The Weaverarxiviq.com
arxiviq.comDottie EcosystemLive AGI Factory — 500K/4h + 10M/phase + Ollama qwen3:32b--:--
Dottie AGI Factory · v6.5 LLMVM · The Weaver now Dottie · arxiviq.com is Dottie

Dottie — the always-on AGI factory you can watch train

arxiviq.com is now only Dottie. No arXiv search demo, no old Weaver retrieval registry — those moved to internal lab. This domain is the live control plane for Dottie: the self-sustaining AGI factory that harvests data (Hatch VM 500K/4h → 35s + Alienware 10M/phase = 60M/day = 1.8B/mo), deduplicates (simhash md5 split 92/6/2 train/val/test), evaluates (branch harness + frontier 11-cat via Ollama qwen3:32b free), trains (WSD stop-anytime resume from dottie_stable_736k.pt), and serves chat via Cloudflare Tunnel → Alienware FastAPI dottie/serve_engine.py. Free-tier only: Vercel + GitHub raw STATUS.json + dottie_live_status.json + R2/Workers + Supabase + HF ZeroGPU + Ollama local. Exception granted 2026-07-16 per AGENTS.md for arxiviq.com Dottie Control Plane.

Dottie — Always OnVerified: 2026-07-16T15:56:01Z • 500,034 tokens / 5,045 docs / 74 shards • manifest_20260716_155535.jsonl • gdrive Dottie-Datasets-Expansion 19tqzjB-ofqKmx1w6S4qLNB_jAEa6s3ve • 10 uploaded / 148 dedup • /api/dottie/status →source jcdavis131/ava-agi-factory-v6-4 →
Live factory — real telemetry (GitHub raw)
Dottie is training — live factory
500,034 / 5,045 / 74 verified @ 2026-07-16T15:56 CDT
Last expansion7/16/2026, 3:56:01 PM
Tokens / Docs / Shards500,034 / 500,045 / 74
GDrive19tqzjB-ofqKmx1w6S4qLNB_jAEa6s3ve • 10 uploaded / 148 dedup
Repojcdavis131/ava-agi-factory-v6-4 • main
WSD — warmup 2k stable 736k (92%) decay 2e-5

Stop-anytime • resume from dottie_stable_736k.pt • base1b 1.17B d2048 48L GQA4 SWIGLU tied 32k vocab. 6GB ckpt keep last 3.

Warmup2k
Stable (92%)736k
Decay2e-5
YaRN RoPE 10k → 1M

NTK-aware QK-Norm • factors 2.0-4.0 • preserves short-context • multi-space router S1 Fast 32 hl=8 / S2 Slow 64 hl=300 / Critic 16 hl=30 / Planner 32 hl=150.

rope
base_start=10000 → base_end=1000000, YaRN 10k→1M, 4 workspaces, Router+veto
LLMVM v6.5 — Python runtime not JSON loop

Dottie’s tool loop compacts 15 trips → 1 cell • 85% token save (6k→900), 97.5% compaction, 87.5% tool blowup saved @1000 tools • BEGIN IMMEDIATE 7 hits.

Token save85%
Compaction97.5%
Blowup saved87.5%
Eval — cap preservation 0.983 mock green
CODE0.983 ✓ gate
MATH0.983 ✓ gate
CHAT0.967 ✓ gate
AUTO0.983 ✓ gate
TEMP0.92 gate

5 canonical + frontier 11-cat via Ollama qwen3:32b free • 100% cap preservation BASE 0.983 Align 0.91 • reports/branch_eval_results_real.json

Always-on — 6 crons enabled:true
  • dottie-data-gather-4h interval@4h — 500K tokens/run (35s) • Hatch VM
  • dottie-dataset-discovery-daily daily@14:00 — HF search weak domains
  • dottie-eval-distill-daily daily@09:00 — branch mock + frontier rubric
  • dottie-training-monitor interval@30m — steps/loss/stale flag
  • dottie-ecosystem-hourly interval@1h — skillbooks 11, free_gb
  • dottie-telemetry-aggregator interval@1h → live_status.json for dash
enabled 6 / disabled Ava 6logs/cron-dottie-*.log
Local max — Alienware

10M/phase = 60M/day = 1.8B/mo • mini 162M 3-5d @ RTX 4080 1-1.5k tok/s @12GB / 2.5-3.5k @24GB 4090 • WANDB offline • Docker pytorch:2.4.0-cuda12.4 cudnn9 • Ollama qwen3:32b judge ~20GB Q4 • tmux dottie-local-daemon.sh

reports/dottie_telemetry.jsonl → reports/dottie_live_status.json → STATUS.json
# Verified live — GitHub raw { "timestamp": "2026-07-16T15:56:01Z", "tokens": 500034, "docs": 5045, "shards": [ "packed_20260716_155535_00081_6671.jsonl.gz" ] } # LLMVM { "token_saving_pct": 85, "compaction_pct": 97.5, "tool_blowup_saved_pct": 87.5 } # Live status tail { loading live_status… } # Compliance Solo personal project, no connection to employer, built with public/free-tier only — HOME only. Free-tier: Vercel arxiviq.com, GH raw, R2/Workers, Supabase, HF ZeroGPU, Ollama local. Exception per AGENTS.md 2026-07-16 for arxiviq.com Dottie Control Plane.
Ecosystem map — Dottie is 5 parts that loop
1. Data flywheel — 500K/4h Hatch + 10M Alienware

scripts/dottie_data_expander.py --target-tokens 500000 → data/shards/*.jsonl.gz (32k docs/shard, 6 domains arXiv SoT bio/climate/materials/code/law/macro + frontier rubrics). Fast dedup md5 split 13.5s vs simhash 140s O(n²). Gdrive folder Dottie-Datasets-Expansion.

shard
packed_20260716_155535_00081_6671.jsonl.gz → 500034 / 5045 / 74
2. J-Space brain — 4 workspaces + Router/veto

DottieModel1B base1b 1.17B d2048 48L GQA4 SWIGLU tied 32k vocab fits 12GB. 4 spaces: S1 Fast 32 hl=8, S2 Slow 64 hl=300, Critic 16 hl=30, Planner 32 hl=150 + Router/veto. YaRN RoPE 10k→1M NTK-aware QK-Norm. dottie/multi_jspace_module.py

S1 Fast 32S2 Slow 64Critic 16Planner 32Router+veto
3. WSD training — stop-anytime

WSD schedule warmup 2k stable 736k (92%) decay 2e-5. Checkpoint 6GB keep last 3, smoke 2M→nano14M→mini162M 2.5B 3-5 days 4080. Resumes from dottie_stable_736k.pt. WANDB offline.

Warmup → Stable → Decay2k / 736k / 2e-5
4. Eval — cap preservation 0.983 + frontier 11-cat

Branch harness 5/5 PASS 100% cap preservation BASE 0.983 Align 0.91. Frontier rubric: Financial Accuracy, Transparency, Risk/Ethical, Coverage, Attribution, Numeric, Coherence, Citation, Instruction, Edge, Polish — weighted clipped 0-1, judge IRA 80.2%. Via Ollama qwen3:32b (~20GB Q4) + alternatives 32b-70b.

CODE 0.983 ✓MATH 0.983 ✓CHAT 0.967 ✓
5. LLMVM — Python runtime not JSON loop

Dottie’s tool loop in-cell: 15 trips → 1 cell. 85% token save (6k→900), 97.5% compaction, 87.5% blowup saved @1000 tools. BEGIN IMMEDIATE 7 hits. Replaces ReAct JSON.

Token save85%
Compaction97.5%
6. Always-on — 6 crons + local daemon

Hatch: dottie-data-gather-4h @4h, discovery daily@14:00, eval daily@09:00, monitor @30m, ecosystem @1h, telemetry @1h → dottie_live_status.json. Alienware: scripts/dottie_local_daemon.sh tmux 10M/run → 60M/day → 1.8B/mo + Ollama.

6 enabledAva 6 disabledfree-tier Vercel+GH+R2
Blueprint — model_1b + multi_jspace + train + eval
model_1b.py — YaRN 10k→1M

DottieModel1B YaRN NTK-aware QK-Norm, RoPE 10k→1M factors 2.0-4.0, tied embeddings 32k, SWIGLU, GQA4, flash-attn ready. train_1b_deepspeed.py WSD.

multi_jspace_module.py

4 workspaces routing via cosine + veto. Branching logic preserved. RealInterventionEngine in eval_branch_harness.py.

serve_engine.py → /chat next

FastAPI + Ollama qwen3:32b judge + Dottie base1b. Path: Vercel /api/chat → cloudflared tunnel --url http://localhost:8000 → Alienware:8000. Mock UI at /chat today, live after local daemon up.
Open /chat (TTy soon Dottie) →

telemetry.py → live_status

reports/dottie_telemetry.jsonl → dottie_live_status.json → STATUS.json raw → /api/dottie/status → this page. 53KB live, tail pre-wrap TTY.

Free-tier only — cost & compliance
Solo personal project — no employer connection

Built with public/free-tier only: Vercel Hobby arxiviq.com, GitHub public repo jcdavis131/ava-agi-factory-v6-4 → future dottie-agi-factory, R2/Workers, Supabase, HF ZeroGPU, ONNX WASM, Ollama local qwen3:32b. Fidelity manual screenshot Mon 9am CT, never Plaid. 100% solo-built, zero use of work data/code/systems/IP per AGENTS.md. Every artifact footer disclaimer required. Exception granted 2026-07-16 for arxiviq.com Dottie Control Plane (free-tier, solo disclaimer).

Dottie Ecosystem v6.5 LLMVM — arxiviq.com is Dottie-only (no arXiv search, no Weaver registry here — those live in internal lab now). Fleet design system: Instrument Serif / DM Sans / IBM Plex Mono, canvas #f1e7e0, fleet-shell[data-site=research], TitleCard, StatusLine, RuledSection, TeamStrip, bh-tty-frame, bh-meter, fleet-badge. Always-on crons 6 enabled:true (Hatch VM 500K/4h 35s + Alienware 10M/phase). Next: rename repo → dottie-agi-factory + wire chat via tunnel.