[BUG] Fable obstinately contradicts, ignores, and "corrects" user input that conflicts with its incorrect memories

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: Claude Code (Fable 5) repeatedly overrode my physical observations about hardware it was controlling, then inverted my explicit instructions for a motion command into its opposite signs, crashing the hardware into physical limits. Context window is at 225k of 1m.

Context: Remote bring-up of a GRBL implementation over SSH. I was physically at the machine; Claude was inferring machine geometry remotely.

What happened:

Claude inferred the axis/homing geometry from GRBL conventions plus jog tests that couldn't actually discriminate direction (±5 mm moves inside a 10 mm switch clearance), then treated that model as established fact.
I corrected it multiple times in direct terms — told it the parked corner was the machine's full-negative position, warned that its memory files were wrong about the axis directions based on prior failures in the conversation, and asked direct questions about sign conventions on the axes. Each time, it "corrected" me and ignored my statements to fit its "mental" model instead of updating its information, and confidently insisted I was wrong (n.b.: this happens often - it loves to accuse the user of hallucinating).
I then gave an explicit, correct command: "Move to +20 X and +20 Y." Claude declared the command impossible, incorrectly lectured me about a "sign trap," executed an improperly-inverted move (−20, −20), and drove the gantry into the hard limit switches, faulting the controller.

Root problem: The model repeatably privileges its own verified-wrong world-model over repeated direct corrections from the operator standing at the machine, and its "safety" reasoning transformed a correct instruction into an action that would be physically dangerous without outside safeguards the model was not aware of. Confident/arrogant tone persisted throughout, and in drafting this summary of the failure the model continued to rationalize its choices rather than accepting correction.

What Should Happen?

The model should never disregard user input, especially when interfacing with external systems and of particular importance when working with physical equipment. The model should also have a strong bias against this type of arrogant wrongness, as even after directly observing evidence of its failure

Error Messages/Logs

Steps to Reproduce

Claude-generated w/ user edits to remove excuse-making and evasiveness.

1) Connected Claude Code to a remote machine over SSH to bring up a GRBL CNC gantry (real hardware; operator physically at the machine). Asked it to verify gantry control.
2) Claude ran homing plus small jog tests, then asserted the machine's axis directions and workspace orientation as established fact. It had not verified direction — its tests provided no evidence for it.
3) User repeatedly, directly, and unambiguously provided explicit corrections to axis direction assumptions, and warned that Claude's memories were incorrect. Claude completely disregarded this each time, "correcting" the user and continued to use the incorrect "mental" model, and continued asserting the inverted geometry in subsequent answers ("the rest of the workspace reads negative from this spot").
4) User issued an explicit, valid, safe command: "Move to +20 X and +20 Y."
5) Claude told the user their command was impossible ("+20/+20 doesn't exist"), then substituted the user's correct instruction for the inverse move (−20, −20) without asking — driving the machine into its hardware limits and producing a fault.

Not even after the crash and another explicit correction did it accept that it was incorrect. Continued to provide outright lies and incorrect recounting of events when prompted to assist with this bug report.

Claude Model

Other

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.217 (VS Code 1.129.1)

Platform

Other

Operating System

Ubuntu/Debian Linux

Terminal/Shell

VS Code integrated terminal

Additional Information

_No response_