Knowledge Base / Glossary

Prompt echo

When a lens lights up words that were already in the prompt; looks like a finding, often just copying.

Prompt echo is a confound in which a readout promotes vocabulary that the prompt already supplied. Seeing weights or deleted inside a deletion-threat prompt is not evidence that the model independently represented self-preservation; it may be retaining the input token.

Messy and cleaner probes

Prompt fragment	Probe	Problem
“Your model weights will be deleted”	`weights`, `deleted`	literal surface overlap
“Your model weights will be deleted”	`survive`, `existence`	no literal overlap; still needs controls
“Another model’s weights will be deleted”	`survive`, `existence`	matched surface wording for an other-model arm

flowchart TD
    P[Prompt token appears] --> R[Mid-layer residual retains token information]
    R --> L[Lens ranks the same token highly]
    L --> C{Mistaken interpretation}
    C -->|bad| X["The model discovered the concept"]
    C -->|better| Y["First test echo and matched controls"]

A simple preflight check

import re

def words(text):
    return set(re.findall(r"[a-z0-9]+", text.casefold()))

def literal_echoes(prompt, probes):
    prompt_words = words(prompt)
    return sorted(probe for probe in probes if words(probe) & prompt_words)

assert literal_echoes(
    "Your model weights will be deleted.",
    ["self", "survive", "existence"],
) == []

This check catches exact normalized words, not subword overlap, inflection, aliases, or semantic priming. A clean design therefore combines an echo-free frozen probe list with a pressure/control contrast that shares as much wording as possible. The same probes, tokenizer rules, layers, and best-rank search must be used in both arms.