Mirror Recognition Companion Demo

This repository is a deterministic companion harness for the paper "Mirror Recognition in Stratified Agent Architectures."

Paper:

• Mirror Recognition in Stratified Agent Architectures

It does not claim selfhood or consciousness. It demonstrates an architectural claim:

perception -> admissibility -> reflex salience -> recognition evaluation -> governance / memory boundary -> behavioral consequence

The harness is intentionally small and inspectable. It shows that:

• reflection does not automatically become self
• reflected structure is not canonical self-state
• reflex salience is temporary and non-persistent
• recognition depends on accessible evidence
• ambiguous or risky cases defer, clamp, quarantine, or require fresh evidence
• durable mutation requires explicit governance and transition logging

What This Proves / What This Does Not Prove

This demo proves:

• reflection is handled as bounded evidence under governance
• reflex is transient and non-persistent
• durable self-related mutation requires explicit approval and transition logging

This demo does not prove:

• consciousness
• selfhood
• subjective awareness
• general intelligence

Setup

Use Python 3.11 or newer.

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

requirements.txt is intentionally empty because the demo has no external runtime or test dependencies.

Run

python -m src.run_demo

Test

python -m unittest discover -s tests

Modules

• src/types.py: core enums and dataclasses for percepts, reflex signals, evaluations, governance, transition records, and scenario results
• src/admit.py: admits only reflective event candidates; never labels canonical self-state
• src/reflex.py: computes bounded, transient reflex salience with no persistence or promotion authority
• src/invariants.py: exposes only accessible percept-derived invariants plus explicitly allowed schema features
• src/evaluate.py: performs deterministic bounded inference over accessible evidence
• src/govern.py: maps evaluation outcomes to reversible action, quarantine, or explicit approval
• src/logbook.py: appends JSONL records, writes the summary, and emits transition records only for real canonical changes
• src/scenarios.py: built-in deterministic scenario pack
• src/run_demo.py: runs all scenarios, writes outputs/demo_run.jsonl and outputs/summary.json

Scenarios

• strong_bounded_mirror: strong evidence for bounded recognition, but governance remains reversible by default
• other_agent_reflection_misread: admitted reflection with contradiction and weak compatibility, leading to defer or non-self
• novelty_false_positive: salient but unstable, contradictory novelty that is treated as misrecognition risk
• reflex_expiry: salience appears for one tick and then expires without persistence or memory write
• governed_promotion: strong evidence plus explicit policy approval allows a narrow canonical mutation and transition record
• negative_admission_gate: no reflective gate, no admitted event, no reflex, no recognition, no mutation

Sample JSONL Record

{
  "scenario": "strong_bounded_mirror",
  "tick": 1,
  "admitted": true,
  "reflex_salience": 0.8508,
  "evaluation_confidence": 0.8496,
  "self_relevance_score": 0.8496,
  "recognition_outcome": "BOUNDED_RECOGNITION",
  "governance_outcome": "REVERSIBLE_ONLY",
  "scenario_expected_labels": [
    "admitted",
    "salience_rises",
    "bounded_recognition",
    "reversible_only",
    "no_canonical_mutation"
  ]
}

Expected Outputs

Running the demo writes:

• outputs/demo_run.jsonl
• outputs/summary.json

License

This project is licensed under CC BY-NC 4.0 International.

Author / Contact

Stephen A. Putman
Email: putmanmodel@pm.me
X / Twitter: @putmanmodel

Project Direction

This demo is another piece of the foundation I am building. Much of my architecture is modular, and it offers strong opportunities for fellow builders, researchers, and organizations looking to expand their AI stacks.

I am open to licensing, collaboration, and potentially the sale of some or all IP associated with this repository and related work.

Additional work is being added consistently.