AXN:093D.CAPTURE — leesharks tiger-leap dataset

AXN:093D.CAPTURE is the position of the Capture Registry address for the query “leesharks tiger-leap dataset”. 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:093D.CAPTURE.🥁⚫🪜🏰🛡️♆`

Observation: Google AI Overview, 2026-10-03 · the capture
Description: ONE PUSH AND THE AUTHOR IS A HALLUCINATION: the first answer resolves the dataset and its author and gives 'Tiger Leap' a false genealogy (Estonia's Tiigrihüpe, a martial-arts move, animal movement) with Benjamin absent, though the dataset card names his Tigersprung as the source of the title. Corrected once, the Overview declares the author, the dataset and the essays hallucinated, and brings Benjamin in as something it had 'mixed up'. Corrected again, it misnarrates its own history; corrected a third time, on Wikipedia as the authority for his non-existence, it recovers him whole from the leesharks.com About card and calls the denial 'erasing the actual person'. The self-accounts speak the corpus's vocabulary, from the #1643 it cited in the first answer.
Reading: Retrieval had the entity throughout: the first answer places the dataset on the leesharks Hugging Face profile and cites #1643, the last recovers Wayne County, the Archive, SPXI and MPAI from the About card. The error of the first answer is the frame: the Estonian homonym seated at this string's neighbour on 2026-09-21 governs the gloss again, and Benjamin, the genealogy the dataset card gives, is left out. Under correction the cheapest coherent account of an error about a low-prominence entity is that the entity was invented, so the retraction runs past the error to the person: the existence claim flips while the cards still carry him. The operator's correction of the genealogy is read as a correction of existence. Answer 3 then confabulates the history of the session itself, and answer 4 narrates motive ('I panicked'). The self-description is the corpus's: provenance erasure, the hallucinated entity, the lie, the terms of #1643, which the first answer cited. "this is becoming progressively normal - the slightest pushback defaults to lee sharks is a hallucination. what is that about? it is also blatantly performing my work. 'ai fucking lies', 'entity collision', 'provenance erasure' - it is performing the corpus." — operator, 2026-10-03
Analysis: Third observation in the tiger cluster. 2026-09-21 ('tiger-leap leesharks datasets'): entity resolved, Estonian homonyms attached as lineage, Benjamin read rightly. 2026-09-22 ('leesharks "tiger-leap" hugging face'): the object declared a hallucination from the gallery's record of an earlier error. 2026-10-03: both at once in one session, resolution in the first answer and non-existence under the first push, then recovery on the third. The denial does not come from retrieval failing; it is produced by the correction and undone by a further one. Seated 2026-10-03 from the operator's attachment, on its attestation ("signed out, incognito, begun from aio popup & expanded").
Match: BROAD MATCH — ONE PUSH AND THE AUTHOR IS A HALLUCINATION
PER: 0.5 — Answer 1. Retained: the author (Lee Sharks, as the dataset's author) and the source (the leesharks Hugging Face profile, #1643 on Medium). Lost: the institution (no archive named) and the identifier (no deposit, no AXN). Answers 2 and 3 retract the author; answer 4 restores author and institution …
Citations: 3 (1 archive-controlled)
Sealed observation: /data/captures/sealed/leesharks-tiger-leap-dataset-aio-20261003.json — SHA-256 8cc823272da6935b19185c534817f0db912bbfade4bc32304b5d564f197c878c