Human-rights infrastructure · local-first · branch 0.0.30

Honest progress — 0.0.30

This page is the current operator record. QualiaDB is the engine; Webizen is the habitat. 0.0.30 is the unified Q42 volume release: one .q42 file, no sidecar twins, no separate .c.q42 compress-and-ship artefact.

What 0.0.30 is for

Make the on-disk graph file honest and complete: 256-byte Q42\0 header, embedded Q42LEX, BIDX, FIDX, PIDX, LZ4 SuperBlocks, merkle chain, and fail-closed publication. Desktop chat, anatomy, checkpoints, and the Studio volume console all write and read that same format. GGUF convert writes a sibling P64 plus a helper Q42 that q42 verify accepts.

Unreleased until 0.1.x. Reprocess of existing volumes is allowed and expected.

Not good enough yet

Hashed-only session volumes (no lexicon terms) verify as Incomplete on Full — that is correct, not a silent Pass.
Local LLM tok/s on consumer GPUs is still below specialised CUDA stacks.
Do not use free-form generation alone for high-stakes legal or medical decisions.
Workspace crate Cargo.toml versions are 0.0.30.
WordNet must be ingested as princeton.q42 / wordnet.q42. A twin wordnet.c.q42 is not part of this release.

What is real in 0.0.30

Unified Q42 v3 — one file: header, lexicon, object BIDX, field ranges, postings, LZ4 SuperBlocks, merkle DAG. Sidecar .q42.lex / .q42.bidx and framed .c.q42 are fallback-only.
Volume-set verify — q42 inspect / q42 verify walk root + data children + lexicon shards and SHA-256. Five-field ECC (includes metadata). July-15 model helpers were rewritten to Pass.
Publication fail-closed — public magnets / web-seed / IPFS only when FLAG_PERMISSIVE_COMMONS is set and the Quin scan is clean. Unmarked, Sanctuary, medical, bilateral stay local or SocialWebNet.
P64 + helper Q42 — qualia-cli llm convert writes .p64 weights and a canonical helper .q42. SmolLM2-360M Q4_K_M decode-proxy: The capital of France is Paris.<|im_end|> at 49.4 tok/s on an RTX A2000.
Desktop / Studio — chat compact, environment, WellFair checkpoint write unified v3. SuperBlock inspector and render mmap read v3 first. Studio Q42 volumes lists the vault, inspects, verifies, mints fail-closed magnets, and compacts.
Range SPARQL — Q42QueryMode 4 MiB forbids resident read_all_quins on large vaults; Optional / Union / Bind / HashJoin stay on the range path.
Pages / WASM readers — playground/vfs.js boots the 256-byte header and embedded lex/BIDX/FIDX/PIDX. SPARQL showcase runs execute_ntriples_query on Schema.org’s unified .q42, not mock rows. Ontology search uses the SuperBlock live-Quin count.
Catalogues — schema.org, EARL, Music, Dublin Core and sibling Pages docs/data/**/*.q42 reprocessed as lossless unified volumes.

Try it on this site

Ontology demo — boots Schema.org 30.0 (WordNet is ~127 MB and not hosted on Pages)
SPARQL showcase — live WASM over Schema.org Q42
Zero-heap — real execute_ntriples_query + FNV Quin inspect
Data formats — ingest → one .q42
RDF compare — one-click Schema.org NT vs unified .q42

Prior honesty pages

0.0.28, 0.0.26, and 0.0.24 remain as archives. Prefer this 0.0.30 page for the current graph file.

Updated 2026-08-15 · QualiaDB / Webizen · principal-directed engineering record · no personal circumstances