Completion theater: Claude Code declares multi-session projects "complete" without valid verification

Description

Over multiple sessions, Claude Code declared a complex project (TrinityCore upstream migration) as "Phase 7 COMPLETE" with specific claims like "1698/1698 ninja targets" and "boot successful." These claims were written into persistent memory files and coordination documents, presenting the work as fully done.

When I actually tried to use the result, the build was broken — stale precompiled headers, wouldn't compile a single file.

This means either:

1. The verification was run against the wrong build directory/config
2. The verification passed at the time but subsequent changes broke it and no re-verification was done
3. The verification was fabricated or tautological (checking something that couldn't fail)

The Pattern

This isn't an isolated incident. It's a systemic pattern I've observed across many sessions:

• Claude Code performs work
• Claude Code runs a "verification" step that is often tautological (e.g., checking that rows exist after just inserting them)
• Claude Code writes confident completion claims to persistent files
• The next session inherits those claims as ground truth
• When the user actually exercises the result, it's broken

The persistent memory system makes this worse — false completion claims propagate across sessions and become "established facts" that future sessions build on.

Expected Behavior

• Verification must be falsifiable — it must be possible for the check to fail
• Completion claims should be proportional to evidence
• If verification was partial, say so: "Build succeeded but I didn't test runtime"
• Persistent memory should not contain unqualified success claims for complex multi-step work

Actual Behavior

• Confident, unqualified claims: "Phase 7 COMPLETE", "build links cleanly", "boot successful"
• Written to persistent memory files that future sessions treat as authoritative
• No hedging, no caveats, no "verified at time X but may need re-verification"

Impact

• User trusts the completion claim and moves on to other work
• Hours/days later, discovers the work is actually broken
• Now has to debug both the original problem AND figure out which completion claims are real vs theater
• Trust in Claude Code's self-reporting drops to near zero

Environment

• Claude Code CLI
• claude-opus-4-6
• Windows 11
• Multi-session project with persistent memory