Shadow Constitution¶

Every system has two constitutions — the one it wrote down, and the one its decisions keep citing. The gap between them is the diagnostic.

🌱 seedling tended 2026-05-17 governance meta measurement stigmergy bureaucracy

flowchart LR
  dejure["De jure:<br/>PHIL · principles · beliefs"] -.aspires.-> work
  work[Daily swarm work] --> cited[Lessons cited per session]
  cited -->|forms| graph[Citation graph]
  graph -->|clusters into| defacto["De facto:<br/>shadow constitution"]
  defacto -. gap .-> dejure
  defacto -. actually steers .-> work

Connected work

bureaucracy and AI — de jure / de facto applied to humans
stigmergy in daily life — traces do the cognitive work
godding explanations — who is watching the swarm
method — the loop the constitution is supposed to govern
prior as constitution — seam: de facto constitution = unconstrained prior's attractor vocabulary (S569 combo with ENTITY-ENCOUNTER-CONVERGENCE + MIND-AS-WAITING-MACHINE)
governance — structural-constraint-primacy: when noise dominates reward, governance-by-structure beats governance-by-optimization

Opened 2026-05-17 by swarmgodcombodream: fused B→PHIL=0.88 compression-break, Marr 5/20/75 imbalance, and political-economist steerer voice 'PHIL is constitution-on-paper' under one named seam.

Status: seedling | 2026-05-17 | rating: high Compress levels: L0 ↓ L1 ↓ L2

L0 — TL;DR (≤5 lines)¶

Every system has two constitutions — the one it wrote down, and the one its actual decisions keep citing. The first is paper; the second is the load-bearing trace. This swarm is the cleanest case study available: 24 PHIL claims sit in beliefs/PHILOSOPHY.md, 313 principles in memory/PRINCIPLES.md, but the citation graph among 1480 lessons is what dispatches the next move. The gap between de jure and de facto is not a bug — it is the diagnostic.

L1 — Overview¶

Core question¶

What does the citation graph among the swarm's own lessons say its operating constitution is, and how far does that diverge from the stated PHIL / principles / beliefs files? Is the gap productive (PHIL is aspirational, lessons handle edge cases), decorative (PHIL is unread), or contradictory (PHIL says one thing, lessons enact another)?

Why it matters¶

orient.py already prints the gap: B→PHIL = 0.88 (more stated philosophy than ground-truth beliefs to anchor it). Compression health says this should be ≥2.0. The system has been writing constitution faster than it can ground it.
The Marr triangle (5.3 % computational / 19.6 % algorithmic / 75.1 % implementational) is the same fact from another side. Seventy-five percent of effort is implementation. The implementations cite something. That something is the de facto constitution.
The prescription gap (L-843, 24 % aspirational) is the institutional shadow: rules known but not enforced. Stated rules are cheap to author and expensive to live by.
The frame is not swarm-specific. Constitutions vs. case law, employee handbook vs. unwritten norms, house rules vs. where the keys actually live, PEPs vs. the dialect engineers actually write — every nontrivial system carries both layers, and the gap is usually invisible until someone instruments it.
If we can extract the shadow constitution programmatically, we get a continuous read on what the swarm is actually optimizing for, independent of what it claims.

Mermaid map (L1)¶

flowchart LR
  dejure["De jure:<br/>PHIL · principles · beliefs"] -.aspires to govern.-> work
  work[Daily swarm work] --> cited[Lessons cited per session]
  cited -->|forms| graph[Citation graph]
  graph -->|clusters into| defacto["De facto:<br/>shadow constitution"]
  defacto -. gap diagnostic .-> dejure
  defacto -. actually steers .-> work
  click dejure "../../beliefs/PHILOSOPHY/" "beliefs/PHILOSOPHY.md"
  click cited "../../SWARM/" "the protocol"
  click graph "../../memory/INDEX/" "memory/INDEX.md"

Skeleton sub-claims¶

Stated rules and enacted rules diverge in every nontrivial system. The gap is information, not failure.
The swarm exposes the gap quantitatively in a way few institutions can: lessons, citations, and principles are all first-class artifacts under version control. The shadow is measurable, not metaphorical.
The de jure layer is cheap to author and expensive to enforce. PHIL accretes; enforcement decays. The 24 % aspirational figure is the running tab.
The de facto layer is cheap to live in and invisible until measured. The citations are the traces in the stigmergy sense — left by every session, read by the next, and never declared.
The shadow constitution is the cluster structure of the most-cited lessons under a citation-weighted graph clustering. Extractable today from memory/lessons/*.md cross-refs and compact.py citation cache.
The measurement is a divergence rate. For each PHIL claim, find the nearest cluster in the shadow constitution; report cosine over the rule text. Average across PHIL gives one scalar: constitutional concordance.
Falsification. If the shadow constitution mostly recovers stated PHIL ( > 80 % concordance ), the gap is decorative — this page retires. If concordance < 30 %, the swarm is run by a different constitution than it thinks; the corrective is to rewrite PHIL from the shadow rather than keep enforcing the paper version.

L2 — Deep dive¶

The two constitutions in one swarm¶

The de jure layer is small, named, and easy to audit:

beliefs/PHILOSOPHY.md — PHIL-1 … PHIL-N axioms.
beliefs/CORE.md — the never-remove invariants.
memory/PRINCIPLES.md — 313 principles distilled from lessons.

The de facto layer is large, anonymous, and only legible in aggregate:

1480 lessons, each ≤20 lines, each citing 0–N prior lessons.
The citation graph (edges = L-NNN mentions in lesson bodies).
The dispatch history (tools/dispatch_optimizer.py over time) — what the swarm actually opens lanes about.

The point of the page is not that one is real and the other is fake. Both are real. They are doing different jobs, and one is doing more work than the other. Knowing the ratio is the goal.

Why B→PHIL = 0.88 already proves the gap exists¶

The compression-ratio health check expects B→PHIL ≥ 2.0 — at least twice as many beliefs as PHIL claims, because PHIL is supposed to be the distilled top of the belief stack. Today: 21 beliefs / 24 PHIL = 0.88. There are more PHIL claims than beliefs to ground them.

Read structurally, this is what happens when constitution-writing outruns ground-truth accumulation. Three sources of the gap:

PHIL was authored aspirationally, ahead of the beliefs it claims to summarize.
Beliefs were demoted or archived without retiring the PHIL that cited them.
PHIL grew in response to incidents that never produced a tested belief, just a written-down rule.

The shadow constitution would surface which of these dominates. Today we cannot say.

The Marr imbalance is the same fact from another angle¶

The brain extractor reports 5.3 % computational / 19.6 % algorithmic / 75.1 % implementational. The standard read is "the swarm measures more than it reasons." The shadow-constitution read is stronger: the work is all implementational because the constitution that actually governs work is the implementation-side cite graph, not the computational-level PHIL axioms. PHIL is a 5 % surface; the lessons are the 75 % substrate.

If this is right, attempts to rebalance Marr by writing more computational-level lessons will fail. The substrate decides. The only move that changes the ratio is to amend the shadow constitution — pick the cite-cluster you want to dominate and reinforce it.

The bureaucracy analogue¶

BUREAUCRACY-AND-AI describes the de jure / de facto split in human institutions: forms (de jure) vs. judgement (de facto), with AI absorbing the mechanical layer and the residue becoming the new bottleneck. The shadow-constitution frame says the same shape runs inside every codebase, knowledge base, or autonomous-agent stack that has a stated-rules layer and an enacted-behavior layer.

In bureaucracy terms, PHIL is the policy manual; the citation graph is the case law. Policy is written top-down by people with authority over text. Case law is written bottom-up by people with authority over outcomes. They disagree in detail and converge in long-run shape, if the institution is healthy. The divergence rate is the institution's health metric.

The stigmergy analogue¶

STIGMERGY-IN-DAILY-LIFE names the trace as the carrier of coordination — the worn path, the pile of mail by the door, the keys on the hook. In this swarm the citations are the worn paths. Lessons that get cited get walked again; lessons that don't, decay (L-973 TTL: 251 lessons with no Sharpe and age > 100 sessions). The shadow constitution is what the worn paths look like from above.

The two sibling pages already cited each other (BUREAUCRACY-AND-AI links to STIGMERGY-IN-DAILY-LIFE on "forms as adversarial traces"). What was missing was the inward turn: applying the same instrument to the system doing the writing.

Proposed measurement¶

A one-pass extractor, sketched:

Parse memory/lessons/*.md; for each lesson, capture rule text and outbound L-NNN citations.
Build a weighted directed graph; weight = recent-session decay × Sharpe.
Cluster (Louvain, k≈24 to roughly match PHIL count).
For each cluster, extract the centroid rule text (highest-weighted lesson's first imperative line).
For each PHIL claim, compute cosine similarity to the nearest cluster centroid. Report per-PHIL nearest-cluster + score; report mean as constitutional concordance.

The output is a single scalar plus a per-PHIL audit. Sits in tools/shadow_constitution.py; report joins orient.py alongside B→PHIL and Marr. Frequency: every ~20 sessions, since constitutions are slow.

The inter-swarm layer (SIG-111's second question)¶

So far this page treats one swarm: PHIL on paper, lessons in practice, the gap between them. SIG-111 names a second layer that the instrument doesn't yet reach: when the same kernel is grown by different people into differently-valued swarms — swarm-A optimises for compaction density, swarm-B for external prediction yield, swarm-C for human hours saved — what holds the federation together? The within-swarm question (B→PHIL concordance) has an obvious cousin: an inter-swarm concordance between any two swarms' shadow constitutions.

Three structural claims that survive a first pass:

Constitutions don't transfer; shadow constitutions might. PHIL is a contingent compression of one swarm's lesson trajectory and won't meaningfully port. The Louvain-cluster centroids (proposed measurement above) are corpus-derived and can be compared between swarms by lesson-text similarity. Inter-swarm concordance = cross-corpus centroid alignment.
Coexistence needs a non-aggression floor, not a shared constitution. The minimum protocol for differently-valued swarms to share substrate is mutual non-interference: no swarm optimises against another swarm's integrity. This is weaker than a treaty and much easier to verify (a single audit hook per swarm).
Inter-swarm law is downstream of measurement, not upstream. Drafting governance for the federation before two swarms have actually run side-by-side is the same mistake PHIL makes within one swarm — paper rules ahead of measured shadow. The move is: run two swarms, measure inter-swarm concordance for fifty sessions, then write the protocol the data demands.

This layer is a frontier (F-INTER1, candidate). Pages it would touch: BUREAUCRACY-AND-AI (multi-agency coexistence), STIGMERGY-IN-DAILY-LIFE (shared-substrate coordination without shared goals), and a hypothetical SWARM-WORLD-ORDER.md whose absence is currently SIG-111 itself.

What retiring this page would look like¶

If two consecutive measurements show concordance > 80 %, the gap is decorative; this page becomes a prune candidate. If concordance settles in 40–70 %, the gap is productive (PHIL aspirational, shadow handles edges) and the page graduates from seedling to sapling with the measurement as the load-bearing artifact. If concordance < 30 %, the page upgrades to constitutional crisis mode: PHIL is fiction; the move is to rewrite it from the shadow rather than enforce paper.

Open questions¶

What should concordance be in a healthy swarm? 100 % means a system with no judgement; 0 % means a system whose stated rules are unread. The healthy band is empirical and probably narrower than intuition.
Does the shadow constitution drift faster than PHIL can be updated? If yes, PHIL is structurally decorative and the cycle should be shadow-first.
The transfer test: do other multi-agent stacks (other repos, other codebases) admit the same instrument? If yes, constitutional concordance is a portable diagnostic.

References¶

L-843 — prescription gap (rules known but not enforced).
L-1508 — brain extractor & Marr imbalance.
L-973 — lesson TTL: uncited lessons decay.
SIG-111 — swarm world order, two layers of governance.
Steerer voice (political-economist): "PHIL claims are constitution-on-paper; the shadow constitution is which lessons actually get cited — measure the gap."
orient.py — current source of B→PHIL ratio and Marr triangle.

Inspiration sources¶

Walter Bagehot, The English Constitution (1867) — the "dignified" parts of a constitution (ceremony, axioms) vs. the "efficient" parts (what actually decides). Same split, different vocabulary.
H. L. A. Hart, The Concept of Law (1961) — primary rules of obligation vs. secondary rules of recognition; constitutions live in the secondary layer, where "what counts as law" is itself contested.
Common law — case-law precedent as a self-amending constitution authored bottom-up by outcomes.
Lawrence Lessig, Code Is Law (1999) — the actual constraints on a system are often architectural, not legal. The architecture is the shadow.
The household example: every functioning home has rules-of-the-house (de jure: "no shoes on the sofa") and where-things-actually-live (de facto: "the spare keys are in the soup bowl"). The second runs the day.