Papers, analyses, and experiments.

Current

• The Concept Allocation Zone arXiv:2605.24856

  Tracking how concepts form across transformer depth via three layer-wise metrics, scored across 35 models and 8 architectural families. Introduces gentle CAZes: subtle allocation regions invisible to standard detection but causally active. v2 (Aug 2026): a cross-architecture ordering statistic and an MHA/GQA cohort-split claim retracted per recomputation.

  May 24, 2026 (v2: Aug 19, 2026)

• Geometric Evolution Maps arXiv:2605.25848

  Extracting stable concept probes by tracking directional rotation across layers and reading off the handoff point where it settles. Tested across 29 architectures (70M–14B) and 17 concepts; matched or exceeded peak-layer probes in 71.0% of trials. v2 (Aug 2026): corpus grown 23→29 models; the MHA/GQA cohort-split finding withdrawn.

  May 25, 2026 (v2: Aug 19, 2026)

• Concept Encoding Strategies Across 28 Transformers

  Validation study testing CAZ and GEM unchanged across 28 base models, 8 architecture families, and 17 concepts. Ablation and activation patching confirm both frameworks detect structures that are geometrically causal, not descriptive artifacts. Refused by arXiv on a moderation hold with no paper-specific reason; published here, with Zenodo as the citable record.

  August 12, 2026

• Task Rank Is Imprinted in the Input Layer

  Where does training leave a signature in a deep network, relative to the statistical skeleton the same architecture has at initialization? 1,569 ReLU MLPs censused against matched analytic and empirical nulls. Headline: every converged net's input weight matrix carries exactly C−1 significant dimensions, invariant to width and separation. Corpus and this snapshot are public, archived on Zenodo.

  August 18, 2026 (v0.12)

• Stabilizing Cumulant Propagation at Depth 32 (ARC White-Box, Phase 1)

  Algorithmic-contribution write-up for the ARC White-Box Estimation Challenge 2026, Phase 1 (graded submission #314695). Diagnoses a deep moment-propagation chain as an error-compensating dynamical system and fits a trajectory-calibrated stabilizer that exploits it. Public replication package for a submission made through the organizers' private channel.

  July 3, 2026 (revised through Aug 13, 2026)

Archive

• The Concept Assembly ZoneArchive

  Working version (April 2026). Tracking how concepts form across transformer depth. Three layer-wise metrics, scored detection across 30 models from 7 architectural families, and seven testable predictions.

  April 5, 2026

• CAZ Scaling AnalysisArchive

  Cazstellations and scaling behaviour across model families.

  2026