De librería técnica a plataforma adoptable

22-32

Release 22-32 · Discipleship ecosystem

11 orthogonal phases that close the active-discipleship loop plus an initial trust baseline.

Ver detalle →

33-38

Release 33-38 · Core at SOTA

6 phases that raise the technical ceiling: embeddings + audio + grammars + VLM + visual + generation. Eighth package (jw-gen) introduced.

Ver detalle →

39-48

Release 39-48 · Runtime trust + community + JS

10 shipped phases: live NLI, provenance, plugin SDK, scaffolding, tracing, judge, chunking, versification, TS port and browser extension.

Ver detalle →

49-56

Release 49-56 · jw-core complete + multilingual SOTA

8 phases that complete jw-core: persistent second-brain, JWPUB/.jwlibrary writers, organized-app schemas, Omnilingual ASR (1672 languages), NLLB-200, multilingual wire-up and JS bridge close-out with anti-drift CI.

Ver detalle →

57-66

Release 57-66 · 2 new packages + live-meeting + biblical KG

13 phases in a single continuous session: clean-room jw-meeting-media (WOL parser from DevTools + Tauri presenter), JW-pure bible-knowledge-graph (607 B.C.E.), opt-in letta-memory with Fernet, marker+markitdown loaders, diarized whisperX + speaker mapping, brain exposed via MCP. Plus 7 MVP+1 follow-ups.

Ver detalle →

65-76

Release 65-76 · Agentic + multimodal + predictive + voice

9 phases that open four new layers: verifiable agentic orchestration with NLI critic (F65-F67), end-to-end multimodal with a physical book camera and a visual quote verifier (F68-F71), diachronic analysis with temporal embeddings (F72) and consented family voice with a 3-layer license gate + Fernet encryption (F76). +480 tests, +15 new MCP tools.

Ver detalle →

77-79

Release 77-79 · Alignment to the published canon

3 phases applying Constitutional AI, RLAIF, and DPO/ORPO to the local fine-tune's concrete problem: not drifting from the published canon. F77 codifies 5 builtin principles in versioned YAML and exposes them to Judge + fidelity_wrap (lazy import, zero cycle). F78 promotes the Judge to a pairwise annotator and introduces SL-CAI (self-critique). F79 adds DPO and ORPO trainers with Unsloth plus three Qwen3.5-0.8B recipes (Apache-2.0) wired into the CLI with a new prepare-preference command. +41 tests, 1,326 passing.

Ver detalle →

80

Release 80 · Mechanistic interpretability, tri-model

Closes the F77-F79 loop: did the model internalise the principles or learn a stylistic shortcut? New jw-interp package with per-principle linear probing (F80.1), steering vectors + activation patching (F80.2), Qwen-Scope adapter (F80.3, TopK residual), Gemma Scope wrapper (F80.4, SOTA JumpReLU with MLP + attention) and a runtime probe store + observational fidelity_wrap Tier 4 (F80.5). Tri-model architecture: production Qwen 0.8B left untouched, lab Qwen 2B and lab Gemma 2B with public SAEs. Cross-family validation. +85 tests, 1,411 passing.

Ver detalle →

81-82

Release 81-82 · Congregational operations + legal cases TJ

Two new application-layer modules. F81 closes the operational loop of the meeting-scheduler with organized-app importer + CRDT diff (F81.0), a roster-editing CLI that honours last_updated so re-imports do not clobber manual edits (F81.1) and a strict Pydantic constraints YAML per congregation with init/lint/show CLI (F81.2). F82 opens the legal module with a Territory catalogue of 30 countries (F82.0 — RU, KP, ER, SG, TJ banned; CN, AZ, BY, VN, MM, TR, CU, KZ restricted; 17 free) that composes the existing LocaleContext without duplicating cultural fields, and the jw-legal plugin registering a legal-cases-tj BrainDomain via entry-point jw_agent_toolkit.brain_domains (F82.1) — 6 NodeTypeSpec + 8 EdgeTypeSpec, CONTRADICTS non-directional sensitive. 2 new packages. +305 tests, 3,318 passing.

Ver detalle →