{
  "abstract": "Transformer inference repeatedly moves and compares large amounts of state. Sparse attention, retrieval, caching, quantization, and small models reduce this burden, but their gains are often designed independently. The Glyphd fractal-computation program asks whether they can be coordinated through a shared multiscale geometry. The proposal combines space-filling curves, recursive tilings, dynamic spatial indexes, foveated context selection, semantic object graphs, and a Q-Frontier that identifies the smallest unresolved set requiring expensive reasoning. Local detail is processed densely; increasingly distant or settled regions are represented more coarsely. Hilbert, Z-order, or Gosper-style orderings provide deterministic one-dimensional traversal of multidimensional neighborhoods. The geometry does not compute answers and does not compress arbitrary data. It organizes locality, candidate sparsity, cache layout, and evidence-preserving omission. The program has two distinct levels. The first is practical runtime architecture around a compact model: structured outputs, PromptCapsules, prefix reuse, foveated retrieval, small deltas, and software scheduling. The second is model-architecture research: multiscale sparse attention and locality-preserving state layout. The first can be benchmarked now. The second requires kernels, trained models, and matched baselines before any performance claim.",
  "author": "Hayden Lindley",
  "category": "Model Systems",
  "claims": {
    "claims": [
      {
        "claim_id": "P13-C01",
        "evidence_refs": [
          "P13-S01",
          "P13-S02",
          "P13-S03",
          "P13-S04"
        ],
        "limitations": [
          "Architecture spans multiple maturity levels."
        ],
        "public": true,
        "scope": "Runtime and model-research program.",
        "statement": "Space-filling curves, recursive regions, FOVEA retrieval, Q-Frontier selection, and structured deltas form a coherent multiscale runtime research architecture.",
        "status": "SPECIFIED"
      },
      {
        "claim_id": "P13-C02",
        "evidence_refs": [
          "P13-S02",
          "P13-S04"
        ],
        "limitations": [
          "No complete accepted end-to-end uplift result."
        ],
        "public": true,
        "scope": "Matched Max-Flex experiments.",
        "statement": "The runtime stack can reduce tokens moved and redundant generation for compact models.",
        "status": "AWAITING_LOCAL_RUN"
      },
      {
        "claim_id": "P13-C03",
        "evidence_refs": [
          "P13-S01"
        ],
        "limitations": [
          "No trained model or kernel benchmark is claimed."
        ],
        "public": true,
        "scope": "Future trained architecture.",
        "statement": "A locality-preserving multiscale sparse-attention model outperforms a dense compact transformer at matched quality and budget.",
        "status": "HYPOTHESIS"
      }
    ],
    "disclosure_note": "Claims marked non-public in the authored ledger are omitted from this public artifact.",
    "paper": "p13-fractal-llm-computation",
    "version": "0.1.0",
    "withheld_claim_count": 0
  },
  "date": "2026-07-27",
  "diagrams": {
    "diagrams": [
      {
        "brief": "Tokens or semantic nodes placed in nested regions, traversed by Hilbert/Gosper paths, with dense local and sparse bridge attention.",
        "format": "svg",
        "id": "P13-F01",
        "path": "/diagrams/p13-fractal-llm-computation/P13-F01.svg",
        "status": "source-generated conceptual architecture",
        "title": "Multiscale compute manifold"
      },
      {
        "brief": "Project graph with verified settled regions faded and unresolved high-authority/evidence nodes highlighted as the compute frontier.",
        "format": "svg",
        "id": "P13-F02",
        "path": "/diagrams/p13-fractal-llm-computation/P13-F02.svg",
        "status": "source-generated conceptual architecture",
        "title": "Q-Frontier"
      },
      {
        "brief": "Exact store → FOVEA → PromptCapsule → structured decode → SurfaceDelta → verifier → renderer, with measured quantities.",
        "format": "svg",
        "id": "P13-F03",
        "path": "/diagrams/p13-fractal-llm-computation/P13-F03.svg",
        "status": "source-generated conceptual architecture",
        "title": "Runtime gains stack"
      },
      {
        "brief": "Sparse attention/retrieval neighborhood retaining mandatory source, authority, blocking, and contradiction anchors.",
        "format": "svg",
        "id": "P13-F04",
        "path": "/diagrams/p13-fractal-llm-computation/P13-F04.svg",
        "status": "source-generated conceptual architecture",
        "title": "Evidence-preserving sparsity mask"
      }
    ],
    "paper": "p13-fractal-llm-computation"
  },
  "evidence_state": "SPECIFIED",
  "id": "p13-fractal-llm-computation",
  "institution": "Glyphd Labs",
  "integrity": {
    "claim_ledger_sha256": "92284a7a3778082ac9a713ee960d66c3d966edfcbfcdcbd482abbe118eb3b082",
    "paper_mdx_sha256": "44d846c4248d25fd6e882115de60c77df56cead26e72ffcf444fd1fc65e1e883"
  },
  "public_disclosure_boundary": "This paper publishes the public-safe research architecture, boundaries, benchmarks, and kill criteria. It excludes unpublished model designs, kernel implementations, private training plans, patent-claim language, and confidential experiment data. No trained fractal transformer or performance advantage is claimed.",
  "publication_class": "public-safe working paper",
  "publication_url": "https://www.glyphd.com/papers/p13-fractal-llm-computation",
  "revisions": {
    "paper": "p13-fractal-llm-computation",
    "revisions": [
      {
        "date": "2026-07-27",
        "summary": "First synthesized public-safe institute edition compiled from the Glyphd project corpus.",
        "version": "0.1.0"
      }
    ]
  },
  "schema_version": "1.0.0",
  "slug": "p13-fractal-llm-computation",
  "sources": {
    "notes": "Internal source paths are provenance references. Public publication must expose only approved public-safe excerpts or archived public artifacts.",
    "paper": "p13-fractal-llm-computation",
    "sources": [
      {
        "date": "2026-07-24",
        "id": "P13-S01",
        "path": "CUTTING EDGE OR NEW / Fractal Frameworks for LLMs",
        "title": "Fractal Frameworks for LLMs project conversation"
      },
      {
        "date": "2026-07-15",
        "id": "P13-S02",
        "path": "ZEKE-4B-MAXFLEX-RUNTIME-MASTER-SPEC-v0.1.md",
        "title": "Zeke 4B Max-Flex Runtime Master Spec"
      },
      {
        "date": "2026-07-12",
        "id": "P13-S03",
        "path": "GLYPHDRIVE-FOVEA-SPEC-v0.2.md",
        "title": "GlyphDrive FOVEA Specification"
      },
      {
        "date": "2026-07-24",
        "id": "P13-S04",
        "path": "How_we_can_make_CAS_lightning_fast_for_any_LLM_by_design_or_adaptation.docx",
        "title": "GlyphCAS Design Report"
      },
      {
        "date": "2026-03-27",
        "id": "P13-S05",
        "path": "PHASE5_CHECKPOINT_SUMMARY.md",
        "title": "MicroScope Resolution Frontier checkpoint"
      }
    ],
    "version": "0.1.0"
  },
  "subtitle": "Space-filling curves, recursive regions, Q-Frontier selection, and evidence-preserving sparsity",
  "title": "Locality-Preserving Multiscale Computation for Compact Language Models",
  "version": "0.1.0"
}