Compile
Turn supported language into a typed operation or an explicit ambiguity/refusal.
Loading Semantic Fractal Topology…
Loading Semantic Fractal Topology…
SFT is an implemented, browser-local reference runtime for typed spatial state, evidence-gated operations, deterministic solvers, and immutable revision history. It is not a demonstrated improvement to RAG or model reasoning.
B18 R1 passed every hard safety gate and failed its predeclared context-efficiency gate.
In this authored illustration, typed intent selects a bounded route-and-clearance context. B18 R1 did not establish a general context-efficiency advantage.
Median SFT context was 78.5827% of full-scene bytes against a predeclared maximum of 60%. No retrieval or model-quality advantage is claimed.
These browser-local examples use authored state and deterministic calculations. They demonstrate fixture behavior only—not production geometry, sensing, packing, safety, or model quality.
An authored bounded-box fixture compares one object with one opening and reports its limiting clearance.
Live bounded-box calculation. A production version would add the car’s interior route and swept-volume geometry.
Turn supported language into a typed operation or an explicit ambiguity/refusal.
Declare required entities, relations, metric facts, precision, revision, and solver capability.
Select fixture world state and stop when the required evidence contract is incomplete.
Run the named deterministic solver and retain its exact trace and scene revision.
Let an interface or model describe the bounded result without becoming its authority.
Typed, recursively addressable semantic and metric world state.
Validation, indexing, deterministic kernels, adapters, and immutable revisions.
A read-only JSON and MCP surface for fixture-bounded operations and evidence contracts.
Across six operation classes and one missing-evidence control, SFT achieved complete required entity, relation, and metric-fact recall; zero false authoritative executions; complete deterministic solver parity; and complete source/revision traceability.
Its median serialized context was 0.785827 of full-scene bytes, above the 0.600000 eligibility target. The frozen model holdout was therefore not run. The lexical and one-hop graph rows are transparent local baselines, not production Vector RAG or GraphRAG implementations.
No RAG, model-answer, scaling, or production-digital-twin improvement is asserted.
The negative comparative result does not erase the implemented deterministic substrate. It narrows the product claim to what the code and receipts actually establish.
Entities, relations, metric frames, source pointers, and explicit precision live in a validated scene contract.
Fit, reachability, constraints, assembly, service access, and other bounded operations route to named deterministic kernels.
Missing facts produce UNKNOWN, INVALID, AMBIGUOUS, or UNSUPPORTED instead of a guessed positive answer.
Accepted state, agent branches, conflicts, and merge candidates retain exact content and ancestry identities.
glTF, OpenUSD, CAD, BIM/IFC, and robot-scene inputs expose only the projections each adapter actually supports.
This scoped list supports technical positioning only. It is not an exhaustive literature review, patent search, novelty conclusion, or freedom-to-operate opinion.
Microsoft GraphRAG structures unstructured text as entities, relationships, communities, and summaries.
Boundary: B18 R1 did not compare SFT with a production GraphRAG system.
Inspect source ↗ConceptGraphs, Hydra-family systems, SG-Nav, and related work represent objects, rooms, places, geometry, and spatial relationships.
Boundary: SFT does not claim that scene graphs, semantic maps, or hierarchical spatial state are new.
Inspect source ↗Schema-guided scene-graph research already lets models query selected graph data instead of receiving an entire environment.
Boundary: The public SFT result establishes a deterministic contract, not a model-answer advantage.
Inspect source ↗OpenUSD provides an extensible, layered scene-description system for constructing and interchanging large 3D worlds.
Boundary: SFT adapters are bounded projections and do not establish complete native interchange.
Inspect source ↗Learned world models reconstruct, generate, render, and predict spatial environments from multimodal inputs.
Boundary: SFT is an explicit state and solver runtime, not a learned renderer or foundation world model.
Inspect source ↗Natural-language scene queries, digital-twin graphs, graph-based building twins, and immutable twin histories have prior publications and filings.
Boundary: The public materials are not a novelty, patentability, or freedom-to-operate opinion.
Inspect source ↗Typed scenes, solver dispatch, bounded adapters, revisions, and branch candidates.
Open instrument →N20 · ARCHITECTUREThe representation, metric boundary, revision model, adapters, and explicit non-claims.
Read note →B18 R1 · FAILEDFrozen oracle, raw rows, passed safety gates, failed efficiency gate, losses, and next experiment.
Inspect evidence →The available conversation is about typed world state, solver routing, evidence gates, revision history, or bounded adapters—not a proven retrieval or model-quality improvement.
Test declared world-state and solver classes against an agreed, bounded corpus.
Normalize a public-safe subset while preserving omissions, precision, and source authority.
Recompute B18 R1 from the frozen packet or define a new, untuned evaluation corpus.
Discuss a defined implementation surface under separate written technical and legal terms.
Semantic Fractal Topology is a deterministic world-state and solver runtime with explicit evidence and revision boundaries. Its first retrieval-efficiency greenlight failed.