[
  {
    "source": "A-007",
    "target": "P-001",
    "type": "generalizes",
    "explanation": "Self-attention generalizes content-based soft alignment",
    "evidenceTier": "direct",
    "citation": "P-001 cites Bahdanau 2014",
    "verificationDate": null
  },
  {
    "source": "A-006",
    "target": "P-001",
    "type": "extends",
    "explanation": "Encoder-decoder transduction framing carried over from seq2seq",
    "evidenceTier": "direct",
    "citation": "P-001 background",
    "verificationDate": null
  },
  {
    "source": "A-009",
    "target": "P-001",
    "type": "depends_on",
    "explanation": "Residual sublayers (x + Sublayer(x)) in every block",
    "evidenceTier": "direct",
    "citation": "P-001 Sec 3.1",
    "verificationDate": null
  },
  {
    "source": "A-010",
    "target": "P-001",
    "type": "depends_on",
    "explanation": "LayerNorm applied around each sublayer",
    "evidenceTier": "direct",
    "citation": "P-001 Sec 3.1",
    "verificationDate": null
  },
  {
    "source": "A-011",
    "target": "P-001",
    "type": "depends_on",
    "explanation": "Adam optimizer with warmup schedule",
    "evidenceTier": "direct",
    "citation": "P-001 Sec 5.3",
    "verificationDate": null
  },
  {
    "source": "A-013",
    "target": "P-001",
    "type": "depends_on",
    "explanation": "Subword (BPE) vocabulary for open-vocab modeling",
    "evidenceTier": "direct",
    "citation": "P-001 Sec 5.1",
    "verificationDate": null
  },
  {
    "source": "A-004",
    "target": "P-001",
    "type": "depends_on",
    "explanation": "Learned token embeddings feed the input",
    "evidenceTier": "direct",
    "citation": "standard practice",
    "verificationDate": null
  },
  {
    "source": "P-001",
    "target": "P-010",
    "type": "applies_to",
    "explanation": "Decoder-only autoregressive LM built on Transformer decoder",
    "evidenceTier": "direct",
    "citation": "P-010 architecture",
    "verificationDate": null
  },
  {
    "source": "P-001",
    "target": "P-013",
    "type": "applies_to",
    "explanation": "Encoder-only masked LM built on Transformer encoder",
    "evidenceTier": "direct",
    "citation": "P-013 architecture",
    "verificationDate": null
  },
  {
    "source": "P-001",
    "target": "P-020",
    "type": "applies_to",
    "explanation": "Encoder-decoder text-to-text model",
    "evidenceTier": "direct",
    "citation": "P-020 architecture",
    "verificationDate": null
  },
  {
    "source": "P-001",
    "target": "P-021",
    "type": "applies_to",
    "explanation": "Encoder-decoder denoising autoencoder",
    "evidenceTier": "direct",
    "citation": "P-021 architecture",
    "verificationDate": null
  },
  {
    "source": "P-010",
    "target": "P-011",
    "type": "scales",
    "explanation": "Same generative-pretraining recipe scaled to 1.5B",
    "evidenceTier": "direct",
    "citation": "P-011 report",
    "verificationDate": null
  },
  {
    "source": "P-011",
    "target": "P-012",
    "type": "scales",
    "explanation": "Scaled to 175B; few-shot in-context learning emerges",
    "evidenceTier": "direct",
    "citation": "P-012 paper",
    "verificationDate": null
  },
  {
    "source": "P-013",
    "target": "P-014",
    "type": "improves",
    "explanation": "Corrective ablation: drop NSP train longer on more data",
    "evidenceTier": "direct",
    "citation": "P-014 abstract",
    "verificationDate": null
  },
  {
    "source": "P-013",
    "target": "P-015",
    "type": "makes_efficient",
    "explanation": "Cross-layer parameter sharing and factorized embeddings",
    "evidenceTier": "direct",
    "citation": "P-015 abstract",
    "verificationDate": null
  },
  {
    "source": "P-013",
    "target": "P-016",
    "type": "improves",
    "explanation": "Replaced-token-detection objective improves sample efficiency",
    "evidenceTier": "direct",
    "citation": "P-016 paper",
    "verificationDate": null
  },
  {
    "source": "P-013",
    "target": "P-024",
    "type": "improves",
    "explanation": "Disentangled content/position attention",
    "evidenceTier": "direct",
    "citation": "P-024 paper",
    "verificationDate": null
  },
  {
    "source": "P-006",
    "target": "P-024",
    "type": "depends_on",
    "explanation": "Relative position representations underpin disentangled attention",
    "evidenceTier": "strongly_supported",
    "citation": "P-024 Sec 2",
    "verificationDate": null
  },
  {
    "source": "P-006",
    "target": "P-002",
    "type": "generalizes",
    "explanation": "Rotary embedding continues the relative-position lineage",
    "evidenceTier": "strongly_supported",
    "citation": "P-002 related work",
    "verificationDate": null
  },
  {
    "source": "P-020",
    "target": "P-022",
    "type": "generalizes",
    "explanation": "Mixture-of-denoisers generalizes span-corruption",
    "evidenceTier": "direct",
    "citation": "P-022 paper",
    "verificationDate": null
  },
  {
    "source": "P-001",
    "target": "P-003",
    "type": "makes_efficient",
    "explanation": "RMSNorm simplifies LayerNorm used in Transformer",
    "evidenceTier": "direct",
    "citation": "P-003 paper",
    "verificationDate": null
  },
  {
    "source": "P-001",
    "target": "P-004",
    "type": "improves",
    "explanation": "SwiGLU replaces ReLU FFN in Transformer",
    "evidenceTier": "direct",
    "citation": "P-004 paper",
    "verificationDate": null
  },
  {
    "source": "P-001",
    "target": "P-005",
    "type": "makes_efficient",
    "explanation": "ALiBi replaces positional encodings for length extrapolation",
    "evidenceTier": "direct",
    "citation": "P-005 paper",
    "verificationDate": null
  },
  {
    "source": "P-001",
    "target": "P-007",
    "type": "makes_efficient",
    "explanation": "Pre-norm placement stabilizes deep Transformer training",
    "evidenceTier": "direct",
    "citation": "P-007 paper",
    "verificationDate": null
  },
  {
    "source": "P-001",
    "target": "P-008",
    "type": "makes_efficient",
    "explanation": "MQA shares K/V heads to shrink KV cache",
    "evidenceTier": "direct",
    "citation": "P-008 paper",
    "verificationDate": null
  },
  {
    "source": "P-008",
    "target": "P-009",
    "type": "generalizes",
    "explanation": "GQA interpolates between MHA and MQA via grouped K/V",
    "evidenceTier": "direct",
    "citation": "P-009 paper",
    "verificationDate": null
  },
  {
    "source": "P-012",
    "target": "P-100",
    "type": "provides_evidence_for",
    "explanation": "GPT-3-era models motivate the scaling-law study",
    "evidenceTier": "direct",
    "citation": "P-100",
    "verificationDate": null
  },
  {
    "source": "P-100",
    "target": "P-101",
    "type": "challenges",
    "explanation": "Chinchilla corrects Kaplan optimal N/D allocation",
    "evidenceTier": "direct",
    "citation": "P-101 abstract",
    "verificationDate": null
  },
  {
    "source": "P-101",
    "target": "P-104",
    "type": "enables",
    "explanation": "Chinchilla token hunger raises the data-scarcity question",
    "evidenceTier": "strongly_supported",
    "citation": "P-104 intro",
    "verificationDate": null
  },
  {
    "source": "P-102",
    "target": "P-103",
    "type": "challenges",
    "explanation": "Mirage shows emergence is partly a metric artifact",
    "evidenceTier": "direct",
    "citation": "P-103 abstract",
    "verificationDate": null
  },
  {
    "source": "P-110",
    "target": "P-111",
    "type": "applies_to",
    "explanation": "MoE moved into the Transformer at scale",
    "evidenceTier": "direct",
    "citation": "P-111",
    "verificationDate": null
  },
  {
    "source": "P-111",
    "target": "P-112",
    "type": "improves",
    "explanation": "Switch simplifies routing to top-1",
    "evidenceTier": "direct",
    "citation": "P-112",
    "verificationDate": null
  },
  {
    "source": "P-112",
    "target": "P-113",
    "type": "extends",
    "explanation": "GLaM applies MoE at frontier quality",
    "evidenceTier": "direct",
    "citation": "P-113",
    "verificationDate": null
  },
  {
    "source": "P-112",
    "target": "P-116",
    "type": "extends",
    "explanation": "Mixtral brings MoE to open-weight ecosystem",
    "evidenceTier": "strongly_supported",
    "citation": "P-116",
    "verificationDate": null
  },
  {
    "source": "P-114",
    "target": "P-117",
    "type": "improves",
    "explanation": "Routing ideas feed DeepSeekMoE",
    "evidenceTier": "strongly_supported",
    "citation": "P-117 related work",
    "verificationDate": null
  },
  {
    "source": "P-115",
    "target": "P-117",
    "type": "improves",
    "explanation": "Stability recipes feed DeepSeekMoE",
    "evidenceTier": "strongly_supported",
    "citation": "P-117",
    "verificationDate": null
  },
  {
    "source": "P-111",
    "target": "P-135",
    "type": "converts_into_infrastructure",
    "explanation": "GShard sharding becomes GSPMD",
    "evidenceTier": "direct",
    "citation": "P-135",
    "verificationDate": null
  },
  {
    "source": "A-013",
    "target": "P-120",
    "type": "generalizes",
    "explanation": "SentencePiece generalizes BPE to raw text",
    "evidenceTier": "direct",
    "citation": "P-120",
    "verificationDate": null
  },
  {
    "source": "P-120",
    "target": "P-121",
    "type": "combines",
    "explanation": "Unigram LM tokenization implemented alongside BPE",
    "evidenceTier": "direct",
    "citation": "P-120",
    "verificationDate": null
  },
  {
    "source": "P-123",
    "target": "P-124",
    "type": "extends",
    "explanation": "RefinedWeb extends open-corpus practice with hard filtering",
    "evidenceTier": "strongly_supported",
    "citation": "P-124",
    "verificationDate": null
  },
  {
    "source": "P-125",
    "target": "P-126",
    "type": "challenges",
    "explanation": "Model collapse bounds synthetic-data optimism",
    "evidenceTier": "strongly_supported",
    "citation": "P-126",
    "verificationDate": null
  },
  {
    "source": "P-130",
    "target": "P-138",
    "type": "extends",
    "explanation": "FP8 extends mixed precision to 8-bit",
    "evidenceTier": "direct",
    "citation": "P-138",
    "verificationDate": null
  },
  {
    "source": "P-132",
    "target": "P-136",
    "type": "extends",
    "explanation": "FSDP is framework-native ZeRO-3",
    "evidenceTier": "direct",
    "citation": "P-136",
    "verificationDate": null
  },
  {
    "source": "P-133",
    "target": "P-134",
    "type": "improves",
    "explanation": "1F1B scheduling reduces pipeline bubble",
    "evidenceTier": "direct",
    "citation": "P-134",
    "verificationDate": null
  },
  {
    "source": "P-200",
    "target": "P-201",
    "type": "applies_to",
    "explanation": "Preference-RL applied to language summarization",
    "evidenceTier": "direct",
    "citation": "P-201",
    "verificationDate": null
  },
  {
    "source": "P-201",
    "target": "P-202",
    "type": "extends",
    "explanation": "InstructGPT extends RLHF-for-text to instruction following",
    "evidenceTier": "direct",
    "citation": "P-202",
    "verificationDate": null
  },
  {
    "source": "P-203",
    "target": "P-202",
    "type": "depends_on",
    "explanation": "SFT/instruction data is stage 1 of RLHF",
    "evidenceTier": "strongly_supported",
    "citation": "P-202",
    "verificationDate": null
  },
  {
    "source": "P-202",
    "target": "P-204",
    "type": "extends",
    "explanation": "Constitutional AI replaces human feedback with AI feedback",
    "evidenceTier": "direct",
    "citation": "P-204",
    "verificationDate": null
  },
  {
    "source": "P-202",
    "target": "P-205",
    "type": "replaces",
    "explanation": "DPO removes reward model and PPO from RLHF objective",
    "evidenceTier": "direct",
    "citation": "P-205",
    "verificationDate": null
  },
  {
    "source": "P-205",
    "target": "P-206",
    "type": "generalizes",
    "explanation": "IPO/KTO/ORPO/SimPO vary the DPO design space",
    "evidenceTier": "direct",
    "citation": "P-206",
    "verificationDate": null
  },
  {
    "source": "P-202",
    "target": "P-207",
    "type": "challenges",
    "explanation": "Overoptimization and sycophancy are RLHF failure modes",
    "evidenceTier": "direct",
    "citation": "P-207",
    "verificationDate": null
  },
  {
    "source": "P-207",
    "target": "P-208",
    "type": "challenges",
    "explanation": "Failure modes motivate scalable oversight",
    "evidenceTier": "strongly_supported",
    "citation": "P-208",
    "verificationDate": null
  },
  {
    "source": "P-012",
    "target": "P-220",
    "type": "enables",
    "explanation": "Chain-of-thought emerges only at scale",
    "evidenceTier": "direct",
    "citation": "P-220",
    "verificationDate": null
  },
  {
    "source": "P-220",
    "target": "P-221",
    "type": "extends",
    "explanation": "Self-consistency votes over many CoT samples",
    "evidenceTier": "direct",
    "citation": "P-221",
    "verificationDate": null
  },
  {
    "source": "P-220",
    "target": "P-222",
    "type": "extends",
    "explanation": "ToT/L2M/PoT structure reasoning beyond a linear chain",
    "evidenceTier": "direct",
    "citation": "P-222",
    "verificationDate": null
  },
  {
    "source": "P-223",
    "target": "P-224",
    "type": "depends_on",
    "explanation": "RLVR uses verifiers/process rewards",
    "evidenceTier": "direct",
    "citation": "P-224",
    "verificationDate": null
  },
  {
    "source": "P-220",
    "target": "P-224",
    "type": "depends_on",
    "explanation": "RLVR learns long chains of thought",
    "evidenceTier": "strongly_supported",
    "citation": "P-224",
    "verificationDate": null
  },
  {
    "source": "P-224",
    "target": "P-225",
    "type": "parallel",
    "explanation": "R1 openly parallels closed o-series behavior",
    "evidenceTier": "probable",
    "citation": "P-225 system card",
    "verificationDate": null
  },
  {
    "source": "P-220",
    "target": "P-226",
    "type": "challenges",
    "explanation": "CoT can be unfaithful; reasoning correctives",
    "evidenceTier": "direct",
    "citation": "P-226",
    "verificationDate": null
  },
  {
    "source": "P-220",
    "target": "P-230",
    "type": "combines",
    "explanation": "ReAct combines reasoning with acting",
    "evidenceTier": "direct",
    "citation": "P-230",
    "verificationDate": null
  },
  {
    "source": "P-232",
    "target": "P-230",
    "type": "combines",
    "explanation": "Tool/function-calling primitives feed the agent loop",
    "evidenceTier": "direct",
    "citation": "P-230",
    "verificationDate": null
  },
  {
    "source": "P-230",
    "target": "P-231",
    "type": "extends",
    "explanation": "Toolformer trains tool use rather than prompting it",
    "evidenceTier": "strongly_supported",
    "citation": "P-231",
    "verificationDate": null
  },
  {
    "source": "P-231",
    "target": "P-232",
    "type": "converts_into_infrastructure",
    "explanation": "Tool use becomes function-calling/MCP infrastructure",
    "evidenceTier": "strongly_supported",
    "citation": "P-232",
    "verificationDate": null
  },
  {
    "source": "P-230",
    "target": "P-233",
    "type": "extends",
    "explanation": "Reflexion adds verbal self-critique to the loop",
    "evidenceTier": "direct",
    "citation": "P-233",
    "verificationDate": null
  },
  {
    "source": "P-234",
    "target": "P-235",
    "type": "combines",
    "explanation": "Memory/skill-library ideas feed coding agents",
    "evidenceTier": "strongly_supported",
    "citation": "P-235",
    "verificationDate": null
  },
  {
    "source": "P-233",
    "target": "P-235",
    "type": "combines",
    "explanation": "Self-reflection feeds SWE agents",
    "evidenceTier": "strongly_supported",
    "citation": "P-235",
    "verificationDate": null
  },
  {
    "source": "P-224",
    "target": "P-235",
    "type": "applies_to",
    "explanation": "Verifiable-reward logic applies to SWE-bench-graded agents",
    "evidenceTier": "strongly_supported",
    "citation": "P-235",
    "verificationDate": null
  },
  {
    "source": "P-001",
    "target": "P-300",
    "type": "applies_to",
    "explanation": "Transformer applied to image patches",
    "evidenceTier": "direct",
    "citation": "P-300",
    "verificationDate": null
  },
  {
    "source": "P-300",
    "target": "P-301",
    "type": "depends_on",
    "explanation": "CLIP uses a ViT image encoder",
    "evidenceTier": "direct",
    "citation": "P-301",
    "verificationDate": null
  },
  {
    "source": "P-302",
    "target": "P-301",
    "type": "parallel",
    "explanation": "ALIGN is concurrent contrastive VL pretraining",
    "evidenceTier": "strongly_supported",
    "citation": "P-302",
    "verificationDate": null
  },
  {
    "source": "P-301",
    "target": "P-303",
    "type": "enables",
    "explanation": "CLIP encoder feeds Flamingo",
    "evidenceTier": "direct",
    "citation": "P-303",
    "verificationDate": null
  },
  {
    "source": "P-301",
    "target": "P-304",
    "type": "enables",
    "explanation": "CLIP encoder feeds BLIP-2",
    "evidenceTier": "direct",
    "citation": "P-304",
    "verificationDate": null
  },
  {
    "source": "P-301",
    "target": "P-305",
    "type": "enables",
    "explanation": "CLIP features projected into LLaVA",
    "evidenceTier": "direct",
    "citation": "P-305",
    "verificationDate": null
  },
  {
    "source": "P-304",
    "target": "P-306",
    "type": "extends",
    "explanation": "Q-Former pattern feeds production VLMs",
    "evidenceTier": "strongly_supported",
    "citation": "P-306",
    "verificationDate": null
  },
  {
    "source": "P-305",
    "target": "P-306",
    "type": "extends",
    "explanation": "LLaVA recipe scaled to production VLMs",
    "evidenceTier": "strongly_supported",
    "citation": "P-306",
    "verificationDate": null
  },
  {
    "source": "P-300",
    "target": "P-308",
    "type": "applies_to",
    "explanation": "DiT uses a Transformer denoiser",
    "evidenceTier": "direct",
    "citation": "P-308",
    "verificationDate": null
  },
  {
    "source": "P-306",
    "target": "P-309",
    "type": "parallel",
    "explanation": "Open VLMs parallel closed native multimodal",
    "evidenceTier": "probable",
    "citation": "P-309",
    "verificationDate": null
  },
  {
    "source": "P-013",
    "target": "P-320",
    "type": "applies_to",
    "explanation": "DPR uses BERT dual encoders",
    "evidenceTier": "direct",
    "citation": "P-320",
    "verificationDate": null
  },
  {
    "source": "P-321",
    "target": "P-322",
    "type": "depends_on",
    "explanation": "RAG builds on retrieval-augmented LM idea",
    "evidenceTier": "direct",
    "citation": "P-322",
    "verificationDate": null
  },
  {
    "source": "P-320",
    "target": "P-322",
    "type": "depends_on",
    "explanation": "RAG uses DPR retriever",
    "evidenceTier": "direct",
    "citation": "P-322",
    "verificationDate": null
  },
  {
    "source": "P-322",
    "target": "P-323",
    "type": "extends",
    "explanation": "FiD/RETRO/Atlas scale retrieval",
    "evidenceTier": "direct",
    "citation": "P-323",
    "verificationDate": null
  },
  {
    "source": "P-320",
    "target": "P-324",
    "type": "improves",
    "explanation": "ColBERT/rerankers improve retrieval quality",
    "evidenceTier": "direct",
    "citation": "P-324",
    "verificationDate": null
  },
  {
    "source": "P-013",
    "target": "P-325",
    "type": "applies_to",
    "explanation": "Sentence-BERT adapts BERT for embeddings",
    "evidenceTier": "direct",
    "citation": "P-325",
    "verificationDate": null
  },
  {
    "source": "P-325",
    "target": "P-322",
    "type": "converts_into_infrastructure",
    "explanation": "Embeddings power production RAG",
    "evidenceTier": "strongly_supported",
    "citation": "P-325",
    "verificationDate": null
  },
  {
    "source": "P-010",
    "target": "P-340",
    "type": "applies_to",
    "explanation": "Codex fine-tunes GPT on code",
    "evidenceTier": "direct",
    "citation": "P-340",
    "verificationDate": null
  },
  {
    "source": "P-340",
    "target": "P-341",
    "type": "extends",
    "explanation": "AlphaCode extends code gen with search",
    "evidenceTier": "strongly_supported",
    "citation": "P-341",
    "verificationDate": null
  },
  {
    "source": "P-021",
    "target": "P-342",
    "type": "generalizes",
    "explanation": "FIM generalizes BART infilling to code",
    "evidenceTier": "strongly_supported",
    "citation": "P-342",
    "verificationDate": null
  },
  {
    "source": "P-340",
    "target": "P-343",
    "type": "extends",
    "explanation": "Open code models follow Codex",
    "evidenceTier": "direct",
    "citation": "P-343",
    "verificationDate": null
  },
  {
    "source": "P-342",
    "target": "P-343",
    "type": "combines",
    "explanation": "FIM used in open code models",
    "evidenceTier": "direct",
    "citation": "P-343",
    "verificationDate": null
  },
  {
    "source": "P-343",
    "target": "P-344",
    "type": "enables",
    "explanation": "Open code models enable execution-feedback loops",
    "evidenceTier": "direct",
    "citation": "P-344",
    "verificationDate": null
  },
  {
    "source": "P-101",
    "target": "P-360",
    "type": "enables",
    "explanation": "Chinchilla-informed over-training shapes LLaMA",
    "evidenceTier": "direct",
    "citation": "P-360",
    "verificationDate": null
  },
  {
    "source": "P-360",
    "target": "P-361",
    "type": "extends",
    "explanation": "Llama 2/3 extend LLaMA",
    "evidenceTier": "direct",
    "citation": "P-361",
    "verificationDate": null
  },
  {
    "source": "P-360",
    "target": "P-362",
    "type": "extends",
    "explanation": "Mistral builds on the open block",
    "evidenceTier": "strongly_supported",
    "citation": "P-362",
    "verificationDate": null
  },
  {
    "source": "P-360",
    "target": "P-363",
    "type": "extends",
    "explanation": "Qwen builds on the open paradigm",
    "evidenceTier": "strongly_supported",
    "citation": "P-363",
    "verificationDate": null
  },
  {
    "source": "P-360",
    "target": "P-364",
    "type": "extends",
    "explanation": "DeepSeek builds on the open paradigm",
    "evidenceTier": "strongly_supported",
    "citation": "P-364",
    "verificationDate": null
  },
  {
    "source": "P-117",
    "target": "P-364",
    "type": "depends_on",
    "explanation": "DeepSeek uses fine-grained+shared-expert MoE",
    "evidenceTier": "direct",
    "citation": "P-364",
    "verificationDate": null
  },
  {
    "source": "P-224",
    "target": "P-364",
    "type": "extends",
    "explanation": "DeepSeek-R1 reasoning extends the family",
    "evidenceTier": "direct",
    "citation": "P-364",
    "verificationDate": null
  },
  {
    "source": "P-360",
    "target": "P-365",
    "type": "extends",
    "explanation": "Broad open ecosystem follows LLaMA",
    "evidenceTier": "strongly_supported",
    "citation": "P-365",
    "verificationDate": null
  },
  {
    "source": "P-001",
    "target": "P-400",
    "type": "makes_efficient",
    "explanation": "FlashAttention makes exact attention IO-efficient",
    "evidenceTier": "direct",
    "citation": "P-400",
    "verificationDate": null
  },
  {
    "source": "P-400",
    "target": "P-401",
    "type": "combines",
    "explanation": "vLLM pairs paged KV with flash attention",
    "evidenceTier": "strongly_supported",
    "citation": "P-401",
    "verificationDate": null
  },
  {
    "source": "P-402",
    "target": "P-401",
    "type": "combines",
    "explanation": "Continuous batching combined in vLLM",
    "evidenceTier": "direct",
    "citation": "P-401",
    "verificationDate": null
  },
  {
    "source": "P-404",
    "target": "P-405",
    "type": "enables",
    "explanation": "4-bit quantization enables QLoRA",
    "evidenceTier": "direct",
    "citation": "P-405",
    "verificationDate": null
  },
  {
    "source": "P-404",
    "target": "P-406",
    "type": "enables",
    "explanation": "Quantization enables edge/GGUF inference",
    "evidenceTier": "direct",
    "citation": "P-406",
    "verificationDate": null
  },
  {
    "source": "P-001",
    "target": "P-420",
    "type": "makes_efficient",
    "explanation": "Sparse attention reduces attention cost",
    "evidenceTier": "direct",
    "citation": "P-420",
    "verificationDate": null
  },
  {
    "source": "P-001",
    "target": "P-421",
    "type": "makes_efficient",
    "explanation": "Linear attention approximates softmax attention",
    "evidenceTier": "direct",
    "citation": "P-421",
    "verificationDate": null
  },
  {
    "source": "P-400",
    "target": "P-422",
    "type": "makes_efficient",
    "explanation": "Ring attention extends flash attention across devices",
    "evidenceTier": "strongly_supported",
    "citation": "P-422",
    "verificationDate": null
  },
  {
    "source": "P-002",
    "target": "P-423",
    "type": "extends",
    "explanation": "YaRN rescales RoPE for context extension",
    "evidenceTier": "direct",
    "citation": "P-423",
    "verificationDate": null
  },
  {
    "source": "P-424",
    "target": "P-425",
    "type": "extends",
    "explanation": "Mamba adds selectivity to S4",
    "evidenceTier": "direct",
    "citation": "P-425",
    "verificationDate": null
  },
  {
    "source": "P-425",
    "target": "P-426",
    "type": "extends",
    "explanation": "Hybrids interleave Mamba with attention",
    "evidenceTier": "strongly_supported",
    "citation": "P-426",
    "verificationDate": null
  },
  {
    "source": "P-421",
    "target": "P-425",
    "type": "parallel",
    "explanation": "Linear attention and SSMs are related",
    "evidenceTier": "probable",
    "citation": "P-425",
    "verificationDate": null
  },
  {
    "source": "P-440",
    "target": "P-449",
    "type": "challenges",
    "explanation": "MMLU saturation/contamination feeds eval crisis",
    "evidenceTier": "direct",
    "citation": "P-449",
    "verificationDate": null
  },
  {
    "source": "P-444",
    "target": "P-449",
    "type": "challenges",
    "explanation": "HumanEval saturation feeds eval crisis",
    "evidenceTier": "direct",
    "citation": "P-449",
    "verificationDate": null
  },
  {
    "source": "P-447",
    "target": "P-449",
    "type": "challenges",
    "explanation": "Judge bias feeds eval crisis",
    "evidenceTier": "direct",
    "citation": "P-449",
    "verificationDate": null
  },
  {
    "source": "A-018",
    "target": "A-019",
    "type": "extends",
    "explanation": "LeNet CNN scaled up by AlexNet",
    "evidenceTier": "strongly_supported",
    "citation": "A-019",
    "verificationDate": null
  },
  {
    "source": "A-016",
    "target": "A-019",
    "type": "depends_on",
    "explanation": "AlexNet used ReLU",
    "evidenceTier": "strongly_supported",
    "citation": "A-019",
    "verificationDate": null
  },
  {
    "source": "A-020",
    "target": "A-021",
    "type": "challenges",
    "explanation": "Elman RNNs exhibit vanishing gradients",
    "evidenceTier": "direct",
    "citation": "A-021",
    "verificationDate": null
  },
  {
    "source": "A-021",
    "target": "A-022",
    "type": "challenges",
    "explanation": "LSTM solves the vanishing-gradient problem",
    "evidenceTier": "direct",
    "citation": "A-022",
    "verificationDate": null
  },
  {
    "source": "A-022",
    "target": "A-006",
    "type": "applies_to",
    "explanation": "Seq2seq built from LSTMs",
    "evidenceTier": "strongly_supported",
    "citation": "A-006",
    "verificationDate": null
  },
  {
    "source": "A-022",
    "target": "A-023",
    "type": "extends",
    "explanation": "BiRNN runs LSTMs both directions",
    "evidenceTier": "direct",
    "citation": "A-023",
    "verificationDate": null
  },
  {
    "source": "A-022",
    "target": "A-024",
    "type": "generalizes",
    "explanation": "GRU simplifies LSTM",
    "evidenceTier": "direct",
    "citation": "A-024",
    "verificationDate": null
  },
  {
    "source": "A-023",
    "target": "P-013",
    "type": "generalizes",
    "explanation": "Bidirectionality realized by BERT",
    "evidenceTier": "strongly_supported",
    "citation": "P-013",
    "verificationDate": null
  },
  {
    "source": "A-024",
    "target": "A-006",
    "type": "combines",
    "explanation": "GRU encoder-decoder sibling of seq2seq",
    "evidenceTier": "strongly_supported",
    "citation": "A-006",
    "verificationDate": null
  },
  {
    "source": "A-028",
    "target": "A-007",
    "type": "generalizes",
    "explanation": "Neural attention is soft learned alignment",
    "evidenceTier": "strongly_supported",
    "citation": "A-007",
    "verificationDate": null
  },
  {
    "source": "A-006",
    "target": "A-007",
    "type": "extends",
    "explanation": "Attention removes seq2seq bottleneck",
    "evidenceTier": "direct",
    "citation": "A-007",
    "verificationDate": null
  },
  {
    "source": "A-015",
    "target": "P-001",
    "type": "depends_on",
    "explanation": "Principled init enables deep Transformer training",
    "evidenceTier": "strongly_supported",
    "citation": "P-001",
    "verificationDate": null
  },
  {
    "source": "A-016",
    "target": "P-004",
    "type": "generalizes",
    "explanation": "ReLU is the parent of gated SwiGLU activations",
    "evidenceTier": "strongly_supported",
    "citation": "P-004",
    "verificationDate": null
  },
  {
    "source": "A-017",
    "target": "A-010",
    "type": "generalizes",
    "explanation": "LayerNorm adapts BatchNorm for sequences",
    "evidenceTier": "strongly_supported",
    "citation": "A-010",
    "verificationDate": null
  },
  {
    "source": "A-019",
    "target": "P-300",
    "type": "challenges",
    "explanation": "ViT re-solves vision without convolution",
    "evidenceTier": "strongly_supported",
    "citation": "P-300",
    "verificationDate": null
  },
  {
    "source": "A-027",
    "target": "A-004",
    "type": "enables",
    "explanation": "Collobert-Weston precedes word2vec",
    "evidenceTier": "strongly_supported",
    "citation": "A-004",
    "verificationDate": null
  },
  {
    "source": "A-029",
    "target": "A-030",
    "type": "extends",
    "explanation": "Q-learning builds on temporal-difference learning",
    "evidenceTier": "strongly_supported",
    "citation": "A-030",
    "verificationDate": null
  },
  {
    "source": "A-029",
    "target": "A-032",
    "type": "extends",
    "explanation": "TD-Gammon applies TD learning via self-play",
    "evidenceTier": "strongly_supported",
    "citation": "A-032",
    "verificationDate": null
  },
  {
    "source": "A-030",
    "target": "A-033",
    "type": "combines",
    "explanation": "DQN combines Q-learning with deep CNNs",
    "evidenceTier": "direct",
    "citation": "A-033",
    "verificationDate": null
  },
  {
    "source": "A-032",
    "target": "A-034",
    "type": "extends",
    "explanation": "Self-play lineage from TD-Gammon to AlphaGo",
    "evidenceTier": "strongly_supported",
    "citation": "A-034",
    "verificationDate": null
  },
  {
    "source": "O-001",
    "target": "O-004",
    "type": "depends_on",
    "explanation": "Boolean algebra realized by switching circuits",
    "evidenceTier": "strongly_supported",
    "citation": "O-004",
    "verificationDate": null
  },
  {
    "source": "O-002",
    "target": "O-003",
    "type": "depends_on",
    "explanation": "Programmable-machine concept precedes the Turing machine",
    "evidenceTier": "strongly_supported",
    "citation": "O-003",
    "verificationDate": null
  },
  {
    "source": "O-003",
    "target": "O-006",
    "type": "generalizes",
    "explanation": "Universal machine realized as stored-program computer",
    "evidenceTier": "strongly_supported",
    "citation": "O-006",
    "verificationDate": null
  },
  {
    "source": "O-004",
    "target": "O-006",
    "type": "depends_on",
    "explanation": "Digital logic underlies the computer",
    "evidenceTier": "strongly_supported",
    "citation": "O-006",
    "verificationDate": null
  },
  {
    "source": "O-014",
    "target": "O-006",
    "type": "enables",
    "explanation": "Wartime computing (Colossus) feeds stored-program machine",
    "evidenceTier": "strongly_supported",
    "citation": "O-006",
    "verificationDate": null
  },
  {
    "source": "O-012",
    "target": "O-006",
    "type": "enables",
    "explanation": "Manhattan Project computing feeds von Neumann's work",
    "evidenceTier": "strongly_supported",
    "citation": "O-006",
    "verificationDate": null
  },
  {
    "source": "O-012",
    "target": "O-013",
    "type": "enables",
    "explanation": "Monte Carlo born from neutron simulation",
    "evidenceTier": "direct",
    "citation": "O-013",
    "verificationDate": null
  },
  {
    "source": "O-003",
    "target": "O-011",
    "type": "extends",
    "explanation": "Turing extends computability to the question of thinking",
    "evidenceTier": "direct",
    "citation": "O-011",
    "verificationDate": null
  },
  {
    "source": "O-008",
    "target": "O-019",
    "type": "extends",
    "explanation": "Algorithmic information theory extends Shannon",
    "evidenceTier": "strongly_supported",
    "citation": "O-019",
    "verificationDate": null
  },
  {
    "source": "O-005",
    "target": "O-010",
    "type": "extends",
    "explanation": "Hebbian learning adds learning to the neuron model",
    "evidenceTier": "strongly_supported",
    "citation": "O-010",
    "verificationDate": null
  },
  {
    "source": "O-010",
    "target": "O-017",
    "type": "extends",
    "explanation": "Perceptron rule elaborates Hebbian learning",
    "evidenceTier": "strongly_supported",
    "citation": "O-017",
    "verificationDate": null
  },
  {
    "source": "O-005",
    "target": "O-017",
    "type": "depends_on",
    "explanation": "Perceptron builds on the artificial neuron",
    "evidenceTier": "strongly_supported",
    "citation": "O-017",
    "verificationDate": null
  },
  {
    "source": "O-017",
    "target": "O-018",
    "type": "challenges",
    "explanation": "Minsky-Papert prove perceptron limits",
    "evidenceTier": "direct",
    "citation": "O-018",
    "verificationDate": null
  },
  {
    "source": "O-005",
    "target": "A-001",
    "type": "enables",
    "explanation": "Artificial neuron ancestor of backprop nets",
    "evidenceTier": "strongly_supported",
    "citation": "A-001",
    "verificationDate": null
  },
  {
    "source": "O-017",
    "target": "A-001",
    "type": "enables",
    "explanation": "Perceptron learning ancestor of backprop",
    "evidenceTier": "strongly_supported",
    "citation": "A-001",
    "verificationDate": null
  },
  {
    "source": "O-018",
    "target": "A-001",
    "type": "challenges",
    "explanation": "Backprop answers the multilayer critique",
    "evidenceTier": "strongly_supported",
    "citation": "A-001",
    "verificationDate": null
  },
  {
    "source": "O-013",
    "target": "A-034",
    "type": "enables",
    "explanation": "Monte Carlo Tree Search in AlphaGo",
    "evidenceTier": "strongly_supported",
    "citation": "A-034",
    "verificationDate": null
  },
  {
    "source": "O-013",
    "target": "P-221",
    "type": "enables",
    "explanation": "Self-consistency is Monte-Carlo voting",
    "evidenceTier": "strongly_supported",
    "citation": "P-221",
    "verificationDate": null
  },
  {
    "source": "O-015",
    "target": "A-036",
    "type": "enables",
    "explanation": "Game theory underlies GAN minimax",
    "evidenceTier": "strongly_supported",
    "citation": "A-036",
    "verificationDate": null
  },
  {
    "source": "O-007",
    "target": "P-322",
    "type": "enables",
    "explanation": "Memex vision underlies retrieval/RAG",
    "evidenceTier": "strongly_supported",
    "citation": "P-322",
    "verificationDate": null
  },
  {
    "source": "P-001",
    "target": "P-366",
    "type": "depends_on",
    "explanation": "GLM is built on the Transformer architecture",
    "evidenceTier": "strongly_supported",
    "citation": "GLM 2103.10360",
    "verificationDate": null
  },
  {
    "source": "P-013",
    "target": "P-366",
    "type": "combines",
    "explanation": "GLM blank-infilling incorporates BERT-style bidirectional masked prediction",
    "evidenceTier": "strongly_supported",
    "citation": "GLM 2103.10360",
    "verificationDate": null
  },
  {
    "source": "P-010",
    "target": "P-366",
    "type": "combines",
    "explanation": "GLM unifies GPT-style autoregressive generation with bidirectional context",
    "evidenceTier": "strongly_supported",
    "citation": "GLM 2103.10360",
    "verificationDate": null
  },
  {
    "source": "P-020",
    "target": "P-366",
    "type": "parallel_development",
    "explanation": "GLM and T5 are parallel unified pretraining-objective approaches",
    "evidenceTier": "strongly_supported",
    "citation": "GLM 2103.10360",
    "verificationDate": null
  },
  {
    "source": "P-366",
    "target": "P-363",
    "type": "parallel_development",
    "explanation": "GLM and Qwen are parallel open-weight bilingual model families",
    "evidenceTier": "probable",
    "citation": "model families",
    "verificationDate": null
  },
  {
    "source": "P-500",
    "target": "P-501",
    "type": "conceptual_ancestor",
    "explanation": "The circuits program on vision networks was extended to Transformers",
    "evidenceTier": "strongly_supported",
    "citation": "source pages",
    "verificationDate": null
  },
  {
    "source": "P-001",
    "target": "P-501",
    "type": "depends_on",
    "explanation": "The framework reverse-engineers the Transformer attention block",
    "evidenceTier": "direct",
    "citation": "P-501",
    "verificationDate": null
  },
  {
    "source": "P-501",
    "target": "P-502",
    "type": "extends",
    "explanation": "Induction heads were identified using the circuits framework",
    "evidenceTier": "direct",
    "citation": "P-502",
    "verificationDate": null
  },
  {
    "source": "P-502",
    "target": "P-012",
    "type": "provides_evidence_for",
    "explanation": "Induction heads give a mechanistic account of in-context learning",
    "evidenceTier": "strongly_supported",
    "citation": "P-502",
    "verificationDate": null
  },
  {
    "source": "P-001",
    "target": "P-503",
    "type": "depends_on",
    "explanation": "Superposition is studied in small Transformer-style networks",
    "evidenceTier": "strongly_supported",
    "citation": "P-503",
    "verificationDate": null
  },
  {
    "source": "P-503",
    "target": "P-504",
    "type": "extends",
    "explanation": "Dictionary learning resolves the superposition described in toy models",
    "evidenceTier": "direct",
    "citation": "P-504",
    "verificationDate": null
  },
  {
    "source": "P-510",
    "target": "P-100",
    "type": "conceptual_ancestor",
    "explanation": "The bitter lesson anticipated compute-driven scaling over hand-design",
    "evidenceTier": "strongly_supported",
    "citation": "book",
    "verificationDate": null
  },
  {
    "source": "P-513",
    "target": "P-100",
    "type": "conceptual_ancestor",
    "explanation": "Data-centric performance foreshadowed the scaling-law regime",
    "evidenceTier": "strongly_supported",
    "citation": "book",
    "verificationDate": null
  },
  {
    "source": "P-514",
    "target": "P-100",
    "type": "provides_evidence_for",
    "explanation": "Documented the compute growth that scaling laws later formalized",
    "evidenceTier": "strongly_supported",
    "citation": "book",
    "verificationDate": null
  },
  {
    "source": "P-100",
    "target": "P-511",
    "type": "depends_on",
    "explanation": "The scaling hypothesis draws on the empirical scaling laws",
    "evidenceTier": "strongly_supported",
    "citation": "book",
    "verificationDate": null
  },
  {
    "source": "P-100",
    "target": "P-512",
    "type": "depends_on",
    "explanation": "Situational Awareness extrapolates scaling and compute trendlines",
    "evidenceTier": "strongly_supported",
    "citation": "book",
    "verificationDate": null
  },
  {
    "source": "P-012",
    "target": "P-520",
    "type": "depends_on",
    "explanation": "WebGPT fine-tunes GPT-3 to browse and answer with citations",
    "evidenceTier": "direct",
    "citation": "P-520",
    "verificationDate": null
  },
  {
    "source": "P-520",
    "target": "P-230",
    "type": "conceptual_ancestor",
    "explanation": "Tool-augmented answering preceded the ReAct agent pattern",
    "evidenceTier": "strongly_supported",
    "citation": "book",
    "verificationDate": null
  },
  {
    "source": "P-012",
    "target": "P-521",
    "type": "applies_to",
    "explanation": "SayCan grounds a language model in a robot value function of affordances",
    "evidenceTier": "direct",
    "citation": "P-521",
    "verificationDate": null
  },
  {
    "source": "P-220",
    "target": "P-522",
    "type": "extends",
    "explanation": "Minerva extends chain-of-thought to quantitative and mathematical reasoning",
    "evidenceTier": "strongly_supported",
    "citation": "P-522",
    "verificationDate": null
  },
  {
    "source": "P-442",
    "target": "P-522",
    "type": "provides_evaluation_for",
    "explanation": "Minerva is evaluated on the MATH benchmark",
    "evidenceTier": "direct",
    "citation": "P-522",
    "verificationDate": null
  },
  {
    "source": "P-523",
    "target": "P-012",
    "type": "challenges",
    "explanation": "Critiques the scale-first paradigm on cost bias and environmental grounds",
    "evidenceTier": "strongly_supported",
    "citation": "book",
    "verificationDate": null
  },
  {
    "source": "P-001",
    "target": "P-530",
    "type": "depends_on",
    "explanation": "DALL-E autoregressively models image tokens with a Transformer",
    "evidenceTier": "direct",
    "citation": "P-530",
    "verificationDate": null
  },
  {
    "source": "P-301",
    "target": "P-530",
    "type": "applies_to",
    "explanation": "CLIP is used to rerank DALL-E samples",
    "evidenceTier": "direct",
    "citation": "P-530",
    "verificationDate": null
  },
  {
    "source": "P-001",
    "target": "P-531",
    "type": "depends_on",
    "explanation": "Whisper is an encoder-decoder Transformer for speech",
    "evidenceTier": "direct",
    "citation": "P-531",
    "verificationDate": null
  },
  {
    "source": "P-020",
    "target": "P-532",
    "type": "depends_on",
    "explanation": "Imagen conditions image diffusion on a frozen T5 text encoder",
    "evidenceTier": "direct",
    "citation": "P-532",
    "verificationDate": null
  },
  {
    "source": "P-308",
    "target": "P-532",
    "type": "parallel_development",
    "explanation": "Imagen and Latent Diffusion are parallel text-to-image diffusion systems",
    "evidenceTier": "strongly_supported",
    "citation": "P-532",
    "verificationDate": null
  },
  {
    "source": "P-308",
    "target": "P-533",
    "type": "depends_on",
    "explanation": "Sora is a diffusion Transformer applied to video",
    "evidenceTier": "strongly_supported",
    "citation": "P-533",
    "verificationDate": null
  },
  {
    "source": "P-001",
    "target": "P-540",
    "type": "depends_on",
    "explanation": "AlphaFold Evoformer and structure module are built on attention",
    "evidenceTier": "strongly_supported",
    "citation": "AlphaFold Nature 2021",
    "verificationDate": null
  },
  {
    "source": "A-034",
    "target": "P-540",
    "type": "conceptual_ancestor",
    "explanation": "Continues DeepMind line of deep learning for hard structured problems",
    "evidenceTier": "strongly_supported",
    "citation": "book",
    "verificationDate": null
  },
  {
    "source": "A-001",
    "target": "P-550",
    "type": "depends_on",
    "explanation": "Trains a student against a teacher soft output distribution via backpropagation",
    "evidenceTier": "strongly_supported",
    "citation": "Hinton 2015",
    "verificationDate": null
  },
  {
    "source": "P-550",
    "target": "P-365",
    "type": "enables",
    "explanation": "Open small models such as Gemma distil from larger frontier teachers",
    "evidenceTier": "strongly_supported",
    "citation": "model reports",
    "verificationDate": null
  },
  {
    "source": "A-001",
    "target": "P-551",
    "type": "depends_on",
    "explanation": "Regularizes weight updates in a backprop-trained network",
    "evidenceTier": "strongly_supported",
    "citation": "EWC 2017",
    "verificationDate": null
  },
  {
    "source": "P-551",
    "target": "P-552",
    "type": "conceptual_ancestor",
    "explanation": "Framed the catastrophic-forgetting problem that test-time memory addresses",
    "evidenceTier": "strongly_supported",
    "citation": "Titans",
    "verificationDate": null
  },
  {
    "source": "P-001",
    "target": "P-552",
    "type": "depends_on",
    "explanation": "Titans augments an attention backbone with a learned long-term memory module",
    "evidenceTier": "direct",
    "citation": "Titans 2501.00663",
    "verificationDate": null
  },
  {
    "source": "P-425",
    "target": "P-552",
    "type": "parallel_development",
    "explanation": "Parallel line of work on efficient sequence memory beyond attention",
    "evidenceTier": "strongly_supported",
    "citation": "Titans",
    "verificationDate": null
  },
  {
    "source": "A-040",
    "target": "A-019",
    "type": "enables",
    "explanation": "AlexNet was trained and evaluated on the ImageNet dataset and challenge",
    "evidenceTier": "direct",
    "citation": "AlexNet 2012",
    "verificationDate": null
  },
  {
    "source": "A-040",
    "target": "P-301",
    "type": "enables",
    "explanation": "Large labelled image corpora underpin later vision-language pretraining",
    "evidenceTier": "strongly_supported",
    "citation": "CLIP",
    "verificationDate": null
  },
  {
    "source": "P-366",
    "target": "P-367",
    "type": "conceptual_ancestor",
    "explanation": "The modern open-weight GLM series continues the GLM lineage",
    "evidenceTier": "strongly_supported",
    "citation": "zai-org HF",
    "verificationDate": null
  },
  {
    "source": "P-364",
    "target": "P-367",
    "type": "depends_on",
    "explanation": "GLM-5 adopts DeepSeek-style sparse attention",
    "evidenceTier": "strongly_supported",
    "citation": "GLM-5 report 2602.15763",
    "verificationDate": null
  },
  {
    "source": "P-112",
    "target": "P-367",
    "type": "depends_on",
    "explanation": "GLM-4.5 and 5.x are large sparse mixture-of-experts models",
    "evidenceTier": "strongly_supported",
    "citation": "GLM-4.5 report 2508.06471",
    "verificationDate": null
  },
  {
    "source": "P-001",
    "target": "P-368",
    "type": "depends_on",
    "explanation": "BLOOM is a decoder-only Transformer",
    "evidenceTier": "strongly_supported",
    "citation": "BLOOM 2211.05100",
    "verificationDate": null
  },
  {
    "source": "P-360",
    "target": "P-368",
    "type": "parallel_development",
    "explanation": "BLOOM and LLaMA are parallel large open language models",
    "evidenceTier": "strongly_supported",
    "citation": "open ecosystem",
    "verificationDate": null
  },
  {
    "source": "P-369",
    "target": "P-308",
    "type": "enables",
    "explanation": "LAION-5B provided the image-text data used to train Stable Diffusion",
    "evidenceTier": "strongly_supported",
    "citation": "LAION",
    "verificationDate": null
  },
  {
    "source": "P-301",
    "target": "P-369",
    "type": "conceptual_ancestor",
    "explanation": "LAION set out to openly reproduce CLIP-scale image-text data",
    "evidenceTier": "strongly_supported",
    "citation": "LAION",
    "verificationDate": null
  }
]