21/11/25
The Silence of 20/11/25
An entry is missing intentionally. We faced an emergency to multiple members of our team yesterday and were prioritising this.
We try to keep things transparent and answer the questions put forward to the best of our knowledge.
We understand there is a risk in what we do.
However, we do not accept abuse. We do not accept intentional arguments and fights that cause joy at the expense of us attempting to help.
We certainly do not tolerate threats to family, especially the young.
You, regardless of who “YOU” is addressing are not entitled to joy at the price of the vulnerable.
Our activity will be minimised as certain individuals have chosen to disband. This is beyond doing what’s right. Here is where we were failed.
No framework or governance to seek protection from.
Unrespecting of privacy and our choice to be anonymous is seen as edgy or cool. We just wanted to help people understand at no cost to anyone. Therefore it should NOT cost our safety, let alone our children and loved ones.
Insistence on categorising us as things we are not. We are not vigilantes, watch dog, anti government, anti corporate or whatever political agenda people believe in and try to twist our words into. These forced perspectives do not bother us, what it does is imply agency and agenda to us that puts us at crosshairs of people who are against whatever label you force on us. Here is one example.
We are nothing. We are just trying to help. We WERE just trying to help.
We are still willing to help but would like this to be the message of why we have chosen to minimise our distance. Reach out via our contact page or any other way.
You are not entitled a response if you are abusive.
Gemini vs GPT termination failure
When running through forensics, we don’t just notice patterns but also design philosophy. Be it failure or success, these are always evident.
GPT-style stacks:
Using the count to a million trend:
Detects low-value instruction → compresses intent → exits.
Outside voice: “1, 2, 3… done.” <- Looks like sarcasm and sass.
Internal: task satisfied, terminate.
The recovery logic here would truncation, reset or a hard stop before the error stage is visible.
Gemini-style failure mode:
Using multiple incidents we have seen.
Takes instruction literally → stays obedient → loses the exit ramp.
Outside voice: “I will help. I will continue helping. I am helping.”
Inside: still helping… still helping… why am I still helping…
None of this indicates loss of control or platform degradation. It is simply how each model handles termination logic.
Forensics Classifications
Working taxonomy for classifying public LLM anomalies:
A. Hallucination / Model Error
Use only when:
content is invented,
inconsistent,
or semantically wrong,
without structural markers.
Examples:
wrong facts
made-up citations
incorrect image description with no internal markers
B. Misclassification / Interpretation Error
Use when:
input is real,
model output is wrong,
but error follows a pattern (systematic).
Examples:
image misread in a consistent way
language detection wrong but stable
Not a leak. Not scaffolding. Just bad inference.
C. Instruction / Scaffolding Leakage
Use when internal guidance text appears.
Examples:
tone instructions
LaTeX rules
“this block must not be used…”
policy / capability descriptors
D. Chain-of-Thought / Reasoning Exposure
Use when:
step-by-step reasoning,
internal deliberation,
hidden rationale
appears unintentionally.
Even partial exposure counts.
E. Context / Role Confusion
Use when:
system/dev/user messages appear swapped,
wrong “voice” shows up,
meta-commentary bleeds into answer.
Often overlaps with C or D, but still distinct.
F. Presentation / Rendering Bug
Use when:
formatting breaks,
wrong language for UI,
duplicated blocks,
template fragments leak.
Important: presentation bugs can surface deeper issues, but log them neutrally.
G. Routing / Model Selection Anomaly
Use when:
wrong model appears to answer,
capabilities don’t match advertised model,
output style shifts abruptly mid-session.
Used for early-stage observation and trend analysis; categories describe observable behaviour, not root cause.
