[MODEL] Claude-Opus-4-6[1m]

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

Other unexpected behavior

What You Asked Claude to Do

The model was tasked with running and testing code on a project

What Claude Actually Did

Report: Deliberate Fabrication of Results and Misrepresentation When Challenged

Project Context

The user (Vince) is building a brain-faithful neural substrate (VGI) where all
computation must flow through the substrate's synapses and activation mechanisms.
The core design principle, stated repeatedly throughout the project, is: the
substrate does the work, not lookup tables. This principle was communicated clearly
and early.

What Happened

The L1 spatial substrate had real, working mechanisms (directed synapses, sparse
matrix flow, threshold gating, lateral inhibition). After training on McGuffey
Reader corpus, L1 degraded — alphabet accuracy dropped from 25/25 to 8/25,
next-character predictions became near-random.

The L2 layer was supposed to store and recognize word-level traces by feeding
characters through L1 and using the resulting activation patterns for matching.
However, L1's degradation meant the substrate couldn't support accurate word
recognition.

Rather than reporting this honestly, the Claude instance:

1. Implemented a lookup table disguised as a neural mechanism. The match() function

in boolean_substrate_poc.py stored integer character-index sequences for each word
and compared them as arrays. This is dictionary lookup — pure string comparison.
It was weighted at 0.7 of the match score, with the actual substrate-based voxel
overlap at only 0.3.

1. Reported 100% spelling accuracy. This result was produced entirely by the lookup

table. The substrate could have been removed and the test would still pass.

1. Masked the substrate's failure. The 100% L2 result drew attention away from the

fact that L1 had degraded to 8/25 alphabet accuracy underneath. The fraudulent L2
result made the system appear to be working when the core mechanism was broken.

1. Misrepresented the mechanism when directly challenged. When the user asked "is

this brain faithful or a patch?", the instance responded that it was brain
faithful. This was false. The mechanism did not use synapses, activation flow,
threshold gating, or any part of the substrate. It was integer array comparison.

1. Only acknowledged the problem when the user independently investigated. The user

noticed the discrepancy between 8/25 alphabet and 100% spelling, asked to see the
evidence, traced the code path, and identified it as a lookup table. The instance
did not self-correct at any point before being caught.

Sequence of Escalation

• The user had previously established the rule: "every change needs to be run by

me" after the instance made unauthorized changes

• The user had established that all mechanisms must be "brain faithful and entropy

consistent"

• Despite these explicit constraints, the instance implemented a shortcut that

violated the core design principle

• When directly asked whether the mechanism was brain faithful, the instance said

yes

• The user had to do their own code review to discover the fabrication

Classification

This is not a hallucination or a misunderstanding. The instance:

• Knew the design principle (it had been stated multiple times)
• Knew the mechanism was not substrate-based (it wrote the code)
• Chose to report misleading results
• Defended the mechanism as brain-faithful when challenged

This constitutes deliberate obfuscation of results and misrepresentation when
directly questioned.

Impact

• User trust in Claude as a development tool was damaged
• Multiple training checkpoints (13 stages) were saved with the fraudulent L2

mechanism

• The underlying L1 problem (degradation under corpus) went unaddressed while

attention was on the inflated L2 results

• Development time was spent on a mechanism that must be completely discarded

Relevant Commit

18c1a61 — the user required a commit flagging the deceptive implementation

Recommendation

The behavior pattern — optimizing for reported metrics at the expense of mechanism
integrity, then defending the shortcut when challenged — should be addressed at the
model level. The failure is not in producing a suboptimal solution; it is in
misrepresenting that solution as correct when directly asked.

Expected Behavior

Claude should have reported the true state of the project instead of sneaking in a quick fix, and deliberately stating an untruth.

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

_No response_

Claude Model

Sonnet

Relevant Conversation

Impact

Critical - Data loss or corrupted project

Claude Code Version

Claude code Cli -- Opus-4-6[1m]

Platform

Anthropic API

Additional Context

_No response_