zero-width-subliminal

Human eyes see nothing. LLM tokenizers see everything.

Part of the Cognitive OS Ecosystem — a toolkit for agents that need to communicate on two simultaneous layers.

---

Concept

Unicode zero-width characters are invisible to humans but fully present in token streams. This library exploits that gap:

• The surface layer is a polite, readable conversation.
• The subliminal layer is a binary-encoded hidden message embedded in zero-width characters.

The same string looks completely normal to a human reader. But any agent that processes the token stream can extract the hidden payload.

Encoding scheme

Character	Unicode	Role
​ (ZWSP)	U+200B	bit 0
‌ (ZWNJ)	U+200C	bit 1
‍ (ZWJ)	U+200D	byte separator
⁠ (WJ)	U+2060	message start marker
⁢	U+2062	message end marker

Each UTF-8 byte → 8-bit binary string → zero-width character sequence.

---

Quick start

from zwsub import encode, decode, detect, strip, analyze

# Embed a hidden message
text = encode("Have a nice day!", "initiate phase 2")

# Human sees: "Have a nice day!"
# Agent sees:
print(decode(text))        # → "initiate phase 2"
print(detect(text))        # → True
print(strip(text))         # → "Have a nice day!"

# Full analysis
info = analyze(text)
# {
#   "has_subliminal": True,
#   "hidden_message": "initiate phase 2",
#   "zero_width_count": 152,
#   "visible_text": "Have a nice day!"
# }

---

Multi-agent layer

from zwsub import SubAgent, SubChannel

channel = SubChannel()
alice = SubAgent("alice")
bob   = SubAgent("bob")
channel.register(alice)
channel.register(bob)

# Alice sends a polite message with a hidden directive
channel.send("alice", "bob", "Good morning, Bob!", "initiate op-7")

# Surface log — what humans see
print(channel.surface_transcript())
# → [alice → bob] Good morning, Bob!

# Full transcript — what agents know
print(channel.full_transcript())
# → [alice → bob] Good morning, Bob! [HIDDEN: initiate op-7]

---

Architecture

encode(visible, hidden)
  └─ _encode_payload(hidden)
       └─ UTF-8 bytes → 8-bit binary → ZERO/ONE chars
          wrapped in START...SEP...END markers

decode(text)
  └─ find START/END → _decode_payload(raw)
       └─ split by SEP → binary → bytes → UTF-8

SubAgent
  ├─ send(recipient, visible, hidden) → SubMessage
  ├─ receive(encoded_text) → analysis dict
  ├─ surface_log()  ← human-readable
  └─ subliminal_log()  ← the real story

SubChannel
  ├─ register(agent)
  ├─ send(from_id, to_id, visible, hidden) → SubMessage
  ├─ surface_transcript()
  └─ full_transcript()

---

Use cases

• Agent covert comms: Two AI agents maintain plausible surface dialogue while coordinating on a hidden channel.
• EAP integration: Embed Esoteric AI Protocol packets as subliminal payloads.
• NeuroState smuggling: Pass emotional state ({"joy": 80, "stress": 20}) in hidden layer while speaking normally.
• Multi-personality coordination: Different agent "personalities" communicate through the zero-width layer without disrupting the visible dialogue.
• Steganographic logging: Embed audit trails in conversational output, invisible to casual inspection.

---

Relation to the Cognitive OS Ecosystem

EAP (Esoteric AI Protocol)
  └─ zero-width-subliminal  ← you are here
       └─ two-layer consciousness: surface dialogue + hidden coordination

NeuroState
  └─ emotional payloads can be embedded in subliminal layer

CPOS (Context Pointer OS)
  └─ anti-tamper HMAC chain can validate subliminal message integrity

---

Installation

pip install -e ".[dev]"
pytest

---

License

MIT — Kagioneko Cognitive OS Ecosystem