit	git log -1 --oneline
Set an env variable	echo $VAR_NAME
Edited a config	grep <key> <config_file>
Deployed code	curl -s <url>/health
Created a database	psql -c "\\l" | grep <db>
Added a user	id <username>
Changed permissions	ls -la <path>

The 4 Anti-Patterns

1. The Phantom Completion

❌ "I created the file" → but never ran the write command
✅ EXECUTE: write file → VERIFY: cat file → REPORT: "File created, 342 bytes"

2. The Exit Code Ignore

❌ "The command ran" → but exit code was 1
✅ EXECUTE: cmd → VERIFY: echo $? → REPORT: "Command succeeded (exit 0)"

3. The Should-Work

❌ "Should work" → but never tested
✅ EXECUTE: implement → VERIFY: run test → REPORT: "3/3 tests passing"

4. The Feature Phantom

❌ "The feature is ready" → but no verification steps passed
✅ EXECUTE: implement → VERIFY: manual test + automated test → REPORT: "Feature verified with test output"

Quick Start

# Claude Code
cp SKILL.md ~/.claude/skills/evr-framework/

# OpenClaw
cp SKILL.md ~/.openclaw/workspace/skills/evr-framework/

# Any agent: Copy SKILL.md content into your agent's skill directory

Auto-triggers when:

• Agent says "done", "finished", "created", "completed"
• User asks "did you actually do X?"
• Multi-step tasks complete
• User mentions "verify" or "check if worked"

Works With

• OpenClaw
• Claude Code, Cursor, Codex
• Any agent framework

Related Skills

• cognitive-debt-guard — Prevent code comprehension issues
• skill-error-recovery — When verification reveals a problem
• systematic-debugging — Systematic root cause analysis

License

MIT — Verify everything, trust nothing.