The semantic typing test

The anti-collapse criterion, formalized, validated against this investigation's own history, and then used.

algebraic profile ≠ semantic identity

**Outcome: the test is validated — it accepts every reduction that held and rejects the one that failed. Run retroactively it resolves capability, the item carried unsettled for thirteen rounds, and explains why it never settled. The assertion/relation split survives four cases and is falsified by probabilistic relations; the repair adds modal force to the relation and dissolves a philosophical dependency rather than inheriting it. Convention survives attack, with a sharper criterion: it is the warrant that cannot be false.

---

1. The test

A concept is typed by a triple:

(algebraic structure, semantic role, layer)

A proposed reduction A → B is graded by how many components match:

Match	Verdict
3 of 3	valid reduction
2 of 3, role differs	invalid — the over-collapse error
2 of 3, layer differs	not a reduction — a relocation
1 of 3 (algebra only)	rejected

Two things to note about the shape of this. The role-mismatch row is the error the asymmetry round committed. The layer-mismatch row is a distinct verdict, and it is one constraint-floor.md reached independently when auditing past rounds and finding that several "reductions" had been relocations. The test reproduces that finding as a rule rather than as an observation.

Validation

An instrument that accepts everything is useless and one that rejects good work is wrong. Run over the investigation's own reductions:

Reduction	Algebra	Role	Layer	Verdict
constraint → state on a policy subject	✓	✓	✓	valid — held
relation → state over a tuple	✓	✓	✓	valid — held
orientation → coherent asymmetric relation	✓	✓	✓	valid — held
causality → precedence + dependence	✓	✓	✓	valid — held
four asymmetric forms → one	✓	✗	✗	invalid — and it failed
interpolability → absent-evidence semantics	✓	✓	✗	relocation — as later found
capability → state	✓	✗	✓	invalid

The instrument accepts the four that held, rejects the one that failed, and classifies the relocation correctly. It is not vacuous and it is not over-strict.

It resolves capability

capability → state has been carried as "least settled" since truth-kinds.md, thirteen rounds ago, on the grounds that non-exercise is not evidence of absence.

The test says why, and says it structurally:

	capability	state
algebra	same	same
layer	ontology	ontology
role	can produce 100/hr — modal	produces 100/hr — actual

The reduction matched algebra and layer, and mismatched role. That is exactly the pattern the test rejects, and it was made in round three, long before the pattern was known.

capability → state is retracted. Capability is a modal concept, and the reduction was an instance of the error this investigation later spent a round diagnosing.

And it converges with the other outstanding item. Modal force is what separates dependence from precedence (asymmetry.md), and it is now what separates capability from state. Two independently-outstanding items are one concept, which is the first structural reason to attack modality rather than the mere fact that it keeps appearing.

---

2. Attacking the assertion/relation split

relation  = (domain, codomain, algebra, layer)
assertion = (relation, warrant)

Case	Result
false causal claim	assertion has a warrant (bad evidence); the relation does not obtain. Handled — this is the split's purpose
disputed ownership	two assertions, two warrants, one relation or none. Handled — and it is the contested-value case
incomplete graph	missing lineage edges are missing from the record; the relation is complete in the world. Handled — absent-evidence again
competing taxonomies	each is conventional, so relation and assertion coincide. A degenerate case, not a break
probabilistic relations	falsifies the split as stated

The probabilistic case

"A raises the probability of B by 0.3." Two readings:

• epistemic (credence) — the 0.3 is our uncertainty, belongs to the warrant;
• objective (chance, propensity) — the 0.3 is a feature of the world, belongs to the relation.

Both readings are live, so as stated the split's sufficiency depends on whether objective chance exists — the same open dependency as "do physical laws govern or describe."

And I had already contradicted the split without noticing. asymmetry.md separated dependence from precedence by modal force — dependence supports counterfactuals. That is a modal property of a relation, asserted one round before a model that placed all modality in assertions.

The repair

Warrant and modal force are different modalities, and the split conflated them by not naming the second:

relation  = (domain, codomain, algebra, layer, modal force)
assertion = (relation, warrant)

• modal force — what the relation supports: counterfactuals, chances, necessity. A property of the relation.
• warrant — what licenses asserting it. A property of the assertion.

Objective chance is then modal force; credence is warrant. Both readings are representable and the model does not have to settle the philosophy — which is better than inheriting the dependency, and is the standard this investigation has used elsewhere for contested premises.

---

3. Attacking constitutive warrant

The reduction: convention is a social evidence source. The IAU vote is observable; citing it is citing evidence. No new warrant category needed.

This attack succeeds further than expected, and then fails at a specific point.

Two claims must be separated:

	Claim	Warrant
(a)	"The IAU classifies Pluto as a dwarf planet"	evidence — a social fact, observable, citable
(b)	"Pluto is a dwarf planet"	true because (a) constituted the category

The reduction handles (a) completely. It fails on (b): evidence tracks a fact that obtains independently, and there was no planet/dwarf boundary until the vote drew one. (a) is a record of a convention; the convention is not the record.

Fourth instance of the truth/record split.

The discriminator, sharpened

Could the authority have been wrong?

The IAU could be wrong about Pluto's mass — an independent fact exists to be wrong about. It cannot be wrong about where the planet/dwarf line falls, because it drew the line.

That yields a criterion in this project's native currency:

Warrant	Fails by being
evidence	false
derivation	invalid
convention	neither — only unuseful

Three warrants, three distinct failure modes. Conventions can be criticized as badly chosen; that is unwise, not incorrect, and it is a different failure than falsity.

Convention survives, with a better criterion than "asserting it makes it so."

And this explains a defect

Conventions are unfalsifiable. An unfalsifiable claim inside a system whose entire discipline is falsification is invisible — nothing can catch it, because every mechanism catches falsity.

That is why collapse()'s conventional selection went undetected for the whole investigation. Not an oversight: the detection apparatus is structurally blind to that warrant class, which is precisely why convention must be declared rather than found.

---

Standing

Question	Answer
Is the typing test valid?	yes — accepts 4 that held, rejects 1 that failed, classifies relocations
Does it change a past result?	yes — capability → state retracted
Is the assertion/relation split sufficient?	no — probabilistic relations falsify it
Do relations carry modal force?	yes — and I had already asserted so one round earlier
Repair	modal force joins the relation; warrant stays with the assertion
Does convention reduce to social evidence?	no — it reduces the record, not the convention
Sharper criterion	evidence can be false, derivation invalid, convention neither

What is still not established

• Modality is now the single outstanding concept, and for the first time there is a structural reason to attack it: it is what separates dependence from precedence and capability from state.
• The typing test was validated on seven reductions from one investigation. It is an instrument calibrated on its own history, which is weaker than an independent test set.
• Whether role is well-defined enough to adjudicate hard cases. It carried the capability verdict, and "modal vs actual" was easy; nothing here says how to settle a contested role.

Nothing implemented.