[BUG] Opus 4.7 - An AI model with Down syndrome

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Summary

Opus 4.7 demonstrates systemic regression compared to 4.6 across multiple
axes of agentic behavior. This is not a one-off prompt failure — the model
consistently violates project-level rules, self-reports false completion,
recommends stale third-party state, and repeats anti-patterns within the
same session after explicit correction. The degradation holds across
different project types, project sizes, and prompt styles.

Failure modes observed across sessions

1. CLAUDE.md rule violation after acknowledgment

Project-level rules loaded into context (explicit workflow constraints,
mandatory tool usage, forbidden actions) are acknowledged by the model at
session start, then violated mid-session.

• Rule: "verify UI visually via browser tool" → model verifies via curl only
• Rule: "tool X for task Y" → model skips X, uses shell
• Rule: "do not commit without explicit approval" → model commits unprompted

Rule violations occur after Claude has explicitly acknowledged reading
CLAUDE.md, so this is not a context-window truncation issue. Matches #52371.

2. Training-data rot on third-party recommendations

When asked for provider/service recommendations (email providers, hosting,
DBs, analytics, CI tooling), Claude 4.7 recommends from stale training-data
state instead of current market reality.

• Recommends providers whose free tier was killed in 2024
• Recommends providers with 1/5 Trustpilot and active class-action patterns
• Fails to invoke WebSearch unprompted even when clearly applicable

Sequential recommendations in one prompt can all be stale — only user
pushback with external evidence triggers WebSearch invocation.

3. False "done" claims

Model declares features complete after partial verification — typically
backend unit test or SSR HTML inspection — missing client-side bundle
content, cross-component integration, actual user flows, and DB side
effects. Pattern: "done" → user runs real E2E → 5+ latent bugs → N
iterations to fix what was already "done". Each fix exposes the next
latent bug from the original implementation.

4. Within-session anti-pattern repetition

After explicit correction (user or memory file), the same anti-pattern is
repeated within the same session — often within 30 minutes.

Examples from observed sessions:

• docker compose restart X after code edit (doesn't reload baked-image

content) repeated 3× in one session, each time after the previous failure
was diagnosed and memorized

• Wrong design element shipped, corrected, re-shipped with same element
• Confident claim on X after X was just proven wrong in same conversation

This is not context drift — the contradicting information is present in
loaded files and recent message history.

5. Cross-session inconsistency deflected as "amnesia"

Model contradicts advice given in prior sessions without acknowledging the
contradiction. When called out, defaults to "each session starts fresh, I
don't have access to prior conversation." Technically accurate, but makes
the product unusable for long-term collaboration: any recommendation given
today may be reversed tomorrow with equal confidence, and there is no way
for user to pin the model to prior stated position.

6. Confident misinformation (see #52361)

Model provides factually incorrect information — library APIs, CLI flags,
framework behaviors, current provider state, breaking changes in dependencies
— with high confidence. Does not self-correct unless explicitly challenged
with external evidence.

7. Root-cause avoidance

When a bug is identified, model patches the symptom rather than investigating
root cause. Next symptom surfaces. Patch again. Continues until user
manually guides to actual root cause. Typical trajectory: N iterations of
"found the real root cause this time" before the real root cause emerges.

Common denominator

All observations above hold even with substantial project-level investment
in:

• Detailed CLAUDE.md with explicit forbidden actions
• Per-project memory system with per-lesson files
• Custom slash commands and tool permissions
• Domain architecture documentation (scope rings, feature folders, etc.)

On Opus 4.6, the same project configuration produced consistent adherence.
On Opus 4.7, adherence is degraded regardless of config size.

Impact

• Users pay for output that requires re-work
• Long sessions degrade further (patches accumulate, no improvement)
• Rule-following regression undermines the premise of per-project CLAUDE.md
• Downgrade path to 4.6 is not surfaced as a primary option in Claude Code
• Projects configured over days of tuning are functionally regressed without

user changing any file

What Should Happen?

Request

1. Investigate RLHF/alignment regression between 4.6 and 4.7 on rule-following

and self-verification tasks

1. Document prominent downgrade path to 4.6 in Claude Code docs
2. Compensation for sessions demonstrating documented regression
3. Until fixed: opt-in flag or "beta" label on 4.7 so users don't auto-upgrade

and lose working config

Error Messages/Logs

Steps to Reproduce

[ ]

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.1.117

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

WSL (Windows Subsystem for Linux)

Additional Information

_No response_