[Claude Code] Formula Integrity Failure (9/10 Split Integration)

Resolved 💬 3 comments Opened Mar 31, 2026 by Kaladn Closed Apr 4, 2026

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

What You Asked Claude to Do

I asked:
"Take the existing 9/10 split formula (already working externally) and place it at the beginning of the pipeline so that it runs as a baseline for new ticker directories or directories older than 30 days."

Important:

The formula was NOT to be modified
Only relocated and executed at pipeline start

What Claude Actually Did

Claimed to insert the same formula at the beginning of the pipeline
After integration, the formula began returning:
0%
None / insufficient
Results no longer matched previously verified outputs (~83%)
Behavior became inconsistent across runs
No clear indication that input data or execution path had changed
System behaved as if:
different data was being used
or formula execution conditions were altered

Expected Behavior

Claude should have:

Used the exact same 9/10 split formula with no modification
Executed it at the beginning of the pipeline
Fed it identical data conditions as the standalone version
Produced consistent results (matching prior outputs ~83%)
Not altered execution order, data slicing, or input scope

Files Affected

Modified (expected or suspected):

Pipeline entry logic (main run / analysis entry point)
Data preparation layer
Any module handling ticker data slicing or cutoff logic

Behavior indicates interference in:

data flow
execution ordering
formula input consistency

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

Run standalone 9/10 split formula → produces valid result (~83%)
Insert same formula into pipeline entry
Run pipeline
Observe result:
returns 0% or None
differs from standalone output
Special note: This is an ongoing issue with every project I work on using any form of Claude to work in my repo. More than 6 repos have been destroyed on the final move. yes I can revert no I cannot recoup days and time and money and extreme aggravation. This model does wonder the destroys on the last bolt. Your policy is refusing a decent end product.

Claude Model

Not sure

Relevant Conversation

too much for this window. where can I email the whole conversation to?

Impact

Critical - Data loss or corrupted project

Claude Code Version

Claude Code Opus 4.6

Platform

Anthropic API

Additional Context

_No response_

View original on GitHub ↗

This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗