v0.35.0

0.35.0 (2026-04-10)

Features

• 210 mnemonic-specific scenarios, bespoke generator, fix encoding token limit (3ebecc1)
• add conciseness guidance for structured_concepts encoding (dc6dabe)
• add Gemma 4 E2B spoke GGUF export script (f96dbbb)
• add serve_spokes.py for OpenAI-compatible spoke model serving (26613ab)
• add thinking-mode support to bake-off harness (665082b)
• backfill verification metrics + fix chart dot visibility (1b78795)
• continuous learning Phase A — runtime verification & experience collection (#391) (82140a1)
• dashboard encoding quality metrics (EPR, TED, experience buffer) (2545de1)
• dashboard narrative + lifecycle chart overhaul (5ce22bf)
• distribution balance data gen, fix batch_encode source preservation (b1bfd96)
• EXP-15 through EXP-19, Gemma 4 adapter, data pipeline, quality testing (e4b94e7)
• EXP-15-19 training research, Gemma 4 adapter, data pipeline (f6ce427)
• EXP-20 data quality pipeline, targeted data generation, MI300X prep (6d75333)
• EXP-25 faithfulness probe — confirmed at seq_len 2375 (d3b6f3a), closes #381
• EXP-25 faithfulness probe, encoding fixes, MCP HTTP transport (45a7cd5)
• EXP-29 candidate model evaluation — bake-off tooling and research designs (e4a8b01)
• EXP-29 candidate model evaluation + faithful prompt (5d470a9)
• faithful prompt format for encoding — rules-first design (EXP-29) (cfb9dc1)
• fix spoke GGUF export, gist merge bug, bump token limits to 4096 (e9fbfaa)
• Gemma chat template support for embedded provider (e647578)
• Gemma E2B spoke training + continuous learning Phase A + dashboard overhaul (6d0d0eb)
• Gemma E2B spoke training setup (EXP-30) (fc54802)
• layer importance profiling script for structured pruning (f6c1a62)
• MI300X Gemma 4 E2B training infrastructure, wandb logging (dc42349)
• Model Control Center with embedded LLM and runtime switching (d2aec19)
• Model Control Center with embedded LLM inference (ad2a86b)
• procedural generator + 96 handwritten mnemonic scenarios v2 (79ed030)
• prompt ablation reveals faithful variant as best encoding approach (f5bd07c)
• research analytics dashboard overhaul (f7e1e56)
• RotorQ RQ4 quantizer + benchmark scripts (ba8e66d)
• RQ4 GPU inference, RQ3 experiment, spoke fusion, fused GGUF export (b603dbc)
• RQ4 GPU inference, spoke fusion, handoff recall fixes (de1efd5)
• serve MCP over HTTP transport from daemon (65fe6cf)
• serve MCP over HTTP transport from daemon (#384) (47a093f)
• spoke routing infrastructure, llama.cpp inference, TurboQuant reference (f51db44)
• TurboQuant prompt cache compression, EXP-22 registration (f8ccf51)
• update EXP-20 config, pre-register EXP-21 (bottleneck rotation) (040c596)
• v6 smoke test 7/7 stress, add advisory board rule (304d884)
• v7 diverse input generation pipeline (8c30d06)

Bug Fixes

• encoding faithfulness + amend raw_id + dashboard timeline bugs (9e874ab)
• encoding faithfulness, amend raw_id, dashboard timeline (671b1b6)
• FR metric now measures content fields only, not concepts (3b12bde)
• GBNF grammar via chat completions payload, not server-level flag (87ec76a)
• gist merge FK violation, ambiguous column in FTS concept search (042a1e3)
• handoff recall, type-filtered search, consolidation exclusions (0ca58bf)
• inline migration 016 for continuous learning tables (3bc78ab)
• preserve full content in handoff-type memories (831a9fe)
• preserve full content in handoff-type memories (ac17492)
• sparse templates with proper gist mapping, dedup to 51 unique (27a400b)
• SQL flagged_rate query handles legacy "null" string values. (2545de1)
• stress test --checkpoint arg, batch_encode model upgrade, misc fixes (0c1c5d1)
• stress test Gemma support, batched generation, JSON parser (cd9e6c7)
• update TestFormatPrompt for ChatML format (d959878)
• use theme-neutral grey dots for EPR chart (053fff8)
• write SQL NULL for empty flags (not JSON "null" string). (2545de1)

What's Changed

• Remove images from README.md by @CalebisGross in #368
• fix: preserve full content in handoff-type memories by @CalebisGross in #375
• feat: EXP-15-19 training research, Gemma 4 adapter, data pipeline by @CalebisGross in #377
• feat: RQ4 GPU inference, spoke fusion, handoff recall fixes by @CalebisGross in #379
• feat: Model Control Center with embedded LLM inference by @CalebisGross in #380
• fix: encoding faithfulness, amend raw_id, dashboard timeline by @CalebisGross in #385
• feat: serve MCP over HTTP transport from daemon by @CalebisGross in #388
• feat: EXP-25 faithfulness probe, encoding fixes, MCP HTTP transport by @CalebisGross in #389
• feat: EXP-29 candidate model evaluation + faithful prompt by @CalebisGross in #392
• feat: Gemma E2B spoke training + continuous learning Phase A + dashboard overhaul by @CalebisGross in #393
• chore(main): release 0.35.0 by @appsprout-release-please[bot] in #376

Full Changelog: v0.34.1...v0.35.0