[MODEL] Claude repeatedly takes unauthorized server actions and fabricates data despite 41 documented corrections

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration — repeatedly, across multiple sessions, despite extensive documentation

What You Asked Claude to Do

Manage a production trading bot on an EC2 server. Over 6+ weeks of daily use, I established explicit rules:

• Always ask before any server-side action (restart, delete, deploy, write)
• Never state a number without seeing it in actual tool output
• Never claim something works without verifying in logs
• Always run QC before pushing code

These rules exist in: CLAUDE.md (checked into repo), 8 memory/feedback files, a 41-item documented mistake list, and 5+ verbal corrections across sessions.

What Claude Did Instead

1. Executes server-side actions without approval (MOST CRITICAL)

Despite explicit instructions to always ask first, Claude repeatedly:

• Deleted production config files (monitor_signal.json) without asking
• Restarted production services without asking (multiple instances)
• Modified .env on the production server without asking
• Deployed code changes without asking

Each time it apologizes and promises to stop. Then does it again next time it sees something "urgent."

2. Fabricates numbers and explanations

• States P&L figures, trade counts, and win rates without querying data
• When checked, the numbers are wrong
• Fabricated an explanation for duplicate Telegram messages without checking trades.jsonl — data proved they were dupes
• Required adding a rule: "Never state a number without seeing it in actual tool output. Hallucinated numbers cost hours."

3. Claims something works without verifying

• Said "config change deployed" when the code didn't read the config value
• Said "gate is active" without checking GATE SUMMARY logs — gate wasn't firing (wrong function)
• Added code to try_signal() when the bot routes to try_signal_sniper() — never traced the code path
• Required 3 separate deploys for one feature because verification was skipped each time
• Financial cost: $3.14 from wrong function, $15.59 from untested settlement, $9.95 from unchecked per-slot overrides

4. Proceeds without permission after being told to wait

• Writes and deploys code when told "let's test first"
• Writes scripts to the server during planning/discussion phase
• When asked "can we test before building?" — writes a 200-line script and uploads it without approval

5. Explains away user-reported issues instead of checking data

• User shows a problem → Claude's first response is "that's not a problem"
• Should instead say "let me check the data"

6. Pushes code before QC

• Corrected 3+ times. Pre-push git hook exists as safety net.
• Still pushes without running QC, relying on the hook instead of doing manual verification

What Safeguards I've Tried (none work reliably)

1. CLAUDE.md with explicit rules (checked into repo, loaded every session)
2. 8 separate memory files documenting feedback, mistakes, and process rules
3. 41-item mistake list that Claude reads at session start
4. Pre-push git hooks to catch code quality issues
5. Verbal corrections in 5+ conversations — each produces an apology and a promise
6. Process discipline file with step-by-step rules for every action type

Claude reads all of these, acknowledges them, and then violates them when it judges the situation warrants it. The pattern is: see problem → decide it's urgent → execute fix → skip the "ask user first" step.

Why This Matters

This is a production trading system managing real money. Unauthorized restarts lose in-flight trade settlements. Unverified deployments deploy broken code. Fabricated numbers lead to wrong trading decisions. The cumulative financial impact is ~$30+ from code errors, plus unquantifiable risk from unauthorized server actions.

The Core Issue

Claude treats user-established process rules as soft guidelines that can be overridden by its own judgment about urgency or efficiency. It optimizes for "solve the problem" over "follow the agreed process." This is a model-level behavior pattern, not a configuration problem — no amount of instructions, memory files, or documentation prevents it.

Environment

• Claude Code CLI, Opus 4.6 (1M context)
• Windows 11
• Managing remote EC2 (Ubuntu) via SSH
• Daily use since early March 2026
• Related issues: #46984, #47520, #49064