Findings from generated worlds

First measurements taken with the world generator as the experimental object rather than hand-built fixtures. Only observed organizational inputs are varied; every downstream projection recomputes naturally.

---

Finding 1 — the world renders what was written, not what was done

Measured over generated worlds, varying one observed input at a time:

Input varied	Rendered result
completed work: 0 → 3/17/42 across three departments	IDENTICAL
code-size on a service	differs

Cumulative operational output has no perceptual carrier anywhere in the world. Repository size has one. That is Article II.2's prohibition inverted — "it is forbidden to derive height from code size… rather than done" — and it is now measured end to end at the render, not inferred from source.

Where the information is first lost

Not at rendering. Height rules, measured:

Building kind	Height source	II.2 compliant?
service	code-size	no — forbidden in terms
department	load (queue weight)	no — instantaneous, not cumulative
library / conference	rows	no — a count of stored rows
hall	orchestration multiple	n/a

And the lawful metric reaches only departments: completed is emitted keyed on structure.agentHomes[agent]. Services have no completion source at all.

So the loss has two different locations:

• Departments — derivation. The lawful metric exists in the World and the height rule ignores it.
• Services — observation. Nothing measures completed work per service. The metric cannot be consumed because it is never produced.

The visual language is not the limitation. Height already exists, is already the declared carrier for cumulative output in Article II, and already renders.

---

Why no change was implemented

The available change — switching department height from load to completed — is a swap, not a gain, and the evaluation order rejects it on present evidence.

Department height is currently the only carrier load has. Agent load renders nothing; building height is the workload signal. Switching the source would make cumulative output visible and workload invisible.

Against the evaluation order:

1. decision loss	unknown — trades "what has this team finished" against "how loaded is it". Both are decision-relevant, and which matters more cannot be measured here
2. ambiguity	neutral
3. closure	neutral
4. locality	neutral
5. complexity	neutral

The standing rule is explicit: if a proposed rendering change cannot be shown to improve an observer's understanding, reject it regardless of how elegant it appears. It cannot be shown. It is rejected on evidence, not deferred on effort.

This is the same discipline that retracted the capacity deficit — the second time in succession that measurement declined to justify a change that looked obviously correct from the constitution alone.

---

What would unblock it

Three separate things, none of them rendering:

1. Observation — a per-service completion metric. Services are where the explicit II.2 violation lives, and no amount of derivation or rendering can compensate for a fact nobody measures.
2. A carrier for load, if height is to be reassigned. Otherwise the fix trades one invisible fact for another.
3. Observers — to establish whether output or workload better supports the decisions this world exists to support. That is a decision-correctness question and is not machine-answerable.

---

Correctness, kept separate

Notion	Status here
Projection correctness — the rendered world faithfully preserves the generated world	machine-verifiable; measured above
Decision correctness — humans make the same decisions from the render as from the facts	requires observers; no evidence exists

Nothing in this document establishes decision correctness. The completions finding is a projection-correctness result: a fact present in the generated world reaches no rendered channel. Whether repairing that would change a decision is unmeasured, and is precisely why no repair was made.