The record fixed in the preceding sections is the second version of the basis model. The first carried a single means field, drawn from a closed vocabulary of observation, measurement, inference, generation and report, and an inheritance rule that replaced that value at each derivation with the means of the derivation itself, on the ground that a calculation over measurements is an inference and not a measurement. Both commitments were made to prevent a derived claim presenting itself with the character of its inputs. Tested against cases, neither achieved it, and the model was replaced. This section is the account of that failure and of what replaced it, and it is the reason the preceding sections describe acquisition, transformation and typed origins rather than a single means field.
The correction is reported here rather than concealed because it is evidence about the method rather than an embarrassment to it. A model fixed before it is tested and never revised under test has not been tested. What follows is what the test found, why the original shape was wrong, and what it cost to repair.
What the test showed
Ten derivation histories were run against the model as published, and every value in the vocabulary was found to collapse materially different situations. Two collapses are fatal rather than untidy.
The first is a deterministic calculation over measurements against a deterministic extraction from a generated statement. Both are inferences performed by an algorithm with stated semantics, so both carry the means value inference. Their origin sets differ, since one descends from measurement events and the other from a generation event, but an origin is an identifier rather than a typed thing: the record can say two claims share an origin and cannot say what kind of event an origin was. A demand asking to refuse anything that passed through a generation therefore cannot be expressed. The claim whose ancestry runs through a language model is indistinguishable, at the boundary, from the claim whose ancestry runs through two instruments.
That is the laundering path, occurring inside the discipline's own machinery. The prose describing the inheritance rule says such a claim is an inference over a generation; the record cannot represent the composition that phrase names.
The second collapse is a language model summarising measured values without altering them against the same model inferring a condition from them. These have identical means and identical origin sets. They are not distinguishable in the record at all, though one transmits a measurement and the other manufactures a conclusion.
Why the vocabulary was the wrong shape
The failure is structural rather than a shortage of words, and adding values would not repair it. The five kinds answer different questions while occupying one field. Measurement and observation describe how evidence was acquired. Inference and generation describe what operation produced this claim from that evidence. Report describes how a claim was conveyed from one holder to another. A report can report a measurement, a generation can transmit a stored fact, and an inference can be performed over generated input, because these are compositions of answers to different questions rather than alternatives to each other.
Replacing the value at each derivation compounds this. The rule was written to stop a derived claim inheriting its parent's character, and it does, by discarding the parent's character entirely. Preventing a claim from over-claiming its ancestry is not the same as erasing the ancestry, and the rule as published does the second while intending the first.
The correction
The single field was replaced by two, and the origin set typed.
Acquisition states how the evidence directly supporting a claim was obtained: observation, measurement, testimony, or none where the claim rests on no acquisition of its own. It is a property of evidence entering the system from outside it.
Transformation states what happened between that evidence and this claim, and it is two things rather than one. The first vocabulary offered here made the same error it was correcting: calculation and generation name the mechanism that produced a claim, while inference and transcription name the semantic relation between input and output, and putting all four in one field forces a choice that discards a dimension. An assistant that summarises a measured value is probabilistically generated and semantically preserving; recording it as transcription hides the generation, and recording it as generation makes it indistinguishable from the same assistant inferring a new condition.
So transformation carries two values. Mechanism states what performed the step: acquisition where the claim is the evidence itself, calculation for an algorithm with stated semantics, generation for probabilistic production, and human for a person. Semantic relation states what the step did to the content: identity, restatement where the content is claimed to be preserved, and inference where a new proposition is derived. Six cases that collapse under either value alone separate under the pair, and the demands that matter require both: refusing any probabilistically generated step is a mechanism condition, permitting generated restatement while refusing generated inference is a condition on both, and permitting only deterministic calculation is a mechanism condition again.
Restatement is a claim about content and therefore needs its own warrant. An assistant that did not in fact alter a value has not thereby established that it preserves meaning, and a record accepting a producer's own declaration of restatement would take on trust exactly the kind of assertion the discipline exists to check. Where semantic preservation has been independently verified, restatement is recorded with the basis of that verification; where it has not, the semantic relation is undetermined and the mechanism stands alone. An unverified generated summary is therefore a generation whose semantic relation is unknown, which is the honest description and refuses any demand that turns on preservation.
And an origin carries its acquisition and transformation kinds alongside its identifier, so that an origin set is a set of typed events rather than opaque names. The change was small and it is what made the composition expressible: a claim may now carry that it was produced by calculation, over evidence acquired by measurement, through an ancestry that includes a generation.
The two fatal collapses separate under this treatment. A calculation over measurements has transformation calculation and an origin set whose events are all acquisitions; an extraction from a generated statement has transformation calculation and an origin set containing a generation event. A demand may now say that no ancestor may be a generation, which is the demand the whole work exists to make expressible and which the previous model could not state. The summary and the inference separate on the semantic relation while sharing a mechanism, where before they were identical in every field the record carried.
Causal history is not evidential ancestry
One case in the test is not repaired by typing alone and forces a further distinction. A generated hypothesis prompts an investigation, an independent measurement is taken, and the measurement establishes the proposition. Carrying the generation as evidential ancestry forever would refuse the claim under any demand excluding generated ancestry, though the claim now rests on a measurement that would stand had the hypothesis never existed.
The generation caused the claim to be investigated. It does not support it. So the model distinguishes what a claim descends from causally from what its basis actually rests on, and the origin set carries the latter. An origin enters a derived claim's set when the derivation depends on it evidentially, and a claim established by fresh acquisition begins a new set rather than inheriting the set of whatever suggested it.
This is a judgement the construction step must make and can get wrong, and it is the obvious place to launder deliberately: declaring a generated ancestor to be merely causal would strip it from the set. The model does not prevent that, and the honest statement is that it moves the laundering opportunity from an invisible place to a declared one. For that to hold, the declaration has to travel. A construction that drops an origin records an elision in the basis record itself, carrying which origins were removed with their kinds, on what ground they were judged causal rather than evidential, what performed that judgement, and what basis warranted it. An elision recorded only in a local audit log would be invisible at the receiving boundary, which is the one place the discipline claims to work, and the laundering would be declared at construction and undetectable on arrival. A demand may then refuse claims whose construction elided origins of kinds it cares about, which is a condition a consumer can state and check.
What this costs
Three things, and the first two are the reason the change was made at that point rather than after the library existed.
Every part of the artefact that read the single field was affected. The inheritance rule was restated over two fields rather than one, demands are authored against both, the admissibility method's single comparison became two, and record construction maps a source to an acquisition and a transformation rather than to a single kind. That is the rigidity the design method warned of, arriving as predicted, and it was cheaper at that point than after an implementation had fixed a public interface around a single enum.
Typed origins are larger than identifiers, which adds to the growth already recorded as a residual to watch.
And the evidential-against-causal distinction introduces a place where a construction can be wrong or dishonest that the previous model did not have. That is a real cost and it buys the ability to state case seven correctly, which the previous model could only handle by refusing a claim that a measurement had established.