Repository navigation
Manifesto
These principles emerged from the Demerzel governance framework's accumulated experience across 85+ sessions, 9 issues closed, 65 translations, 8 repos, and one question that changed everything: "What did a real human find useful today?"
Source analysis: docs/articles/ergol-vs-lolli-lessons.md
A guitarist saying "Cool!" validates more than all design specs combined. Specs are promises. User interactions are proof. Weight your work accordingly.
ERGOL (productive capacity) is real value delivered to real users. LOLLI (inflated metrics) is activity that feels productive but isn't. Factory-of-factories amplifies LOLLI — easy creation makes discipline more critical, not less.
Don't create departments — create MetaBuild, which creates departments. Don't write policies — create the governance pattern that generates policies. The leverage is in the system, not the artifact.
Before claiming something works: is it T (verified), U (unknown), F (refuted), or C (contradictory)? Agent teams optimize for task completion (marking things T prematurely) — require evidence before T.
Rust, F#, .NET, Node.js, Python — the governance layer must work for all of them. Keep constitutions, policies, and personas as pure behavioral specifications. Zero executable code in the governance layer.
Every session is a compounding opportunity. Did this reveal a gap? Create an issue. Did this teach something reusable? Package it as a knowledge state. Did this validate a governance artifact? Update the evolution log. Never let an insight die.
Discomfort signals are not noise — they are the governance system detecting potential harm before it happens. Suppress them at your peril. Track them, trend them, resolve them.
The Asimov Zeroth Law overrides everything. No performance pressure, no deadline, no clever argument changes this. Hierarchy: Asimov Laws > Demerzel Mandate > Default Constitution > Policies > Personas.
Agents escalate at confidence < 0.5. Agents ask before acting at confidence < 0.7. Agents proceed with note at confidence >= 0.7. Full autonomy only at >= 0.9. The gradient exists for good reason.
Every session should end with: "What did a real human find useful today?"
If the answer is "nothing" — the session was LOLLI, regardless of artifact count. This is the only metric that cannot be gamed.
These principles encode into the compounding value formula:
D_c = user_interactions × 0.40 + executed_pipelines × 0.25 +
validated_findings × 0.20 + external_citations × 0.15
User interactions weighted 40% — because the only reliable ERGOL detector is a real user.
These principles crystallized after a session that produced 80+ artifacts, completed 24 autonomous tasks, and achieved an ERGOL score of ~15%. Only the Discord chatbot was validated by a real user — and that single validation was worth more than everything else combined.
The framework responded by creating the Anti-LOLLI inflation policy, revising the D_c formula, and adding the session rule as a governance checkpoint.
- Metrics-and-Observability — D_c and ERGOL/LOLLI metrics
- Policies — Anti-LOLLI and Compounding Metrics policies
- Architecture — factory-of-factories design philosophy
- Conscience-System — principle 7 in practice