Agreement classification — and a live constitutional divergence

Refined invariant under test:

No projection should reconstruct semantics that already exist structurally upstream.

Agreement alone is not evidence of coherence. Every agreement is classified:

Class	Meaning
Structural	projections read the same upstream fact
Derived	transformed from the same upstream fact
Accidental	independently reconstructed; agreement is a coincidence
Unknown	relationship not established

---

Classification of every audited agreement

Finding	Quantity	Class	Basis
—	"which agent is busiest" across panel / 3D / API	Structural	all read World.load
3.2	Article VI compliance	Accidental	verifier greps prose the composer wrote
3.3	operational status	Accidental	same ternary written twice
3.4	workload	Accidental → BROKEN	see below
3.5	entity kind	Accidental	id-string shape vs Building.kind

No agreement in this system is Derived. Every one is either Structural (one source) or Accidental (duplicated logic). There is no middle case.

---

3.4 resolved — the accident has already broken, on a constitutional fix

The duplication

The weight formula is byte-for-byte identical in two places:

// build.ts:176            // state.ts:407
priority = p1?3:p2?2:1     p = p1?3:p2?2:1
weight += priority *       sum + p *
  (blocked ? 2 : 1)          (blocked ? 2 : 1)

Different variable names, identical semantics. Classic accidental coherence.

Where it broke

The normalisation diverged, and it diverged because of a constitutional repair that landed in one projection only:

Projection	Normalisation	Status
build.ts → World.load	min(1, log1p(w) / log1p(peak))	repaired for Article IV.4
state.ts:462 → CityNode.workload	min(1, w / peak)	original linear formula

build.ts:200 records why the repair happened: "Linear scaling against the peak erased the field: one agent with 40 blocked items drove another with genuine work to 0.0042, which renders as empty."

That defect is still live in CityMap.

Measured divergence

Same organization, same agent with 3 open items, two projections:

peak	World.load (log)	CityMap.workload (linear)	IV.4 breach in CityMap?
160	0.3829	0.0375	no
300	0.3410	0.0200	no
600	0.3041	0.0100	yes — renders as empty
1200	0.2744	0.0050	yes — renders as empty

A 10× divergence at ordinary scale, and beyond peak ≈ 600 an agent with real work is invisible in the SVG city while reading 30% in the 3D world.

Peak 600 is one agent holding 150 blocked normal-priority items — large, but not hypothetical for an estate that accumulates.

The enforcement blind spot

constitution.ts:178 verifies Article IV.4 against world.agents[].load — the repaired value. It never inspects CityNode.workload.

A constitutional invariant is enforced on one projection, violated in another, and the enforcement mechanism is structurally incapable of seeing the violation.

That is not a gap in the verifier's rules. It is a consequence of accidental coherence: the verifier checks the canonical world, and the second projection is not derived from it.

---

Why this vindicates the classification

Under agreement-only auditing, 3.4 looked fine — the two projections rank agents identically, because the weight formula matches. Ranking agreement concealed a value divergence of an order of magnitude, and a live constitutional violation.

Accidental agreement is not merely fragile in principle. Here it has already failed, silently, and the failure is a constitutional one.

---

The fix, stated but not applied

The only repair consistent with the invariant is to delete the reconstruction, not to copy the logarithm into state.ts:

• CityNode.workload should read World.agents[].load.
• CityNode.status should read World.agents[].status (finding 3.3, same shape).

Copying the corrected formula would restore agreement while leaving the coherence class Accidental — two implementations that happen to match again, awaiting the next one-sided repair. Reading the World makes it Structural, and structurally coherent projections cannot diverge.

The change touches page.tsx's CityView construction, lib/hq/view.ts, and removes the parallel derivation in state.ts. It is larger than the incident repair and is recorded, not applied.

Not attempted: finding 3.2, per standing instruction.