start.sh: LOW_MEM=1 short-circuits to status server only (kill all bg daemons)

start.sh: fix repo clone paths arkashira/* → axentx/* (5 silent 404s)

start.sh: gate 5 heavy harvest launchers behind LOW_MEM

fix(discord-bot): direct 11-provider LLM chain (drop surrogate CLI dep)

v18: simplify hub naming to surrogate-1-{SIZE}B-v1.5 (owner directive)

v18: HUB_ID base-aware + T4 FP8 guard

v18: add Qwen3.5/3.6 aliases (2026-04 releases) — Qwen3.5-9B = sweet spot for T4x2

v18: BASE_MODEL alias resolver — short names just work

v18-safe-defaults: flip SUR_LORA_INIT=loftq + DISABLE_AL=1 as defaults

v18b: add GLM-family Kaggle-feasible bases (glm-4-9b-chat, 4.7-Flash)

v18(round6): Phase 78-96 + Thai/gen-intel datasets + Qwen3 base option

v17(catchup): 18 phases + multi-teacher distill + 5-step merge + EAGLE-3 + TTC

v16(round4-mega): 21 new phases + 30 datasets + 9 tokens + Granite/RoPE-NoPE/YaFSDP

v15(round3): 13 new phases + 25 datasets + 20 swarm tokens + Kimi/DS/GLM techniques

fix(bundle): prefer HF_TOKEN_PRO_WRITE to dodge HF_TOKEN 2500-req/5min cap

v14(unified): Phase -1 ingest + Phase 23-25 daemon/auto-feat/multicloud

v13(into-model): 22 phases + 30+ datasets + multi-agent tokens + frontier kernels

v12(into-model): wire ALL techniques as 14 env-toggle training phases

v11(into-model): add 9 ingest datasets + Phase 0 hygiene + TruthRL ternary GRPO

pivot(v9): SRE-specialist trainer prep — knowledge corpora + 6 role personas

feat(overnight): scoring rubric + report generator + multi-window orchestrator

feat(release): add file:/local-path support + arkship to allowlist

fix(v8/kaggle): bootstrap Kaggle Secrets → os.environ before SFT setup

feat(v8/lora-init): add loftq+pissa hybrid + corda modes (init can't co-exist trivially)

feat(v8+autonomy): research-driven trainer + 4 daemons + 9-layer safety gate

feat(v7): Spectrum-lite + Magpie + active-learning + full swap-and-bench chain

feat(v1.1-extended): pivot to 7B base + EXTENDED stack on Kaggle T4×2

feat(post-bench): 3-branch decision pipeline ready before bench fires

perf(harvest): bump worker fleet 1→16 + push cadence 2× — use the headroom

feat(civo): L40S 48GB training launcher with auto-teardown

fix(watcher+bench): retarget 14B v1.5-mid (Kaggle T4×2 reality)

fix(kaggle-trainer): pick 14B (not 32B) for T4×2 — OOM diagnosis

feat: GitHub Actions harvest worker — 3rd-tier compute leverage

feat(synth-puller): round-robin between 2 PRO ZeroGPU endpoints

fix(auto-bench-watcher): bash structure (else clause was nested wrong)

fix(auto-bench-watcher): single-API-call check (avoid pipe-stdout dup)

feat(auto-bench): watcher fires bench-v1-vs-v15 on first v1.5 checkpoint

feat(kaggle-trainer): hardware-aware base model auto-pick

fix(kaggle-trainer): defensive schema handling — drop interleave_datasets

rename: drop '-lora-' segment from all model names + capitalize v1.5 size

feat(eval): axentx-eval-50 — 50-prompt in-domain DevSecOps eval suite

feat(v1.5): rewrite Kaggle train.py with full R1-R12 technique stack

feat(harvest): lift source-side length caps 6K/8K → 100K/200K chars

feat(v1.5): 32B SFT config + 3-way benchmark (v1 vs base32B vs v1.5)

feat(v2): upgrade base model 32B → 72B (Qwen2.5-72B-Instruct)

fix: disable chutes ladder + zero-gpu cold-start retry

fix(zero-gpu-bridge): /api/predict → /call/respond SSE polling

fix(synth-puller): switch /run/ → /call/<api>/<event_id> SSE poll

fix(bridges+cron): repair 3 broken bridges + wire build-data-pipeline weekly

fix(data-feed): repair dataset feeding gaps