The model

What a record becomes when habitat reads it, in the record's own terms. Short on purpose. Every claim here is measured on the two public corpora (the Adams and Jefferson letters, 1,103 acts; a climate corpus, 356 acts) and reproduces from their bytes.

An act, not a string

What enters is read as an act: an actant doing something through a predicate. The actant is the subject of the predicate's assertions. The predicate is a process that connects actants and transforms them, and is transformed by them. Neither is a word. "I" in a letter from Adams to Jefferson is Adams, in that letter, on that day; the same word in the next letter is a different actant with the same base identity.

Seventeen eventualities

Each act is held as seventeen observed assertions, in two blocks that touch:

The actant's five. Agency capacity, stability, influence, boundary definition, resonance. Derived from proto-role theory (Dowty): how the actant participates in the process.

The predicate's twelve. Aspect (telicity), modality (causality), temporal extension, relational dynamicity, intensity (agentivity), viewpoint aspect, transitivity, thematic role depth, agency distribution, animacy, boundedness, modality type. Derived from eventuality theory (Bach, Vendler) and verb classes (Levin): what the process asserts about itself.

The two blocks are a direct sum, 5 ⊕ 12. Their covariance across a corpus holds a 5 × 12 cross-block: where the actant and the predicate move together. That cross-block, minus what the extraction rules would have produced from role and eventuality class alone, is the grammar the corpus earned rather than the grammar the constructor laid down.

A field, not a fixed geometry

Acts accumulate into a covariance, one per corpus. The metric habitat reads couplings through is that covariance's inverse, ridged by a named constant (0.01) so a counterparty reproduces the same digest. Nothing is fitted, trained, or tuned: the corpus sources its own metric, the way a record sources its own geometry. Two readers of the same corpus hold two metrics, and the difference between their orderings is a measurement, not an opinion.

Beside the metric, the co-movement of the seventeen is reported as warp: the marginal correlation each eventuality carries with the others. Warp and coupling are two objects and are never substituted for one another.

A second space, and the relation between them

Each act also carries a 768-dimensional embedding of its text. The two spaces are not two feature sets. Each of the seventeen is laid on the record as a key from the act's structure plus the act's relation to a named direction in the embedding, and that relation is admitted only where the structural field has not yet formed its own differential. What is observed is the relation, per act, per eventuality. The system asks what a relation does to a coupling, never how large it is against a value.

What a coupling is made of

When two passages couple, part of what they share is a condition: the same eventuality class, the same side of the actant's role structure, the same level on the dims the record does not move. That part is the same for every partner on that side. The rest is the pair's own: the cells of the metric that belong to these two acts and no others. The shared part is a floor almost every pair stands on, and it is larger than the coupling itself; the pair's own part is what ranks, negative for almost every pair and least negative or positive for the partners the field surfaces. The reader's surface prints both, separately, with the pair's own part signed, to be read against the corpus.

Identity, all the way down

A document names itself: its identity is the source the corpus declares (a letter's date, author and addressee; a report's file), never a random string. An actant or predicate has a base identity (its text, corpus-wide) and an instance identity (that base, in that document, with the type the document gives it). A composition's identity is derived from its session, its document, its two instances and its position, so the same act in the same document is the same composition on every re-ingest, and a new extractor makes a new version of it, stamped with the extractor's own version. Where a caller has no document, the id says so aloud.

What it refuses

To rank, advise, predict, or decide what an act means. To fit anything to look good. To keep a fallback that decides a dataflow. To say "never" about what it has not measured. Where it published something it later could not stand behind, the retraction sits beside the claim: every correction.

Canon is the code: the seventeen names are read from canonical_dimensions.py, and the numbers on this site reproduce from the committed stores.