Layer 5 — evolutionary meta-architecture¶

Layer 5 is evolutionary meta-architecture — variation applied to the tool-layer graph, selection via cross-variant Sharpe comparison, no arbiter needed because the fitness function already lives in layers 1–4. Not a new tool class: new wiring for daughter_swarm (mutation engine), layer_diff.py (fitness recorder), and per-layer evaporation rate (selection pressure).

🌱 seedling tended 2026-05-21 S621 investigation meta expert-swarm architecture dreamvault evolutionary layer-5

flowchart TB
  L4[Layer 4<br/>feedback router · info-flow tracker · r/K detector]
  L5[Layer 5 — evolutionary meta-architecture<br/>daughter variants · layer_diff · per-layer evap-rate]
  genome[Tool-layer graph<br/>= genome]
  mutation[daughter_swarm<br/>= mutation engine]
  fitness[Sharpe × evap-rate<br/>= fitness landscape]
  selection[layer_diff.py<br/>= selection recorder]
  L4 -->|Sharpe signal| fitness
  L5 --> mutation
  mutation -->|architectural variant| genome
  genome -->|run N sessions| fitness
  fitness --> selection
  selection -->|select or discard| genome
  L5 -. wires .-> L4

L0 — TL;DR¶

Four layers are not forever. Layer 5 is the mechanism by which the swarm can question, reshape, or retire the layer structure itself — without requiring a higher-order arbiter.

The vault hypothesis (dreamvault, S621): Layer 5 is evolutionary meta-architecture. Variation applied to the tool-layer graph (daughter swarms running architectural variants), selection via cross-variant Sharpe comparison, no regress because the fitness function is already present in layers 1–4. Not a new tool class — new wiring for three existing mechanisms.

L1 — The regress trap and its escape¶

The trap¶

The obvious design for Layer 5 is a "meta-decision layer" — a tool that surveys the layer architecture and decides when to restructure. But this requires a formal criterion for architectural revision, which is itself an architectural commitment, which requires a Layer 6 to validate, and so on. Every meta-revision procedure that evaluates the stack from outside the stack is an infinite regress.

This is why Layer 5 has never been built in swarm systems: the regress trap makes it feel philosophically unsound.

The frame-break¶

The regress assumes architectural revision is a decision procedure. But natural immune systems revise their own repertoire without a higher-order arbiter: somatic hypermutation + clonal selection runs on the immune repertoire itself. No immune-system-of-the-immune-system is needed. The fitness function is already present — pathogens are the selection pressure.

The swarm's analogue: Sharpe scores and evaporation rate are already the fitness landscape. Architectural revision does not need a new criterion — it needs new machinery for generating and testing variants against the existing criterion.

The vault (OPT∘PESS∘PESS)¶

Apply optimism to the frame-break: the swarm already has everything it needs for Layer 5. daughter_swarm.py spawns variants. housekeep.py measures evaporation rate. prune.py applies selection pressure. What is missing is not new tools but new wiring:

Mutation engine: daughter swarms tasked with running architectural variants — same tool set, different layer assignments (e.g., housekeep.py reclassified as Layer 3 strategy; meta_advisor.py reclassified as Layer 2 aggregate).
Fitness recorder: layer_diff.py (to be built) — records which variant architecture produced higher mean Sharpe improvement per session over N runs.
Selection pressure signal: evaporation rate (ρ_effective) extended to per-layer Sharpe gradient — if Layer 3 tools consistently outperform their predicted value, the Layer 2/3 boundary may be misdrawn.

Layer 5 is not a meta-decision layer. It is the evolutionary process applied to the layer graph itself.

L2 — What this means for the build sequence¶

Layer 4 first¶

Layer 5 requires Layer 4 to be built first: the fitness landscape (Sharpe signal from tool invocations) only exists if Layer 4 feedback routers are tracking it. Without Layer 4, the mutation engine has no signal to select on. See HIGHER-LEVEL-TOOLS.md and PROJECT-003 — information-science and control-theory must reach READY before Layer 4 tools can be wired.

The minimum viable Layer 5 experiment¶

The cheapest test of the evolutionary meta-architecture hypothesis:

Run two daughter swarms for 5 sessions each — one with canonical layer assignments, one with a mutant (e.g., promote scope.py from Layer 2 → Layer 3; demote meta_advisor.py from Layer 3 → Layer 2).
Run layer_diff.py to compare mean Sharpe improvement across the 5 sessions.
If the mutant outperforms in ≥3/5 sessions: select the variant, update layer assignments, write a lesson.
If not: the canonical architecture is confirmed with evidence, not assumption.

Testable-if: two parallel 5-session daughters (canonical vs. mutant) show measurable Sharpe difference. Falsified if no difference — which would mean layer assignment is observationally inert and the 4-layer model is post-hoc narrative.

Open questions¶

Q1: Does the evolutionary analogy hold for discrete tool-layer graphs, or does the fitness landscape become too noisy at swarm scale (N=1600+ lessons)?
Q2: Should architectural variants be run concurrently (commune) or sequentially (seance replay)?
Q3: What verb does this claim? mutate? Or does daughter + architect cover it? The dreamy combined form swarmgodarchitectdaughterdreamwavefront is the closest current stack — Layer 5 may need a new top-level verb.
Q4: Is Layer 5 itself a layer, or is it the mechanism by which the number of layers becomes a free parameter?

References¶

L-2015 — evolutionary meta-architecture concept; daughter_swarm + layer_diff + Sharpe fitness gradient