Index of the top-level documentation. One line per file; specs that live
with their crate are indexed in specs.md. Subdirectories:
adr/ (architecture decision records), audits/ (review
reports), diagnoses/ (root-cause write-ups),
ffn/ (FFN backend docs — weight, sparse, walk,
distributed).
| Doc |
One line |
| format.md |
LARQL graph format specification (v0.1.0) |
| vindex3-format.md |
VINDEX3 model-system container format — the living spec (plan/encode/verify semantics), companion to the 3.0 Candidate Specification |
| vindex3-runtime.md |
VINDEX3 runtime stack — Vindex3Runtime, LogitsSession, the KV seam, V3 serving over /v1/completions, /v1/chat/completions, /v1/responses |
| vindex3-experiments.md |
Pre-registered VINDEX3 experimental programme (the V2-0..V2-4 gates) |
| vindex3-ontology-drill.md |
The four-architecture ontology drill (candidate §17.4) — run 2026-08-30, findings F1–F16 |
| lyrw-v2.md |
LYRW v2 — the K3 routed-layer physical-layout gate (storage half of K3) |
| specs.md |
Pointer page: which spec lives with which crate |
| knowledge-pipeline.md |
Stub — placeholder for the knowledge pipeline spec |
| Doc |
One line |
| inference-engine.md |
Inference engine — compute substrate (ADR-0022 layout), attention, FFN backends |
| ffn-graph-layer.md |
FFN graph layer — mmap walk faster than dense (517 ms vs 535 ms) |
| ffn-cache.md |
FFN activation cache — skip recomputation of repeated feature sets |
| ffn/README.md |
FFN backend family — WeightFfn, SparseFfn, WalkFfn, distributed sharding |
| kv-residency-contract.md |
The KV residency contract — window vs storage vs residency, disentangled |
| kv-attention-scaling.md |
KV attention scaling — measurement schema + run hygiene rules |
| metal-kernel-capabilities.md |
Metal kernel capability table (Phase B ground truth audit) |
| mech-interp.md |
Mechanistic-interp surface — hooks, lens, ablation, steering, patching |
| residual-trace.md |
Residual stream trace — decomposition, storage, tiered context |
| multi-modal.md |
Multi-modal support — Phase 0–2 shipped, phases 3–6 design-only |
| virtual-experts-dispatch.md |
Virtual experts — bounded routing into typed, sandboxed WASM compute units |
| confidence.md |
Confidence scoring for query results |
| Doc |
One line |
| vindex-factory.md |
Vindex Factory — recipe-driven, verified, remote-executed builds |
| model-publishing.md |
Republishing models — the 2026-08 manual recovery and the recipes it demands |
| k3-funnel.md |
K3 adapter ladder — GPT-OSS-20B → Kimi Linear → K3 |
| glm5-flash-funnel.md |
GLM-5.3-Flash funnel — admission, census and the two tracks (321 B, KDA + DSA + mHC) |
| dec-funnel.md |
DEC funnel (v0.5, current) — decoupled attention/weights serving |
| dec-funnel-v0.4.md |
DEC funnel v0.4.1 — superseded by dec-funnel.md |
| dec-funnel-v0.2.md |
DEC funnel v0.2 — archived; control plane and gates inherited by reference |
| tts-funnel.md |
TTS funnel — audio-token output (MOSS-TTS-Realtime), ~1.9× realtime CPU |
| quant-obs.md |
Quant-Obs — observer-metric ladder for quantisation error allocation |
| fleet-routing-extensions.md |
Fleet routing extensions FR1–FR4 — spec + frozen pre-registrations |
| fhg.md |
FHG — Fourier heuristic graph programme (behavioural, model-agnostic) |
| authority-control-plane.md |
Authority control plane (EXP-26..38) — layer-mechanism branch closed |
| kimi-precision-topology.md |
Kimi Linear 48B — the PRECISION-1 topology, the first complete REPRESENT chain |
| represent-optimizer-mcp.md |
REPRESENT as a queryable optimiser — map DAG, MCP surface, and the search ladder to MCTS (design) |
| Doc |
One line |
| positioning.md |
LARQL vs ollama, vLLM, llama.cpp — what it is and is not |