Ontave developer preview
The working developer preview is 0.1.0-alpha.2, distributed as a versioned GitHub archive. Earlier archives are preserved. There is no npm registry publication. No general production-readiness or arbitrary-data accuracy claim has been established.
Available in alpha.2
- Immutable evidence, scoped authorization and fenced durable workers with append-only attempt accounting.
- Real process-kill/restart, bounded shutdown, lost-acknowledgement retries and repeated-crash budget checks.
- Whole-bundle review, authority-scoped identity decisions, corrections and replay respecting withdrawals.
- Additive candidates, isolated extraction/identity/relationship/observation/read evaluation, and atomic activation/replay scheduling.
- Pinned contracts, typed queries, bounded links, signed cursors and actual CLI workflows.
- Operator restore checks across all 18 tables, clean Node/Next.js package consumers and workspace-scoped metrics.
The complete solar/rental lifecycle demonstrations use a deterministic fictional adapter. They establish database behavior, rather than autonomous discovery accuracy.
Try the preview
The browser sandbox explores fictional SDK reference runs. The one-command local playground runs real SDK calls against temporary PostgreSQL, with a fixture adapter and no provider key. You can approve or decline the model, confirm identity matches, accept or reject the interpretation, and inspect the resulting record. Neither is a live-model accuracy demonstration.
Qualification still open
The 60-input maintainer dataset has 179 expected field assertions and 107 relationship checks. Its labels have not been independently reviewed. Twenty executable read tasks are available; they are not an independently authored held-out set.
Live Gemini and Mistral experiments now have recorded precision/recall and failure results. Earlier configurations failed the 95% precision / 90% recall targets. Every experiment is retained, including incomplete runs. New schema or scoring policies produce separately frozen trials. A new three-trial ministral-14b-2512 profile met the aggregate field targets: 97.7–97.8% precision and 96.1–97.2% recall. It used an explicit revised description-equivalence rule, frozen before fresh collection. It is a development result on the same maintainer dataset, not independent complete-MVP certification. Each trial still rejected one invented year-zero date. See measurements and benefits.
Independent label review, live identity/discovery qualification, the remaining per-case acceptance audit and a fresh-developer walkthrough remain open. The npm publication guard requires successful source- and archive-bound evidence. There is no registry publication yet.
Deliberate boundaries
The alpha does not implement physical business DDL, destructive model changes, automatic established-entity merges, a hosted database, Studio, browser/edge runtime, React hooks, MCP integration or autonomous acceptance of facts.
Source uses Apache License 2.0. Follow the installation instructions for the alpha.2 preview archive or a locally built source package.