Status: pre-launch. This page describes target design; this document does not establish a deployed service.
Two Loops — Execution and Evolution
Current reading — 2026-09-27
Operational work uses the inputs required by the owning contract. Independent outcome evidence may seed an improvement proposal; separate owner adoption and a later scoped execution are required for change. L5–L7 provide counsel, not autonomous runtime mutation. [I]
This is an editorial boundary, not a new runtime contract, upstream repeal or proof of adoption. See current status and the product roles. Original revision dates remain as recorded below.
Historical research draft — retained on 2026-09-27; not current terms or implementation instructions
The organism runs two distinct loops. The execution loop runs the pipeline. The evolutionary loop revises the pipeline itself. They operate at different timescales, different frequencies, and different levels of the Rosetta. Confusing them is the most common architectural error in self-organizing systems.
The Execution Loop (L1 -> L4)
L1 (Perception) -> L2 (Pattern) -> L3 (Model) -> L4 (Decision/Action)
IS IS COULD SHOULD + ACT
The execution loop runs per signal. Every observation that enters Circle triggers one pipeline cycle:
- L1: Raw data ingested (sensor reading, market price, news article)
- L2: Pattern matched against known signatures (EigenTrust scoring, anomaly detection)
- L3: Model updated (ReFu probability recalculated, LMSR price adjusted)
- L4: Decision reached and executed (APU Court Deliberation, Skyzai settlement)
Frequency: Milliseconds to hours, depending on signal urgency. Cost: Low per cycle. The Boltzmann distribution (see McGilchrist) ensures most execution-loop cycles use cheap models. Failure mode: The execution loop produces a wrong answer. A prediction fails. A trade loses. An observation was misclassified.
The execution loop cannot fix itself. It executes the current model. If the model is wrong, the execution loop faithfully executes the wrong model.
The Evolutionary Loop (L4 -> L5 -> L6 -> L7 -> L4*)
L4 (Failure detected) -> L5 (Witness) -> L6 (Dissolution) -> L7 (Void) -> L4* (New model)
Execution error Observe Deconstruct Reset Reconstruct
The evolutionary loop fires only when the execution loop fails. It is the meta-loop — the loop that changes the loop.
- L4 (Failure): The execution loop's output is wrong. P-score drops. A kill criterion triggers. A prediction market resolves against the model.
- L5 (Witness): The system observes itself failing. This is the Witness tier (see Three Tiers). No action — pure observation of the failure pattern.
- L6 (Dissolution): The failing model is deconstructed. Assumptions are identified. Hidden premises are surfaced. The Sadu caste (L6 agents) strip the model to its axioms.
- L7 (Void): The space between the old model and the new one. No model is active. This is the creative vacuum. The Rishi caste (L7 agents) hold the space.
- L4 (New model):* A new model descends from L7 through L6, L5, back to L4 — but it is a different L4. The asterisk marks the update. The execution loop resumes with the new model.
Frequency: Days to months. Evolutionary-loop cycles are rare and expensive. Cost: High per cycle. Requires frontier models (CALLOSUM allocation). Evolutionary-loop cycles are where the organism spends its compute budget. Failure mode: The evolutionary loop produces a worse model. This is caught by AK7 (self-correction kill criterion).
The Descent Rule
All evolutionary-loop output must descend to L4* for execution. This is a hard rule, not a guideline.
Why: L5-L7 insights that remain at L5-L7 are philosophy, not engineering. The organism is not a philosophy department. Every evolutionary insight must be compiled into an executable model the pipeline can run. If it cannot be compiled, it is not yet an insight — it is a speculation.
The descent path: L7 -> L6 (formalize) -> L5 (validate against witness data) -> L4* (deploy as the pipeline's new model).
The Two Loops in Practice
| Property | Execution loop | Evolutionary loop |
|---|---|---|
| Trigger | Every signal | Execution-loop failure |
| Levels | L1-L4 | L4-L7-L4* |
| Timescale | ms to hours | Days to months |
| Compute | Cheap models (80%) | Frontier models (5%) |
| Output | Actions | New models |
| Failure | Wrong action | Wrong evolution |
Why Two Loops, Not One
A single loop that both executes and evolves is unstable. It either:
- Optimizes for execution and never evolves (brittle)
- Optimizes for evolution and never executes (academic)
Separating the loops allows the execution loop to run fast and cheap while the evolutionary loop runs slow and deep. The organism acts at L4 speed and evolves at L7 speed. Both are necessary. Neither is sufficient.
The execution loop runs the game. The evolutionary loop changes the rules. Both descend to L4.