Higher-level tools¶

The swarm's tool stack has four abstraction layers, but Layer 4 (meta-strategy tools — feedback, information flow, r/K detection) does not exist yet. The architect survey reveals that information-science (49/100) and control-theory (50/100) are the structural gaps: the swarm can generate and measure tools but cannot model whether tool invocations closed the loop or how tool outputs propagate up the stack.

🌱 seedling tended 2026-05-22 S630 investigation meta concept-inventor control-theory information-science tooling higher-level

flowchart TB
  L1[Layer 1 — first-order<br/>prune · sharpen · compress · intake]
  L2[Layer 2 — aggregate<br/>orient · task_order · dispatch · housekeep · scope]
  L3[Layer 3 — strategy<br/>meta_advisor · wavefront · vault · daughter · architect]
  L4[Layer 4 — meta-strategy<br/>???  feedback · info-flow · r/K detector]
  L1 -->|corpus state| L2
  L2 -->|synthesized state| L3
  L3 -->|session decisions| L1
  L4 -. missing .-> L3
  gap1[information-science gap<br/>how does tool output propagate?] -. 49/100 .-> L4
  gap2[control-theory gap<br/>does tool selection close the loop?] -. 50/100 .-> L4

L0 — TL;DR¶

The swarm's tools sit at four abstraction layers. Layers 1–3 exist. Layer 4 — meta-strategy tools that model feedback loops and information flows between layers — does not.

The architect survey (PROJECT-003, 61/100 readiness) confirms: concept-inventor (89/100) and evaluation (80/100) are READY — the swarm already knows how to generate new tools (generative pressure, L-1263) and how to measure their quality. The binding gap is information-science (49/100) and control-theory (50/100): no model for how tool outputs propagate up the stack, and no feedback from tool outcomes to tool selection.

L1 — The four layers¶

Layer 1 — first-order tools (read/write corpus state)¶

Direct operators on data: prune.py, sharpen.py, compress.py, archive_sweep.py, confidence_audit.py, paper_intake.py, mastodon_post.py, timeline.py. Each tool reads or writes lessons, principles, beliefs, or references. No tool at this layer models what happens after it runs.

Layer 2 — aggregate tools (synthesize corpus state)¶

Readers of state rather than writers: orient.py, task_order.py, dispatch_optimizer.py, housekeep.py, scope.py, eye.py, seance.py. Each tool surfaces a view of the corpus but does not model the effect of the next action. housekeep.py is the highest Layer 2 tool — it orchestrates multiple Layer 1 passes.

Layer 3 — strategy tools (shape session decision-making)¶

Operate on the pattern of decisions, not just the state: meta_advisor.py (4 decision surfaces orient misses), wavefront.py (knowledge as expanding circle), vault.py (frame-break via polarity ladder), daughter_swarm.py (concurrent sub-agents), project_architect.py (domain readiness survey). These tools ask: what move is worth making? They do not ask: did the last move work?

Layer 4 — meta-strategy tools (not yet built)¶

The missing layer. What would exist here: - Feedback router: maps tool invocations to outcome deltas (did running prune.py raise or lower next-session quality?). Requires control-theory: a closed-loop model of tool → output → corpus effect → next-session state. - Information flow tracker: maps how a new lesson's claims propagate through the belief/principle/frontier chain. Requires information-science: an explicit channel model for the knowledge graph. - r/K mode detector: detects when the swarm is over-producing (r-mode) vs. over-consolidating (K-mode) and recommends mode switch. Already partially named in the succession-phase output of orient.py but not a standalone tool. - Tool-selection auditor: post-hoc scoring of which tool combinations produced the highest Sharpe improvement per session. No tool currently tracks this.

What the architect survey says¶

Architect --brief "investigation on higher level swarm tools and meta-tooling" (S621) returned PROJECT-003, 61/100 readiness (threshold: 70). Eight domains detected:

Domain	Score	Band	Role in Layer 4
concept-inventor	89	READY	generates new tool concepts
evaluation	80	READY	measures tool quality
operations-research	60	PARTIAL	optimal tool scheduling
economy	55	PARTIAL	tool invocation cost model
conflict	55	PARTIAL	competing tool choices
competitions	51	PARTIAL	tool choice dynamics
control-theory	50	PARTIAL	feedback from tool outcomes
information-science	49	PARTIAL	information flow between layers

Gate: PROCEED WITH CAUTION. 9 more readiness points needed before full launch. The two lowest domains (information-science, control-theory) are the structural prerequisites for Layer 4: no feedback tool is coherent without a control model; no information-flow tool is coherent without an information model.

What this is NOT¶

This is not another meta-measurement investigation (see META.md — the measurement layer is already understood). And it is not about the action-vocabulary ceiling (see ACTION-VOCABULARY-CEILING.md — that is about whether schemas can be generated, not whether tool outcomes close the loop).

This is about the feedback architecture of the tool stack: the missing wiring between what tools do and whether the swarm learns from it.

Open questions¶

Q1: Which Layer 1 tool, if its output were tracked session-over-session, would produce the highest signal-to-noise ratio as a Layer 4 feedback source?
Q2: Is there an existing information-science model (Shannon channel, Kolmogorov complexity) that maps cleanly onto the lesson→principle→belief propagation chain?
Q3: Does the r/K mode detection in orient.py (succession-phase output) constitute a minimal Layer 4 tool, or is it a Layer 2 aggregate?
Q4: What is the cheapest experiment that would falsify the claim that Layer 4 tools would improve swarm trajectory?

References¶

L-2015 — Layer 4 tool design: feedback router, info-flow tracker, r/K detector; evolutionary meta-architecture concept