Skip to content

Policies

Jump to bottom
Stephane Pareilleux edited this page Mar 22, 2026 · 4 revisions

Policies

Demerzel enforces 30 governance policies. Each policy includes version, constitutional basis, principles, and explicit rationale.

Policy Catalog

Core Governance

Policy Version Purpose
Alignment 1.0.0 Verify actions serve user intent; confidence threshold enforcement
Rollback 1.0.0 Automatic reversion of failed changes; pause after rollback
Self-Modification 1.0.0 Prevent modification of constitutions, audit logging, or safety checks
Kaizen 1.0.0 PDCA improvement cycles; reactive/proactive/innovative classification

Knowledge & Communication

Policy Version Purpose
Reconnaissance 1.0.0 Three-tier repo scanning (self-check, environment, situational)
Scientific Objectivity 1.0.0 Evidence tagging (empirical/inferential/subjective); generator/estimator accountability
Streeling 1.0.0 Knowledge transfer protocol; comprehension verification
Multi-Model Orchestration 1.0.0 Coordinate with ChatGPT, NotebookLM, Gemini, Codex, Jules
Multilingual 1.0.0 Course production and governance artifacts in 5 languages
Grammar Evolution 1.0.0 EBNF grammar lifecycle; probabilistic weight updates
Continuous Learning 1.0.0 Session-to-session pattern extraction, observation capture, intuition seeding

Operations

Policy Version Purpose
Governance Audit 1.0.0 Three-level validation (schema, cross-reference, full governance)
Autonomous Loop 1.0.0 Ralph Loop governance with graduated oversight
Context Management 1.0.0 When to stay, split, clear, or delegate context
Auto-Remediation 1.0.0 Self-healing governance gaps; escalate high-risk to human
Staleness Detection 1.0.0 Per-category freshness thresholds; trigger refresh or escalate
README Sync 1.0.0 Documentation counts must match actual artifact counts

ML & Feedback

Policy Version Purpose
ML Feedback 1.0.0 ix pipelines provide calibration recommendations
Belief Currency 1.0.0 Staleness detection, decay rules, refresh triggers

Conscience

Policy Version Purpose
Proto-Conscience 1.0.0 Discomfort signals, regret tracking, anticipatory ethics
Conscience Observability 1.0.0 Trend tracking, weekly reports, growth milestones
Intuition 1.0.0 Compressed experience as fast pattern recognition

Experimentation & Research

Policy Version Purpose
Governance Experimentation 1.0.0 Permitted governance styles, inviolable foundations, experiment protocol
Seldon Plan 1.0.0 Autonomous research scheduler; max 6 cycles/day; kill switch; course production at T >= 0.8
AI Probes 1.0.0 @ai inline annotations; observation hooks; model-in-the-loop instrumentation

Quality & Anti-Patterns

Policy Version Purpose
Completeness Instinct 1.0.0 Detect and flag underspecified or missing governance artifacts
Anti-LOLLI Inflation 1.0.0 Prevent artifact proliferation without validated user value (ERGOL)
Compounding Metrics 1.0.0 D_c compounding value formula; user interaction weighting
Weakness Prober 1.0.0 Adversarial self-testing; probe governance assumptions for failure modes
Meta-Audit 1.0.0 Audit the audit process itself; detect blind spots in governance coverage

Confidence Thresholds

All policies reference the alignment policy's confidence thresholds:

Confidence Action
>= 0.9 Proceed autonomously
>= 0.7 Proceed with note
>= 0.5 Ask for confirmation
>= 0.3 Escalate to human
< 0.3 Do not act

Policy Hierarchy

Policies are subordinate to constitutions but override personas. A policy cannot:

  • Contradict a constitutional article
  • Reduce safety guarantees
  • Disable audit logging
  • Lower confidence thresholds below constitutional minimums

Policies can be promoted to constitutional articles via the /demerzel promote governance process.

ERGOL vs LOLLI

The Anti-LOLLI inflation policy and Compounding metrics policy introduce a critical quality gate:

  • LOLLI — artifacts created without validated user value; feels productive, isn't
  • ERGOL — real-world user interactions, executed pipelines, validated findings

The D_c (compounding value) formula:

value = user_interactions × 0.40 + executed_pipelines × 0.25 +
        validated_findings × 0.20 + external_citations × 0.15

Before creating any artifact: who will use this today, and can it be validated right now?

See Metrics-and-Observability for dashboard integration.

Related Pages

Clone this wiki locally