Prannesshkva/ISOM-R1-Reasoning-1.5B-Instruct-Beta

Refactor: Register standalone native ISOM-R1 class architecture and auto_map

Refactor: Register standalone native ISOM-R1 class architecture and auto_map

Docs: Update cross-lineage table link to ISOM-R1-Coder-16B-MoE-Beta

Fix(1.5B): Clear holographic table chunks/proj/step cache in reset() to prevent cross-session memory leak

Add authentic real-world massive data benchmark runner and Colab notebook: benchmarks/README.md

Add authentic real-world massive data benchmark runner and Colab notebook: benchmarks/ISOM_R1_Real_World_Massive_Benchmark.ipynb

Add authentic real-world massive data benchmark runner and Colab notebook: benchmarks/benchmark_real_world_massive.py

Fix F1-F4: INT8 scale overhead docs, 15.8x peak label, dynamic _first_seen_layer, CyclicManifoldVerifier heuristic disclaimer

Fix F1-F4: INT8 scale overhead docs, 15.8x peak label, dynamic _first_seen_layer, CyclicManifoldVerifier heuristic disclaimer

Fix F1-F4: INT8 scale overhead docs, 15.8x peak label, dynamic _first_seen_layer, CyclicManifoldVerifier heuristic disclaimer

Audited architectural harmonization: benchmarks/README.md

Audited architectural harmonization: benchmarks/audited_systems_benchmark_t4.json

Audited architectural harmonization: README.md

Audited architectural harmonization: NOTICE

fix(core): multi-head saliency, step-level holographic caching, and scientifically grounded benchmark documentation

fix(core): multi-head saliency, step-level holographic caching, and scientifically grounded benchmark documentation

fix(core): multi-head saliency, step-level holographic caching, and scientifically grounded benchmark documentation

fix(core): multi-head saliency, step-level holographic caching, and scientifically grounded benchmark documentation

perf: replace O(L^2) Gram matrix with O(L*M) landmark entropy saliency (125x memory reduction)

docs: update repository links to Beta

fix: bounded cache invariant, forward cache injection, typing annotations

fix(core): expose configurable slack_tokens and track autoregressive tokens in holographic table

fix(core): expose configurable slack_tokens and track autoregressive tokens in holographic table

refactor(core): streamline modeling_isom.py, reduce cyclomatic complexity by 57% and eliminate dead code

Move Notice of Non-Endorsement & Lineage to footer for academic readership

Add formal LICENSE: CC BY-NC-ND 4.0 and Enterprise Commercial Restriction

Enforce CC BY-NC-ND 4.0 license restriction in model card YAML

Update README with Kaggle Tesla T4 verified empirical benchmarks & non-endorsement notices

docs: update to ISOM-R1 sovereign signature with non-endorsement lineage notes

docs: update to ISOM-R1 sovereign signature with non-endorsement lineage notes

docs: update to ISOM-R1 sovereign signature with non-endorsement lineage notes

docs: update to ISOM-R1 sovereign signature with non-endorsement lineage notes

docs: update to ISOM-R1 sovereign signature with non-endorsement lineage notes

docs: update to ISOM-R1 sovereign signature with non-endorsement lineage notes

docs: add unified hybrid signature lineup and base model lineage notes

privacy: replace email with LinkedIn in benchmark JSON

privacy: replace email with LinkedIn contact

docs: fix all math rendering in table cells, add Qwen2.5-Coder to lineup

docs: update benchmarks README with LinkedIn research integration link

docs: add mandatory LinkedIn research integration protocol and badge

docs: add benchmarks documentation with empirical audit tables and reproducibility specs

data: upload audited 4-pillar physical benchmark results on Tesla T4 GPU

docs: revamp README with audited 4-pillar benchmarks, Zenodo DOI, and clear developer marketing

Update model card with audited 4-pillar benchmark results: 100% NIAH (5/5), O(1) 32K scaling

Fix SDPA bias dtype to match query.dtype (model.dtype)

Update default isom_budget to 8192

Add exact 4D causal attention mask to chunked prefill and set isom_budget to 8192

Fix Pillar 3 batch concurrency: Native chunked prefill + decode loop, eliminate super().generate cache slicing bug

feat: Add prefill_chunk_size=2048 to config.json

feat: Add native chunked prefill loop to eliminate O(S^2) activation spike