Compile
Turn language into one of the runtime's typed spatial operations.
Operational context before retrieval commitment.
The Context Resolver turns a natural-language spatial request into a typed operation, determines which world-state facts and relationships are required, names the deterministic authority allowed to decide the result, and returns an immutable evidence contract alongside a bounded fixture context it can justify.
SFT is available as bounded, read-only infrastructure.
D25 already contained a typed graph, metric frames, deterministic solvers, and immutable revisions. This release adds the missing adoption boundary: a resolver service that can be called through ordinary JSON or through a stateless Streamable HTTP MCP endpoint.
Contract
Declare the entities, relations, frames, measurements, and precision required before retrieval.
Resolve
Construct bounded context with full-scene, lexical, graph-neighborhood, or operation-directed logic.
Gate
Return UNKNOWN when required evidence is absent rather than allowing a model to invent it.
Execute
Route the operation to a named deterministic solver and preserve its proof trace.
Integrate
Expose the same behavior to web apps, agents, ChatGPT, Claude, Cursor, or another MCP-capable host.
IMPLEMENTED · JSON API + STATELESS MCP
Compile the operation before selecting context.
Pick a world and ask a bounded spatial question. The resolver identifies the required entities, relationships, measurements, revision identity, and factual authority before the selected context is handed to a model or solver.
PUBLIC ENDPOINTS
POST /api/sft/resolvePOST /api/mcpOne endpoint for MCP-capable model hosts.
{
"mcpServers": {
"glyphd-sft": {
"url": "https://www.glyphd.com/api/mcp"
}
}
}The public server is read-only and fixture-bounded. It exposes tools for operation resolution, strategy comparison, entity inspection, and typed-command execution. It exposes the four authored fixture worlds as resources. No tool can commit accepted world state.
B18 R1 passed safety and failed the efficiency gate.
The frozen receipt runs six operation classes plus one missing-evidence control under full-scene, lexical top-k, one-hop graph, and SFT operational retrieval. It records required-fact recall, irrelevant context, serialized bytes and tokens, local latency distributions, exact solver parity, false-authority behavior, and source/revision completeness.
SFT achieved complete required-fact recall, zero false authoritative executions, complete solver parity, and complete revision traceability. Its median serialized context was 78.5827% of full-scene bytes, above the predeclared 60% maximum. The pinned-model holdout was frozen but not run because the deterministic efficiency gate failed. No Vector RAG, GraphRAG, or model-quality advantage is claimed.