46. The chimera's three donors are not three sources

Context

ADR 0045 built two of the three pos rows and declined the third in writing: chimera "is kind: constructive over three source texts, which is a genuinely new parameter shape and deserves its own record rather than being carried in on the back of this one". Issue #22 says the same thing earlier and more plainly — the row is blocked because "SourceParams carries one source and chimera needs three". This is the record ADR 0045 asked for, and its first finding is that both statements are false.

SourceParams does not mean a text this procedure reads. It means the text this one was made from — homoconsonantism's original, n_plus_7's original, homosyntaxism's original. A chimera is made from its frame: the text whose word order, function words and punctuation survive the operation. The frame is the source, SourceParams already carries it, and parse_apply_params already supplies it from apply's own text argument, so a caller who passes source to apply is refused as they are for every other source row.

The three donors are not sources in that sense at all. Nothing of a donor's structure reaches the output; what reaches the output is words. A donor is lexical stock — a word list that happens to be written as a text, and read as one only because the tagger needs sentences to tag. Two texts made from the same frame with different donors are the same operation on the same original; two texts made from different frames are not.

So the question the issue framed — how does one procedure take three sources? — does not arise. The real question is what to call three word lists beside an inherited source, and that is an ordinary parameter question.

Decision

ChimeraParams(SourceParams) adds three plain string fields: nouns_from, verbs_from, adjectives_from. core/base.py is untouched. No mixin, no SourcesParams, no change to the spine, and nothing about this row is reflected in a shared shape.

Two alternatives were considered and rejected.

An ordered sources: list[str]. Position would encode which donor fills which word class, and params_schema() — the surface every non-Python caller reads, and the only documentation an MCP or CLI caller gets — cannot say that. A caller would have to learn from prose that element 1 is the verbs, and a caller who got the order wrong would get a plausible wrong answer rather than an error. Role-keyed names say it in the schema, and the test test_a_content_word_from_the_wrong_donor_is_a_violation pins the distinction: river is a noun sitting in verbs_from, and a checker treating the donors as one bag of words accepts the text that one rejects.

A SourcesParams mixin in core/base.py. It would generalise a shape with exactly one caller, and it would generalise the wrong thing — a list of sources is not what this row has, per the whole of the Context above. Naming it in core would make the false framing permanent and invite the next row to inherit it. source_compare's docstring already records the house rule: extract from real callers rather than ahead of them.

The premise being wrong is itself the decision's content, so it is recorded in the row's notes: and in the module docstring as well as here. Issue #22's grouping of "the eight checkability: source rows" behind one parameter contract was already contradicted once, by ADR 0045's finding that the other six are gated on capabilities rather than on a contract. This is the second half of the same correction: the one row that looked like it did need the contract did not either.

Consequences

Nine catalogued rows minus three: pos unblocked all three of its rows, and no parameter-shape work was needed for any of them. What issue #22 recorded as a shared contract problem was, in the end, one capability and no contract.

A row may take several texts without them being sources. That is now demonstrated rather than asserted, and the next row wanting a word list as a parameter has a precedent that does not reach into core. What it must do instead is say in each field's description what the field is for, because the schema is the only place a non-Python caller reads it.

The caller-facing caveats live in the definition, not only in notes:. describe() returns the definition, the prompt hints and the param schemas; it returns notes: only under scholarly=True, a flag the MCP tool's own docstring describes as being for "the form's source, attribution and history", and which the CLI's human-readable output ignores entirely — denckring describe chimera --scholarly prints exactly what denckring describe chimera prints. So by default no caller sees a row's caveats, and a caller who wanted them would have no reason to look under provenance. The first draft of this row put four rulings, an estimate caveat and the fact that apply can refuse entirely in notes: and the module docstring, and left a definition that overclaimed twice: it said the content words are "replaced by" the donors' (ruling 2 permits a drawn word to equal the one it replaced, so a frame whose content words all appear in its donors satisfies itself) and it said "three different source texts" (nothing requires or checks that the donors differ). The definition now says what is true, following iambic_pentameter, which carries its own caveat — "the checker verifies strict stress" — in the definition for the same reason. That describe() omits notes: at all is a cross-cutting gap, shared with homosyntaxism, and is filed separately rather than fixed here.

seed does less than the parameter list suggests, and each candidate says how much less. A donor pool is the tagger's reading of the donor, and a word that reads as its class in the donor need not read as that class where it is drawn. Measured: pools of four NOUN, two VERB and one ADJ over The quick boy opened the door. give fourteen distinct outputs over twenty seeds, with the adjective bright in all twenty. Candidate.metrics therefore carries target_positions and forced_positions, the second counting the positions of that output where exactly one donor word survives tagging — the undecided_words / PosTag.known / Production.truncated idiom applied to choice rather than to confidence. Only the sentence a position falls in is re-tagged per trial, which on a 680-word frame is 0.34 s against 21.1 s for whole-text re-tagging, for the identical count.

Known remaining

apply's search is greedy and incomplete, and a refusal is not a proof of unsatisfiability. _produce draws a word per position, re-tags its own output and redraws only the positions whose class did not survive, from the words not yet tried there. That is coordinate descent: a redrawn position keeps every other position's word, and a word once tried at a position is never tried there again. Because a word's class is a fact about its context, a filling that this walk cannot reach may still exist.

Measured, not supposed. Frame The old lighthouse kept a steady light. Sailors watched the dark water and counted every slow turn. with nouns_from A grocer weighed salt., verbs_from They carried the ladder, painted the shutters and left. and adjectives_from A thin mist hung over the green fields. — apply at seed 0 raises NoCandidateWord, and The thin grocer carried a thin salt. Grocer carried the thin salt and carried every green grocer. satisfies the same parameters with score 1.0. A sweep of 400 random (frame, donors) pairs gave 21 refusals of which at least 8 were false, so the false share is 38% or more.

Not fixed, and the ruling is that it should not be. A complete search over three pools is exponential and this row does not need one; an incomplete search that says so is honest, and a backtracking search with a budget would be a second thing to explain and would still refuse satisfiable cases. What was wrong was the claim. The NoCandidateWord message now says that this greedy search found no word and that a filling it cannot reach may still exist, the module docstring says the search is incomplete, and test_apply_refuses_when_its_greedy_search_runs_out_at_a_position holds the refusal and the satisfying filling side by side. docs/audit/chimera.md records the measurement. If a caller ever needs the refusal to mean unsatisfiable, that is a new decision and a different search.

A third check/apply asymmetry, from ruling 2. Where every pool collapses to the word already standing in its position — frame The cold wind blows. with pools {wind}, {blows}, {cold} — _produce has nothing else to draw, returns the frame, and the spine raises DegenerateOutput, while check reports that same text satisfied. Both are right: the text does satisfy the constraint, and allow_identity is how a caller asks for the degenerate case. Documented in ruling 2 and tested, rather than made to disagree quietly.