Skip to content

Latest commit

 

History

History
82 lines (69 loc) · 5.64 KB

File metadata and controls

82 lines (69 loc) · 5.64 KB

docs/

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).

Formats and specs

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

CLI, language, bindings

Doc One line
cli.md Full larql CLI reference
lql-guide.md LQL quick-start guide
larql-python.md Python bindings for the vindex

Engine and runtime

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

Extraction and knowledge

Doc One line
weight-extraction.md Weight extraction pipeline — model weights → vindex, no bulk forward passes
training-free-insert.md Training-free knowledge insertion — residual capture + feature writes
validation.md Graph validation — extraction faithfulness checks
circuit-types.md Circuit type analysis — gate/down cosine classifies feature roles
findings.md Research findings from querying Gemma 3 4B weight vectors
walk-boundary-sweep.md Walk boundary sweep — correctness across all layer boundaries

Programmes and funnels

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)

Positioning

Doc One line
positioning.md LARQL vs ollama, vLLM, llama.cpp — what it is and is not