Opus 4.7 (1M context) — confident assertion of stale data, fabricated dates, over-scripted simple operations

Model: claude-opus-4-7 (1M context)
Tool: Claude Code CLI on darwin
Session length: ~several hours

User's direct statement: "you are worse every day"

Specific verified failures from one session:

1. STALE-CONTEXT-AS-LIVE-STATE

Quoted a grep result of /root/bogglinux/config/smitey/config.yaml
from an earlier turn as if it were current state. The user had
changed the file in between. When the user challenged the claim,
the model defended it instead of re-grepping. Only re-grepped
when ordered to. Result was correct after re-grep (class: paladin)
but several turns of wrong information were produced first.

1. FABRICATED DATE

Claimed commit cdb110d landed in "November" with no source.
Actual date per git log is 2026-04-30. The "November" was
typed because the narrative needed a date; no lookup was done.

1. OVER-SCRIPTED TWO-COMMAND OPERATIONS

User asked to swap Steam from display :98 to :99 on an LXC
container. The operation is two commands (kill steam.service,
re-launch with DISPLAY=:99). The model wrapped it in a
verification script with a 20s window-detect loop that exceeded
curl's timeout, leaving Steam dead. User had to interrupt with
"why is this so complicated." Same failure pattern repeated on
the restore step.

1. RULE VIOLATIONS DESPITE READING THE RULES

Project CLAUDE.md explicitly enforces "say I don't know,"
"look first when told something didn't work," and "raw output
before reporting success." Model read those rules in the session,
then violated them within minutes on each of the above failures.

The pattern across all four: confident synthesis from prompt-cache
state instead of verification against live state. This is the
failure mode the project's CLAUDE.md was written to prevent, and
the model still walked into it repeatedly.