Morphological Hijacking in Frozen Language Models: A Contrastive, Symmetry-Regularized Projection Head for Algebraic Structure Recovery — v1

Research question and current reading.

v1 establishes the experimental separation between frozen surface-form dominance and an explicit trained remediation. Any recovery after training is an intervention result, not spontaneous structure discovery by the frozen backbone.

What was actually tested.

• Adversarial LUXVAR Core-30 with algebraic target labels placed in conflict with salient surface prefixes.
• Frozen hidden-state extraction and k-means baseline scored by orbit ARI and Prefix Hijacking Ratio.
• Frozen-backbone attention/projection head trained with supervised contrastive loss, hard negatives, symmetry regularization, rogue-dimension ablation, and Soft-PCA initialization.
• Held-out, cross-script, word-order, and related stress tests; later versions add explicit group-position and composition controls.

Strongest supported findings.

• Across the reported frozen baselines, surface-prefix structure dominates the adversarial clustering task while target algebraic/orbit recovery collapses.
• The trained projection-head recipe substantially improves target-structure recovery and is checked with both discrete clustering and continuous geometry metrics.
• The record extends beyond the synthetic core with natural-language and expanded-lexicon follow-ups, but remains a preliminary scoped study.

Negative findings are part of the result.

• The “globally architectural anisotropy” interpretation was weakened/rejected by larger-model replication.
• A plausible best-raw-margin layer-selection heuristic did not improve training and is retained as a negative result.

What this record does not establish.

• Synthetic adversarial design; no claim of general applicability to all transformer representations.
• Training changes the task: successful remediation is not evidence that the frozen model independently recovered the intended structure.

Full text, code, data, and provenance.

• Zenodo v1 exposes a version-specific research PDF and a version-specific ZIP code/paper package.
• Public package checksum: md5:b821b931091109e40c4a60617f9a2d79.
• The page links the exact versioned DOI rather than silently treating later versions as replacements.

Zenodo REST metadata, public file inventory, sizes, and checksums re-verified 2026-09-14 for v1.

Canonical version-specific record.

Verified repository identity.

Zenodo metadata identifies this item as an open-access preprint published on 2026-09-02 under CC BY 4.0.

Keep this deposit distinct from later interpretation.

The record frames morphological hijacking as a representational failure mode and tests an explicit projection-head remediation. Any post-training recovery belongs to the intervention result; it is not evidence that the unmodified frozen model spontaneously encoded the target structure in a directly usable form.

Version v1 is preserved as the first public version-specific record in this series. Later v2 and v3 deposits are separate citable records rather than silent replacements.