Skip to content

Changelog

Record of structural and conceptual changes to the collection. Trivial typographical fixes may not appear here.

2026-10-16 — v1.1.0: Bilingual technical corrections

Corrected and clarified

  • Corrected the Top-K support direction, the CTC alignment example, selection indices, probability components, and temporal/patch indexing.
  • Separated integration steps from model evaluations, codec code counts from sequential decoding steps, and prediction horizons from execution horizons.
  • Qualified complexity, uncertainty, voting, Best-of-N, GRPO, graph operators, and risk statements with their required assumptions; distinguished scientific claims from conceptual diagrams and empirical tendencies.
  • Distinguished off-policy from offline training and frozen parameters from mathematical derivatives; clarified optimizer, adaptation, and chat-serialization semantics.
  • Preserved the contrastive formula as an explicitly identified decoupled contrastive loss (DCL), with its domain and primary reference.
  • Corrected graph-generation and link-prediction task definitions, schema-prefix constraints, retrieval conventions, and the strength of a cited time-series claim.
  • Aligned terminology, conditioning tags, portable-link conventions, and relevant English/Portuguese wording while preserving each edition's editorial voice.

Release v1.1.0 consolidates these corrections and the preceding mathematical-rendering repairs without adding chapters. The archived v1.0.0 release and its DOI remain unchanged. The review and corrections were AI-assisted and checked against primary sources and reproducible examples; they do not constitute a guarantee that no further errors remain.

2026-10-01 — Lifecycle & Governance Layer

Added

16_TRAINING_ALIGNMENT_ADAPTATION_DATASHEET.md

New lifecycle layer covering:

  • data mixture/curation;
  • objectives;
  • optimizer mechanics;
  • full fine-tuning;
  • SFT;
  • reward/preference optimization;
  • PPO/DPO/KTO/ORPO families;
  • GRPO/DAPO/GSPO;
  • RLVR;
  • LoRA/QLoRA/DoRA;
  • distillation;
  • model merging;
  • pruning;
  • continual learning;
  • training memory/compute.

17_EVALUATION_BENCHMARK_EXPERIMENTATION_DATASHEET.md

New evidence-engineering layer:

  • System Under Test definition;
  • metric/harness separation;
  • uncertainty;
  • paired experiments;
  • bootstrap;
  • benchmark contamination/saturation;
  • judge bias;
  • runtime metrics;
  • local-model Pareto fronts;
  • regression suites.

18_SECURITY_ROBUSTNESS_FAILURE_MODES_DATASHEET.md

New cross-cutting risk/control layer:

  • prompt injection;
  • RAG/memory poisoning;
  • supply chain;
  • agent security;
  • authorization;
  • ACS runtime control;
  • adversarial robustness;
  • local AI attack surface;
  • observability/incident response.

19_PARAMETER_GLOSSARY_AND_REGISTRY.md

Canonical data dictionary:

  • provenance tags;
  • symbol conventions;
  • parameter collisions;
  • canonical names;
  • backend aliases;
  • domain registries;
  • units;
  • deprecation/semantic-drift rules.

CONVENTIONS.md

Added editorial and maintenance rules to prevent arbitrary taxonomic growth.

CHANGELOG.md

Created this history.

Changed

00_README_INDEX.md

  • expanded tree to 00–19;
  • added META / MODEL / SYSTEM / LIFECYCLE / RUNTIME layers;
  • updated reading guide;
  • expanded provenance tags.

01_GUT_GENERATIVE_AI_UNIFIED_ENGINEERING.md

  • added the distinction between the execution model tuple and the lifecycle envelope;
  • training/evaluation/security now wrap the model ontology instead of being treated as “model types.”

2026-10-01 — Foundation Model Expansion

Added

  • 10_MULTIMODAL_VLM_OMNI_DATASHEET.md
  • 11_EMBEDDING_RETRIEVAL_RERANKER_DATASHEET.md
  • 12_REWARD_VERIFIER_JUDGE_DATASHEET.md
  • 13_WORLD_MODEL_VLA_EMBODIED_AI_DATASHEET.md
  • 14_TIME_SERIES_TABULAR_FOUNDATION_MODELS.md
  • 15_GRAPH_FOUNDATION_MODELS_DATASHEET.md

Changed

  • 01_GUT_GENERATIVE_AI_UNIFIED_ENGINEERING.md: expanded representation/objective/backbone axes.
  • 02_LLM_TRANSFORMER_INFERENCE_DATASHEET.md: incorporated masked/diffusion/parallel language generation.
  • 06_AUDIO_MUSIC_TTS_GENERATION_DATASHEET.md: incorporated ASR/speech understanding.
  • 09_LOCAL_AI_INFERENCE_RUNTIME_DATASHEET.md: incorporated multimodal, retrieval, and parallel-language cost models.
  • 00_README_INDEX.md: expanded the collection and consolidated cross-links.

2026-10-01 — SOTA++ Rebuild

The collection was rebuilt from the legacy 2025 datasheets.

Central change:

before:
  Transformer vs Diffusion

now:
  Representation
  × Objective
  × Backbone
  × Conditioning
  × Inference
  × Persistent State
  × Decoder
  × Modality Topology

Major corrections:

  • seed is no longer treated as a guarantee of determinism;
  • CFG is no longer treated as mathematically equivalent to temperature;
  • sampler and scheduler were separated;
  • diffusion and flow matching were separated;
  • Transformer is no longer treated as synonymous with autoregression;
  • KV/cache/runtime received explicit physical modeling;
  • 3D now starts from representation;
  • agents are treated as control systems with policy/tool/environment;
  • model snapshots were separated from the mathematical core.

Legacy source set — 2025-12-04 / 2025-12-05

The original materials that motivated the rebuild included datasheets for:

  • LLMs;
  • code generation;
  • image diffusion;
  • video generation;
  • audio generation;
  • 3D generation;
  • action models;
  • a generative-AI GUT.

They remain historically useful as a snapshot of the previous mental model, but have been superseded by the current SOTA++ collection.