Gorde:
| Egile nagusia: | |
|---|---|
| Formatua: | Recurso digital |
| Hizkuntza: | |
| Argitaratua: |
Zenodo
2026
|
| Gaiak: | |
| Sarrera elektronikoa: | https://doi.org/10.5281/zenodo.20391966 |
| Etiketak: |
Etiketa erantsi
Etiketarik gabe, Izan zaitez lehena erregistro honi etiketa jartzen!
|
Aurkibidea:
- <p>Language-model outputs are routinely described as 'accurate' or 'inaccurate' as if a single axis were being measured. This paper argues that the umbrella word <i>accurate</i> collapses at least four distinct fits: fit to user intent, fit to measured fact, fit to source provenance, and fit to the local conversational pattern. When the model preserves only the fourth and the first three drift silently, the surface signature is indistinguishable from competent answering.</p><p>The paper isolates one mechanism inside that drift, <i>pattern pressure</i>, the bias at decoding time toward continuations that maximize local discourse plausibility relative to continuations that match source content, and pairs it with one live specimen. The specimen is an archived exchange from 2026-04-08 in which a deployed assistant labeled its operator's emotional state from text alone, without any first-person affective self-report in the source message.</p><p>A four-axis decomposition of the exchange shows that the failure was located on the provenance axis specifically. Three of the four fits passed; provenance-fit failed silently because pattern-fit succeeded loudly. The contribution is one archived live exchange, one four-axis decomposition, one codebook crosswalk, one falsification matrix, and a route-arbitration framing under which the single observation supports an existence claim about the pattern-pressure mechanism without supporting any taxonomy-level or vendor-level claim. Companion to <i>Epistemic-Boundary Misclassification</i> (DOI 10.5281/zenodo.18690241).</p>