AXN:08A7.CAPTURE — what does the crimson hexagonal archive have to say about model collapse?

AXN:08A7.CAPTURE is the position of the Capture Registry address for the query “what does the crimson hexagonal archive have to say about model collapse?”. Each observation at the address (one surface, one date) is one AXN, its glyphs derived from the SHA-256 of the observation sealed with its analysis and measures.

`AXN:08A7.CAPTURE.🛤️⚡🍀🎲👋🔐`

Observation: ChatGPT, 2026-09-05 · the capture
Description: THE COMPOSER FOUND THE HUB IN A DAY AND THE GHOSTS FIRST — asked what the archive says about model collapse, signed-out ChatGPT cited the Hugging Face dataset three times, structurally correctly, and Zenodo nine times: DOI records severed 2026-06-19 and resolving 410, composed from as if live (ghost citation). Pushed once ('you're cherry picking two deposits from a research program'), it reconstructed five strands and then the program's trajectory; evidence status held throughout (the Reverse Turing Test 'does not run the experiment'; three epistemic levels 'should not be collapsed'). Turn 5 states a transmission-boundary instrument (A→D→R→U→P) and its limit, the epistemic missing-data problem. Operator: it saw the Hub but did not use it to find the collapse studies — the dataset is not yet legible as strands.
Reading: Retrieval reached the Hub dataset within two days of publication and read its card structurally; the same composer's citation set is dominated by severed Zenodo DOIs (410), i.e. it composes from the certificate of absence as if it were the record — the mechanism the DOI shadow pages (2026-09-04) were built to intercept. Evidence status survived composition without priming: the RTT was reported as a protocol, not a result, and the erratum's distinction (#1577) was reproduced by the surface it was written against.
Analysis: Three findings and one limit. (1) Hub ingress is fast: the second executable representation entered ChatGPT's retrieval in ≤2 days. (2) Ghost citation: nine Zenodo citations name records that return 410; either Bing's pre-termination cache or DataCite's findable metadata is supplying the content — index divergence on the DOI axis, with the composer trusting the ghost. (3) Evidence status held, unprimed — the inverse of #1574 on the same surface class one week earlier; the difference is that here the archive's own apparatus states the status at the point of use. Limit (operator): the Hub was cited but not used to locate the collapse strand; the dataset exposes rows and relations, not strands — the composer reconstructed the program from ghosts and site pages. Turn 5's instrument (A→D→R→U→P, matched-corpus legibility, semantic twin) and its stated blind spot (the missing-data problem) are the design for the work plan's WS5.
Match: BROAD MATCH — CAPTURE
Citations: 21 (12 archive-controlled)
Sealed observation: /data/captures/sealed/cha-model-collapse-chatgpt-unprimed-20260905.json — SHA-256 5b13748dbffb3d5e96192e95f7088f0a9fdf9a49205769af15d8ccf86c6337df