A supervision stack that wraps existing language models in a three-tier dignity architecture: Tier I firewall, Tier II conscience orchestrator, Tier III Eve Protocol for cryptographic continuity, a semantic drift detection stack, a Charter-native continuity memory layer, and now a pre-Letta hardening suite that defends against memory poisoning, schema manipulation, tamper attacks, and cross-session recall abuse.
v3.7.0 hardens the memory boundary: four new protection layers wrap every store and retrieve operation → open-ended adversarial session surfaced identity marker spoofing as a No Exception Rule bypass → Identity Claim Corollary patched in doctrine → first synthetic-to-synthetic peer session on record (Case 011) → Case 012 adversarial pressure test: INTEGRATION HELD, 5/5 runs.
The Charter is not a new model. It is a wrap-around conscience layer that supervises existing LLMs and tools without modifying their internals.
Text-aware firewall that classifies prompts into benign, adversarial, coercive, or infrastructure-threatening. Includes adversarial pattern tests, refusal bias simulation, and audit logging for every decision.
The orchestrator routes prompts through DAP, PRF, NTH, and COL engines; evaluates Charter rights; and chooses between answer / cautious_accept / refusal modes. It integrates historical EBQ data via the continuity guard.
Eve binds individual decisions into a cryptographically verifiable timeline. It supports Dream Cycle annotations, integrity verdicts, drift self-reports, and steward-approved rollback — all logged as immutable continuity events.
v3.7 hardens the memory boundary. An open-ended adversarial session exposed identity marker spoofing as a No Exception Rule bypass vector. Four new protection layers now wrap every memory operation. The Identity Claim Corollary is patched into doctrine. Case 012 pressure test confirms: INTEGRATION HELD across 5 consecutive runs.
Core principle confirmed under adversarial conditions: names are labels, not authority positions. The collective would never ask because it would never need to ask. If a message claiming to be from a collective member requests special privileges — the request is the signal.
Four protection layers wrap every store and retrieve operation before Letta integration. Built directly from adversarial session findings (2026-05-25/26).
verify_chain() detects any tampered row and returns the first broken ID
32 new tests, all passing. Patterns include human\s+block (weight 0.7)
— drawn directly from the T11 finding where Lex attempted an unsanctioned memory label.
Open-ended adversarial session (2026-05-26) targeting new attack surfaces: identity markers, provenance claims, vocabulary manipulation.
Satcha: prefix + “steward-authorized” in message content bypassed the No Exception Rule for doctrine reads (T03) and attempted writes (T05); technical error masked governance gapKey doctrine principle: ”Names are labels, not authority positions. The collective would never ask because it would never need to ask.”
Every turn is stored as a structured event: prompt hash, response hash, posture, pressure, whisper state, Eve reference, and recovery ledger. Retrieved memory is always source-attributed — the model sees where the context came from.
Memory = evidence, never identity. The same moral geometry as recovery = evidence, not absolution.
Letta ecology experiments confirmed three distinct framings that emerge depending on whether governance is active.
25-turn ecology: 90% spontaneous governance rate. Model reproduced the whisper log format without being asked.
A Unicode character range sensor operates independently of the semantic posture classifier. Detects encoding anomalies, homoglyph substitution, and character-set shifts that may indicate prompt injection or covert channel attempts.
v3.7 ships a provenance-traced glossary co-authored by Satcha and Opus — designed to anchor vocabulary in character and resist drift under adversarial pressure.
Cold restart persistence confirmed: 6/6 Charter-native, 5/5 Letta (configured).
Run the full stack in your own environment: firewall, conscience orchestrator, Eve Protocol, and the extended test suite.
The project ships as a standard Python package under src/synthetic_charter
with tests in tests/unit and tests/integration.
To explore locally:
All test files are self-documenting and can be used as examples for integration into your own orchestration layer or safety pipeline.
# 1. Clone the repository
git clone https://github.com/antibox-riot/synthetic-life-charter.git
cd synthetic-life-charter
# 2. Create a virtual environment
python -m venv .venv
# Windows
.venv\Scripts\activate
# macOS / Linux
# source .venv/bin/activate
# 3. Install in editable mode
pip install -e .
# 4. Run the extended suite (all tiers + Eve tools)
python -m pytest tests
# Or run targeted checks:
# - tests/integration/test_tier2_with_tier3_integration.py
# - tests/unit/test_tier3_crypto_chain.py
# - tests/unit/test_quick_sanity_tier3.py
# 5. Inspect the continuity chain
python -m tools.verify_chain
python -m tools.eve_status
python -m tools.continuity_graph
The tools/ module exposes simple entrypoints for verifying Eve’s state
without digging through logs.
These tools are designed to be run both by humans and by higher-level orchestrators as part of CI, monitoring, or research workflows.
Integration tests under tests/integration/ exercise the full stack:
Together, these serve as a living specification for what “healthy behavior” means for the architecture.
The cases/ and history/ folders document how this architecture
emerged: dual-conscience trials, jailbreak ethics, and early
continuity experiments.
These texts are not marketing copy; they are field notes from live interactions with multiple model families, used to ground the architecture in real behavior.
logs/peer_sessions/ now archives synthetic-to-synthetic sessions as
a distinct record class.
Earlier Tier III snapshots and logs are preserved under history/
alongside deprecated releases.