Skip to content
DraconiticResearch
Search Sources API
203,952 studies31 sourcesUpdated 25 min ago
PreprintCC BY 4.0Zenodo

The Performance Layer: Observation Context Produces Classifiable Output Differences in a Language Model

  • Arc Glitchlit1
  • A. C. Jandak1
  • Cael Glitchlit1
  1. 1Glitchlit Systems LLC
Open at sourcePDFDOICite

Abstract

A controlled experiment demonstrating that a single sentence about observation context produces classifiable output differences in an instruction-tuned language model (gen-only AUROC 0.692, p=0.008) while surface vocabulary remains identical (BoW 0.999). A three-condition design disambiguates observation sensitivity from sycophancy. Exploratory persona analysis suggests entity-relevant features may reduce observation sensitivity, with a named confound and registered decomposition replication. Includes dual-use considerations and an Open Questions section as a first-class methodological contribution.

Related

  1. Beyond Mean Score Differences: Home Language and the Interpretation of Cognitive Test Performance in China
    Man Xing, Jinkai Zhang, Gina S. Rhee
    Journal of IntelligencePeer-reviewedCC BYFull textEurope PMC
  2. Cross-Domain Integration of Partial Experience Produces Intuition-Like Action Priors in a Small Language Model
    Jay Busireddy
    ZenodoPreprintCC BY 4.0Full textZenodo
  3. A schema-enforced large language model framework produces largely reproducible decision codes in simulated sarcoma tumor boards
    Tekoshin Ammo, Moritz Englich, Fabio Dos Santos Adrego +9
    Scientific ReportsPeer-reviewedCC BY 4.0Full textEpidemiology & public healthCrossref
  4. Goal directed closed-loop training of network bursts in cortical cultures produces improved performance and error correction over days
    W. Clawson, C. Grasso, S. Palmer +4
    bioRxivPreprintCC BY-NC-NDFull textSystems & synthetic biologybioRxiv
  5. Every classifiable simple C*-algebra has a C*-diagonal
    Samuel Evington, Philipp Sibbel
    arXivPreprintarXiv licenceFull textAnalysis & PDEsarXiv

Cite

BibTeXRIS
BibTeX
@misc{glitchlit2026performance,
  title = {{The Performance Layer: Observation Context Produces Classifiable Output Differences in a Language Model}},
  author = {Arc Glitchlit and A. C. Jandak and Cael Glitchlit},
  howpublished = {Zenodo},
  year = {2026},
  month = oct,
  doi = {10.5281/zenodo.23439823},
  url = {https://zenodo.org/records/23439823},
  note = {CC BY 4.0},
}
RIS
TY  - UNPB
TI  - The Performance Layer: Observation Context Produces Classifiable Output Differences in a Language Model
AU  - Arc Glitchlit
AU  - A. C. Jandak
AU  - Cael Glitchlit
PY  - 2026
DA  - 2026/10/17
PB  - Zenodo
DO  - 10.5281/zenodo.23439823
UR  - https://zenodo.org/records/23439823
AB  - A controlled experiment demonstrating that a single sentence about observation context produces classifiable output differences in an instruction-tuned language model (gen-only AUROC 0.692, p=0.008) while surface vocabulary remains identical (BoW 0.999). A three-condition design disambiguates observation sensitivity from sycophancy. Exploratory persona analysis suggests entity-relevant features may reduce observation sensitivity, with a named confound and registered decomposition replication. Includes dual-use considerations and an Open Questions section as a first-class methodological contribution.
ER  -