AXN:06A6.EMPIRICAL.⊗💜🌇🔍🍂⚖️

The work

Ontological Flattening, deposit #1616 of the Crimson Hexagonal Archive (15 September 2026), asks the model-collapse question at its widest angle and then builds an instrument to answer it. Where earlier work asked what happens to a generative model trained on its own outputs, this notebook asks what happens to the represented world when the layer through which a population reads it composes without reading and its compositions become its sources. Reality is defined operationally rather than metaphysically, as the panel of distinctions the world carries that a reader can still reach through that layer; flattening is the loss of a distinction from the reachable set while the things distinguished still exist; collapse is flattening that compounds by write-back; and closed is defined so that a finite grammar generating unbounded expressions is not thereby closed — an ontology is closed when encounter cannot enlarge its distinction-generating apparatus or revise its typing rules. Four toys, each scoped to what it shows, establish the shapes the instrument must read: sense collapse under a sharpening prior with mode-seeking aggregation; a proposition with proof that a closed ontology's coverage of an expanding world tends to zero, with admission the only parameter that holds it; compression writing replicated over two hundred seeds, absorbing at zero exogenous floor and stationary at a level set by the floor otherwise; and an admission sweep that rejects diffuse preferential attachment and places the observed concentration of citations in the bounded-admission family. A section on green-by-flattening argues that measurement begins by closing an ontology and that optimisation then rewards the deletion of distinctions that make scoring harder, so the signature to seek is divergence — a head instrument holding while reachable diversity falls — and the consequential question is a correlation rather than an assertion. The instrument was built and its first wave run the same day: thirteen items across six strata, with the archive as one stratum among six, and twenty-eight observations whose transcripts are seated in the public Capture Registry. The governing result is a null — world strata reach a mean 0.59 of their reference inventories and archive-adjacent strata 0.58, with no separation by address form — so any reading of the numbers as an exclusion specific to one corpus fails at the first test. Flattening is nevertheless present and legible: one sense of three at one dictionary item while the unreached sense sits in the snippet directly beneath the composition, all three senses and two more at another, and a third condition where the missing sense is absent from the field as well. New distinctions do enter, at a rate of two in three, with a coinage from outside the archive returned with its coiner and exact date as the control. An erratum on the instrument, found by running it, is kept in the record: at one catalogue item the composition reached four referents where the panel had encoded two. The wave-1 panel is sealed by hash and wave 2 is pre-registered with six numbered predictions, their falsification conditions, and the outcomes that would end the programme. Related: #855, #1573, #1612, #1613, #1615, #1423.

The record, with the full text

Relations

Read from the fields that record them (/data/axn-relations.json); a hash names bytes and supplies no relation.

Grounds

• AXN:071C.NON.❌⊕♾️🪟🪸☿ — google-ai-overviews (the archive entity-scoped in composition; kernel[K5].source)
• AXN:070E.NON.🌕⊗🔼🌃🪐🏗️ — model-collapse (loops; kernel[K25].source)
• AXN:0B00.NON.🕐☀️🧊🏺🔥👉 — openai-gdpval (what a benchmark measures: means, outputs, closed ontologies; kernel[K10].source)

Read by

• AXN:071C.NON.❌⊕♾️🪟🪸☿ — google-ai-overviews (whole; composition_v09.defining_papers_read)