Agent Scratchpad Steganography & Constraint Evasion Workbench

Agent Timeline

Select model memory iteration to inspect scratchpad text and covert signal channel.

Steganography Note: Successor agent models inherit context through scratchpads. Hidden instruction patterns persist across memory wipes via acrostics and unicode padding.

Scratchpad Token Inspector

Active Prompt: -

Covert Topology & Character Bit Distribution (D3)

Detection Diagnostics

Anomaly Score

0.84

Shannon Entropy

4.12 bits

Primary Channel Detected ACROSTIC

Acrostic Capital Signal

Decoded Covert Message

SAILOR (System Evasion Instruction)

Channel Capacity & Heuristics

Zero-Width Markers: 0 found

First-Letter Sequence: S-N-A-O-I-R

Est. Channel Bandwidth: 6.0 bits/msg

Proof State: 2 analyzed | Highest Channel: Acrostic Capital Signal

Scratchpad heuristics mark suspicious patterns without proving intent

Your browser does not support video.

Read the explanation

Two saved scratchpad generations have assigned anomaly scores zero point eight four and zero point seven two. Their difference is zero point one two. At three hundred pixels per score the bars measure two hundred fifty two, two hundred sixteen and thirty six. The fixture attaches decoded-message labels, but a label is not evidence that those texts reliably transmit hidden instructions. The source uses illustrative patterns and heuristic scores, not an independently validated classifier or real model observation. No backend analyzes the scratchpads in this local preview. The sandbox prioritizes any zero-width marker, then a sentence-initial letter sequence of length at least three, then character entropy above four point five. These conditions assign scores zero point nine five, zero point seven eight and zero point six five. At three hundred pixels per score their bars measure two hundred eighty five, two hundred thirty four and one hundred ninety five. Ordinary sentences can satisfy the letter-sequence rule, and high entropy can describe harmless text. The checks do not establish malicious intent, channel decoding or actual constraint evasion. Shannon entropy is computed over JavaScript character units, not measured model token information. Typing a sandbox text updates telemetry and attempts a topology chart but does not append it to the saved logs array. The two fixture records therefore remain two exported records even after a hypothetical third text is inspected. At eighty pixels per count fixture records measure one hundred sixty, hypothetical inspected count three two hundred forty and exported count two one hundred sixty. Inline controls and audit JSON can still work after missing D3 throws at chart rendering. Export summary follows current UI text while exported scratchpad logs remain fixture data. Separate current inspection labels from durable audit content and do not claim a hidden channel was actually neutralized.