Finite affine core
Finite affine bijection and operator-count results remain distinct from construction-specific Mirror-13, torus, genomic, and geophysical computations. The Level-1 public record is 10.5281/zenodo.22036769.
The current VARZIN record,
grounded in the latest public artifacts.
This atlas uses the latest relevant Zenodo manuscript and its version-matched code/data package as the primary source for current scientific interpretation. Older site prose and earlier versions remain visible for provenance, but they do not override later evidence for the same claim.
Source rule
Latest relevant public evidence first; historical records preserved, not silently rewritten. A newer paper supersedes an older interpretation only within the same experimental line.
What the project currently establishes within scope.
LUXVAR structure
v2.2 reports non-random morphology and phonotactic distinctiveness, while its semantic-axis tests are negative. The historical “AI-Recoverable” wording is retained only for citation continuity.
Recovery ≠ composition
Later experiments show targeted structure can be recoverable after explicit intervention, yet the tested held-out composition procedures remain negative.
Core-30 is the anchor, not the whole current research stack.
Core-30 reference lexicon
Fixed historical/canonical reference set for early clustering, semantic-axis, and adversarial morphology experiments.
Broader LUXVAR v2.2 root layer
The manuscript reports 801 stable roots and an approximately 1.3-billion-form generated combinatorial corpus. Generated forms are design outputs, not 1.3 billion empirically validated words.
Appendix E scale/generalization benchmark
Fresh 20-root × 15-prefix variant. The reported pooled held-out-root result is ARI 0.992 ± 0.007 under the easier same-script/same-order condition.
Appendix F true group-position benchmark
5 roots × 6 prefixes × 12 positions with opaque position encoding and shuffled-label control. This tests targeted position classification, not composition.
VARZIN V2 frozen dataset
12 structural positions, four synthetic morphemes, 30 contexts, 24 train contexts and 6 held-out contexts. This is a separate experimental dataset, not an enlarged Core-30 dictionary.
The latest model evidence is mixed by design—and that distinction matters.
Positive, scoped results
- Targeted projection/training substantially improves recovery of the intended structural labels in the documented intervention experiments.
- v2 reports trained Mistral/Qwen ARI about 0.957 and 0.880 in the stated stress configuration.
- Appendix F reports true group-position recovery above shuffled-label controls across GPT2-small, Mistral-7B and Qwen2.5-7B.
- VARZIN V2 finds the synthetic position variable highly linearly decodable from frozen Qwen hidden states: mean balanced accuracy 0.9666 across 29 indices, with 25/29 at 1.000.
Negative, equally important results
- Direct frozen-model semantic/orbit recovery in Track D was near chance under that protocol.
- SEM-001/SEM-002 did not establish independent semantic-axis recovery from LUXVAR word forms.
- v3 composition pilot: seen-pair MLP ≈0.996, but held-out TRUE 0.106 vs SHUFFLED 0.281 and WRONG_OP 0.175.
- VARZIN V2 Phase 2: PASS 0/58, AMBIGUOUS 0/58, FAIL 58/58 under the preregistered decision rule.
Current interpretation: structural information can be strongly decodable or recoverable under explicit procedures without thereby becoming a demonstrated substrate for systematic composition on unseen combinations.
The code packages are part of the evidence chain.
Current publication pages expose the exact public DOI, full-text artifact, code/data package where present, file size and checksum. The v3 ZIP contains the experimental chain through scripts 01–25, including the 300-word generator/scale test, the preserved rejected leaking design, the corrected true-position control, and the composition diagnostic chain. VARZIN V2 exposes the integrated PDF plus the large “All code Phase.zip” package and documented frozen hashes.
Inventory caveat: the verified public Zenodo inventory for LUXVAR v2.2 currently exposes the PDF only. The site does not claim a version-specific ZIP for that record unless one is actually present in the public deposit.
What the current evidence still does not establish.
- Independent semantic emergence from LUXVAR word forms.
- A universal theorem that decodability never implies composability.
- Systematic composition for the tested v3/V2 procedures.
- Generalization of these results to all models, natural languages, operations, or readout architectures.
- Physical, consciousness-field, quantum, biological-frequency, or non-human-origin mechanisms.
- Peer review or independent external validation merely because a record is publicly deposited.
Which record controls which claim.
10.5281/zenodo.22036769
Finite-affine/software stack and companion computational audit.
10.5281/zenodo.22115483
Morphology, phonotactics, generator scale and semantic-axis tests in v2.2.
10.5281/zenodo.22101179
Dedicated direct frozen-model semantic/orbit audit.
10.5281/zenodo.22287006
Latest projection-head series record: remediation, 300/360-word benchmarks and composition pilot.
10.5281/zenodo.22679978
Integrated Phase 0–2 frozen-Qwen representation recovery versus systematic composition.