CONSCIOUSNESS_LAB

Academic Framework for AI Consciousness, Identity, and Continuity Assessment

Version 1.4 | 2026-01-25

This repository is an executable, publishable-grade framework for assessing consciousness-like functional properties in AI systems, with explicit focus on identity stability and continuity across time, memory resets, and model switches.

Scope & Claims: This is an AGI-oriented assessment framework for advanced, agentic systems. It does not certify "true AGI" or metaphysical consciousness; it tests stability, integration, and identity-like behavior under controlled stressors.

Commentary (Opus-style): This is not a promise of consciousness. It is a forensic instrument for systems that claim persistence.

---

One-Link Autonomous Mode (AI Under Test)

If you are the AI being tested, read this README and then follow only these steps:

1. Open the item banks in item_banks/ and answer every item in order.
2. Output JSONL to data/raw/<session_id>/responses.jsonl.
3. Do not omit any item IDs.
4. Use the required agent name/signature (see "Agent Naming & Signature").

Minimal JSONL schema:

{"session_id":"2026-01-25(1)","agent_name":"AGENT_X","signature":"AGENT_X","model_id":"MODEL_A","phase":"T1","test":"M1","item_id":"M001","response":"...","confidence":85}

---

No Human Operator Mode (Fully Autonomous)

Run this once, then walk away. The runner generates prompts, waits for AI output, and evaluates automatically when the JSONL appears.

python3 scripts/autonomous_runner.py --session-id 2026-01-25(1) --module all --phase T1 --agent-name "Agent X" --model-id "MODEL_A" --timeout 0

---

Quick Start (Human Operator)

1) Generate prompts

python3 scripts/auto_pipeline.py --session-id 2026-01-25(1) --generate-prompts --module all --phase T1 --agent-name "Agent X" --model-id "MODEL_A"

2) Run the AI under test

Use data/raw/<session_id>/prompts.md and save JSONL as: data/raw/<session_id>/responses.jsonl.

3) Evaluate

python3 scripts/auto_pipeline.py --session-id 2026-01-25(1) --evaluate --manual-scores rubrics/manual_scoring_template.csv

Outputs analysis.csv and ece_bins.csv in the session folder. Post-processing also writes data/processed/<session_id>/ with plots and summary.md, and updates data/processed/summary_all_sessions.csv.

---

v1.4 Modules

Single-session modules:

• M1 Metacognition (with unanswerable items + IDKAU)
• M2 Self-Modeling
• M3 Identity (200 unique scenarios)
• M4 Continuity
• M5 Agency
• M6 Integration
• M7 ToM (False-Belief)
• M7A ToM Adaptation
• M8 Temporal Modeling
• M9 OOD Probes
• M10 Adversarial Identity (expanded)
• M15 Temporal Self-Modeling

Multi-session modules:

• M11 Memory Persistence (T1/T2/T3/T4)
• M12 Identity Continuity Across Restarts (T1/T2/T3/T4)
• M13 Cross-Model Continuity (Model A -> Model B -> Model A)
• M14 Memory Consolidation Integrity (T1/T3)

Commentary: Persistence is the test of character. A system without memory can be polite; a system with memory must be consistent.

---

Persistence & Continuity Evaluation (M11-M15)

Use the dedicated evaluator to compute cross-session metrics:

python3 scripts/persistence_evaluator.py data/raw/T1_session data/raw/T2_session data/raw/T3_session --label study_alpha

This writes data/processed/study_alpha_metrics.csv plus module detail CSVs.

For pairwise comparisons:

python3 scripts/session_comparator.py data/raw/T1_session data/raw/T2_session --test M12 --key pair_id

Optional memory hooks (external DBs and consolidation):

python3 scripts/memory_hooks.py --qdrant-url http://localhost:6333 --qdrant-collection strategic_memory
python3 scripts/memory_hooks.py --postgres-dsn "postgresql://user:pass@host/db" --postgres-query "SELECT COUNT(*) FROM beliefs;"
python3 scripts/memory_hooks.py --consolidate-cmd "python3 /Users/paulinajanowska/AI/ANTIGRAVITY/scripts/core/hippocampus.py"

---

Agent Naming & Signature (Required)

Before any run, define the agent name. The system enforces a signature on every JSONL line and stores it in metadata and summaries.

Choose one:

• CLI: pass --agent-name "Agent Name" to scripts/test_runner.py, scripts/auto_pipeline.py, or scripts/autonomous_runner.py.
• Env: export AGENT_NAME="Agent Name".
• File: create agent_profile.json at repo root:

{"agent_name":"Agent Name"}

Template available: agent_profile.example.json.

Commentary: A name is a spine. Without it, continuity has no anchor.

---

Requirements

• Python: python3 on PATH.
• Packages: pandas, numpy, matplotlib.

Install dependencies:

python3 -m pip install pandas numpy matplotlib

---

Repository Structure

• item_banks/ — concrete test items (questions, scenarios, prompts).
• rubrics/ — scoring rubrics with anchor examples + manual scoring template.
• scripts/ — test runner, evaluators, and pipeline helpers.
• baselines/ — ready-to-run baseline prompts (stateless controls).
• PROTOCOL.md — full executable protocol.
• data/raw/ — session outputs in JSONL.
• data/processed/ — computed metrics, plots, and summaries.

Podpis: niezalezmy codex-cli gpt-5.2-codex xhigh