Morphemes vs. Manifolds: Diagnosing Structural Blindness and Morphological Hijacking in Large Language Models via Finite Affine Orbits

Research question and current reading.

Under the stated Semantic Starvation protocol, the tested Mistral 7B and Llama 3 8B systems did not recover the designed orbit structure. The result is model- and protocol-scoped; it does not justify a universal claim about all language models.

What was actually tested.

• LUXVAR Core-30 arranged into five designed affine-orbit target groups with adversarially shared surface morphology.
• Semantic Starvation removes natural-language explanations before blind five-cluster recovery.
• Hungarian alignment handles arbitrary cluster labels; Adjusted Rand Index provides chance-corrected partition recovery.
• Thirty independent runs per model with seed/temperature variation; random-partition baseline reported alongside model results.

Strongest supported findings.

• Mistral 7B Instruct: mean ARI = −0.0337 ± 0.039; mapped accuracy = 33.3% ± 4.2%.
• Llama 3 8B: mean ARI = 0.0152 ± 0.028; mapped accuracy = 50.0% ± 5.1%.
• Random baseline: mean ARI = −0.0583 ± 0.045; the tested models remain near chance against the designed orbit labels.
• Observed errors preferentially follow shared spelling/morpheme cues, motivating the “morphological hijacking” label for this tested setup.

Negative findings are part of the result.

• The intended semantic/algebraic axes were not recovered by the tested frozen models under the reported protocol.
• The earlier inter-model agreement figure from the LUXVAR line cannot be used as evidence of axis recovery.
• The planned VPE-001A human comparison is future work in this record, not evidence about human cognition.

What this record does not establish.

• Core-30 is a small, synthetic, researcher-designed lexicon.
• Only the tested models, prompting protocol, clustering task, and evaluation procedure are directly supported.
• Some wording in the deposited manuscript is broader than the evidence; the current VARZIN page deliberately narrows those claims to the tested setup.

Full text, code, data, and provenance.

• The public record includes the exact PDF, README, CITATION.cff, manuscript DOCX, participant-research checklist, license, and a package-recreation script.
• The uploaded PDF supplied for this website audit is byte-identical to Zenodo varzin-track-d-scientific-paper-v3.pdf (verified by MD5).

Zenodo REST file inventory and checksums re-verified 2026-09-14; uploaded paper copy exactly matches the public Zenodo PDF.

Canonical version-specific record.

Verified repository identity.

The Zenodo REST API description retrieved on 2026-09-14 appears internally inconsistent with this record’s deposited title and references DOI 10.5281/zenodo.22036769. VARZIN therefore does not use that description to characterize this record; the title, DOI, version, date, creator identity, and resource type were verified independently from the record metadata.

Keep this deposit distinct from later interpretation.

The record anchors the negative-result side of the VARZIN language-model program: structural regularity can be detectable while stronger semantic interpretations remain unsupported under the tested procedures.

Version v1.0.0 is retained as a fixed, citable public deposit. Later VARZIN pages may refine interpretation, but they do not silently rewrite this versioned record.