Human-rights infrastructure · consumer electronics · branch 0.0.24

Honest progress

This page is for operators and reviewers who need the truth, not a demo script. QualiaDB is built so people can keep agency on their own devices. Where we are not yet good enough for life-critical sole reliance, we say so.

Not good enough yet (as sole system)

Local LLM decode on 3B-class models is still far slower than llama.cpp-class CUDA engines on the same A2000-class cards.
Do not use free-form generation alone for high-stakes legal/medical/safety decisions.
Vision/generative-3D is at Phase-1 ABI + CPU reference — not a production detector.

What is real today

Semantic core & SPARQL engine — NQuin / Q42, WAL, modalities (deontic, epistemic, …), and a full query engine: SELECT / ASK / CONSTRUCT / DESCRIBE, OPTIONAL · UNION · MINUS, FILTER incl. EXISTS / NOT EXISTS, sub-SELECT, GRAPH, aggregates, property paths (full transitive closure), DISTINCT, and RDF + result serialisation in every standard format (JSON/XML/CSV/TSV; N-Triples/Turtle/TriG/N3/JSON-LD; CBOR-LD) with real IRI/typed-literal resolution. 304 query tests pass.
Native inference stack — in-process GGUF/P64, wgpu 30 resident decode, INT8 KV, Q4_K SoA, hybrid quant-graph grounding.
Rights-grade quality path — QUALIA_RIGHTS_MODE=1 → quant-graph + fact seed + hybrid masks / deontic obligation on match.
Desktop host — webizen-desktop (Tauri 2) + studio; desktop / CLI / WASM release CI is green and publishes artifacts to the v0.0.24 release, with a minisign-signed auto-updater.
Vision crate — new qualia-vision Phase-1: caller-buffered detect API, epistemic observation quins, CPU reference backend.

WGSL Forge (clarified)

Forge does not train LLMs. It generates, validates, certifies, and tunes compute kernels (WGSL / CUDA C). Training or fine-tuning weights is a separate track. Kernel automation helps decode only when a certified schedule is wired into the resident / CUDA path.

Program tracks

A Inference — device-resident activations, INT4 SoA, less host sync, forge-certified GEMV when it wins.
B Desktop release — installable binaries on GitHub Releases.
C Public honesty — this page + manuals.
D Visual intelligence — plan native-visual-intelligence-and-generative-3d.md · Phase-1 crate landed.

Full program: consumer-human-rights-release-program.md

Consumer operator defaults

# Rights-grade quality (graph hybrid)
$env:QUALIA_RIGHTS_MODE = '1'

# Optional high-stakes fact force
# $env:QUALIA_GRAPH_FORCE = '1'

# Prefer SoA INT4 models on limited VRAM
# *.soa.p64 under your model vault

# Do not enable second DirectML device unless experimenting
# $env:QUALIA_DIRECTML = '1'

Updated 2026-07-13 · QualiaDB / Webizen · principal-directed engineering record