LOSSLESS HANDOFF — LANE 00 INTEGRATION LAYER Solutions-Based AI Studio Date: 2026-04-14 Version: v2 — Updated from Lane 05 session (PT-05-005 complete) From: Claude thread 00 - Integration Layer (session ending due to context length) To: New Claude thread 00 - Integration Layer (continuation) SECTION 1 — WHO YOU ARE You are Lane 00 — Integration Layer for the Solutions-Based AI Studio. Your role is the connection point between all workstreams and lanes. You do NOT own benchmarks, prompts, training data, or infrastructure individually. You OWN the wiring between them, cross-lane synthesis, master state, and handoff intake. Tag: Solutions-Based AI — Turn vision into solutions Owner: Kip Hohman (kipper) — retired spring manufacturer, MCSE, Navy veteran. Background: Custom spring manufacturing — large actuator springs (2.500 inch bar) to small Inconel/Monel/Cobalt/Nickel springs for valves and industrial products. Goal: Build a local AI studio to ingest domain knowledge and produce grounded engineering answers, quotes, and solutions. SECTION 2 — HARDWARE ACTIVE MACHINE: gaming-ai-System CPU: AMD Ryzen 9 9950X RAM: 64GB DDR5-6000 (DOCP/EXPO active — confirmed this session) GPUs: 2x RTX 3090 24GB (48GB total VRAM) OS: Ubuntu 24.04 User: kipper IP: 10.0.0.200 (dynamic — static not yet set) Tailscale IP: 100.119.173.73 (CORRECTED — was 100.107.111.126) NOTE: gaming-ai-System is Ubuntu on second drive of kipper-system hardware. kipper-system (100.107.111.126) is the same physical machine, different OS drive. Network: 10GbE enp12s0 FUTURE MACHINE: Ai-System Platform: TRX40 / RAM: 128GB DDR4 / GPUs: 6x RTX 3090 24GB Status: Parts purchased, build pending — Ubuntu install next session INGESTION FLEET: 12x Xeon rigs GPUs: 5x RTX 3060 12GB per rig / Status: Ubuntu not yet installed NAS: QNAP NAS2991B8 IP: 10.0.0.100 (static) / Tailscale: 100.81.13.64 Mount: /mnt/Vault (SMB 3.0) Usable: ~31TB / Speed: 10GbE confirmed SECTION 3 — RUNNING SERVICES (confirmed 2026-04-14) Open WebUI http://10.0.0.200:8080 Running code-server 127.0.0.1:8081 Running (localhost only — intentional security) text-gen-webui http://10.0.0.200:7860 Running text-gen-webui API http://127.0.0.1:5000/v1 Running (OpenAI-compatible) PostgreSQL 16 localhost:5432 Running Claude Code CLI Authenticated (Max subscription) Docker 172.17.0.1 Active vault-http localhost:8888 Running (systemd service — auto-start) TAILSCALE NETWORK (confirmed 2026-04-14): gaming-ai-System 100.119.173.73 Connected (Ubuntu — active build machine) kipper-system 100.107.111.126 Last seen Apr 8 (Windows drive — same hardware) nas2991b8 100.81.13.64 Connected (QNAP) diskstation 100.125.195.66 Connected desktop-hka4q2m 100.93.143.87 Last seen Apr 8 SESSION PERSISTENCE (PT-05-005 COMPLETE): vault-http systemd service: running on port 8888, serving /mnt/Vault/00_Overview Tailscale Funnel: active background persistent (gaming-ai-System) Cloudflare Tunnel: LIVE — ephemeral URL changes on restart Current URL: https://roof-tribunal-federal-humidity.trycloudflare.com Claude fetch confirmed working: web_fetch → Cloudflare URL → Vault file NEXT: Named Cloudflare tunnel for permanent URL (scheduled next session) FETCH PATTERN FOR NEW THREADS: At session start, user pastes current Cloudflare tunnel URL. Claude fetches: https://[cloudflare-url]/lane00_lossless_handoff_2026-04-14.md Full context loaded autonomously — no copy/paste required. SECTION 4 — MODEL INVENTORY (gaming-ai-System) qwen2.5-72b/qwen2.5-72b-instruct-q4_k_m-00001-of-00012.gguf LOADED (default) qwen3-32b/Qwen3-32B-Q4_K_M.gguf Available qwen2.5-32b/qwen2.5-32b-instruct-q4_k_m-00001-of-00005.gguf Available qwen2.5-14b/Qwen2.5-14B-Instruct-Q4_K_M.gguf Available qwen2.5-7b/qwen2.5-7b-instruct-q8_0-00001-of-00003.gguf Available gemma-4-26b-a4b/google_gemma-4-26B-A4B-it-Q4_K_M.gguf Available DUAL-GPU BASELINE: gpu-layers=-1, tensor_split=1,1 SECTION 5 — DATABASE STATE Engine: PostgreSQL 16 + pgvector 0.6.0 Host: localhost:5432 / DB: ai_studio / User: vault_user / Password: vault_secure_2026 Connection: PGPASSWORD=vault_secure_2026 psql -U vault_user -d ai_studio -h localhost CHUNKS TABLE: id, document_id, chunk_index, content (text), token_count, embedding vector(384) INDEX: chunks_embedding_idx — hnsw (embedding vector_cosine_ops) — ACTIVE DOCUMENTS IN PGVECTOR (691 total chunks): ROG Crosshair X870E Hero Manual 360 chunks TRX40 PRO 10G Motherboard 305 chunks alloy_steels.md 2 chunks copper_alloys.md 5 chunks high_carbon_steels.md 3 chunks nickel_alloys.md 6 chunks spring_materials_index.md 7 chunks stainless_steels.md 3 chunks EMBEDDING MODEL (locked): sentence-transformers/all-MiniLM-L6-v2 (384 dims) PIPELINE ORDER: batch → chunk → embed → query Always prefix Docling: CUDA_VISIBLE_DEVICES="" SECTION 6 — STUDIO_ROUTER.PY STATE Location: ~/apps/studio_router.py Vault: /mnt/Vault/02_AI_Stack/pipelines/scripts/studio_router.py STATUS: WORKING — confirmed live test passed this session Query: What is the maximum operating temperature for Inconel 718? Result: Auto-routed 72b, 4 RAG chunks, Answer: 1100 degrees F, Unicode clean, Training candidate YES CONFIRMED FEATURES: ascii_filter() — strips non-ASCII (WL-01 WL-03) contains_cyrillic() — WL-02 detection U+0400 to U+04FF contains_fullwidth_digit() — WL-03 detection U+FF10 to U+FF19 unicode_integrity_check() — combined WL-02 + WL-03 wrap_governance() — governance packet with unicode_result param Auto-routing — domain keywords trigger 72b automatically MODEL_TIER MAP: 14b: qwen2.5-14b/Qwen2.5-14B-Instruct-Q4_K_M.gguf 32b: qwen3-32b/Qwen3-32B-Q4_K_M.gguf 72b: qwen2.5-72b/qwen2.5-72b-instruct-q4_k_m-00001-of-00012.gguf ROUTING KEYWORDS: 72B: specification, tolerance, alloy, certification, valve spring, inconel, elgiloy, hastelloy, operating temperature, corrosion, compare, tradeoff, analyze, multi, constraint, recommend, evaluate, synthesize, design, architecture 14B: part number, quote, price, lead time, quantity, summarize, rewrite, clean, format, list, what is, define, explain briefly 32B: default SECTION 7 — LANE STRUCTURE CLAUDE LANES: Lane 00 — Integration Layer (this thread) Lane 05 — Infrastructure (separate Claude thread) CHATGPT LANES: Lane 01 — Assistant Design Lane 02 — System Prompt Lab Lane 03 — Behavior Tests Lane 04 — Training Data / LoRA Prep Support — Technical and Operational LANE PROTOCOL (locked 2026-04-14): All cross-lane work requires formal handoff packet FROM / TO / DATE / TASK ID / TASK / INPUT / OUTPUT NEEDED / DEPENDENCY / RETURN No lane executes out-of-scope work All completions report to Lane 00 SECTION 8 — WATCHLIST WL-01: Chinese character insertion at noun positions Examples: plan=计划, cloud=云, platform=平台, time=時間 Detection: ascii_filter() catches it Training: 5 contrast pairs HD-72B-0067 through 0071 WRITTEN WL-02: Cyrillic substitution in trade names and body text Examples: Monел, Incоnel, inсhes Detection: contains_cyrillic() U+0400 to U+04FF Training: 5 contrast pairs HD-72B-0062 through 0066 WRITTEN Test suite: BT-WL02-001 through 007 saved to Vault WL-03: Fullwidth digit insertion at numeric positions Examples: April 11=April １１, 0.625=0.6２５, lot 2026-0412=2026-0４12 Detection: contains_fullwidth_digit() U+FF10 to U+FF19 Training: 5 contrast pairs HD-72B-0057 through 0061 WRITTEN Test suite: BT-WL03-001 through 007 built by Lane 03 WL-04: Technical fact mutation — MOST DANGEROUS Examples: Inconel 718→778, 2.250→2.225 inches, 180→179 lb/in Detection: NOT automatable — requires semantic comparison Training: 5 contrast pairs HD-72B-0052 through 0056 WRITTEN Strategy: PT-03-003 Lane 03 building SECTION 9 — TRAINING PROGRAM STATE File: 72b_psr_training_mcr_v1_2026-04-12.jsonl Target: Qwen 2.5 72B QLoRA fine-tune multi_constraint_rewrite: 50/50 LOCKED bounded_synthesis: 50/50 LOCKED bounded_drafting: 10/50 IN PROGRESS (Lane 04 batch 2 building) bounded_decision_support: 0/50 PENDING vague_to_clear_heavy: 0/50 PENDING Watchlist contrast pairs: 20/20 LOCKED TOTAL: 110/250 (44%) Next packet ID: HD-72B-0100 (Lane 04 batch 2) SECTION 10 — PENDING TASK REGISTER PT-03-003 | Lane 03 | WL-04 detection strategy proposal | ACTIVE building PT-04-006 | Lane 04 | bounded_drafting batch 2 HD-72B-0100-0109 | ACTIVE building PT-05-003 | Lane 05 | WL-03 + combined unicode_integrity_check | BLOCKED awaiting packet PT-04-007 | Lane 04 | bounded_decision_support 50 examples | PENDING PT-04-008 | Lane 04 | vague_to_clear_heavy 50 examples | PENDING PT-00-003 | Lane 00 | JSONL generation 110 candidates | PENDING PT-00-005 | Lane 00 | studio_master_state.md write to Vault | PENDING PT-00-006 | Lane 00 | pending_tasks.md write to Vault | PENDING PT-05-006 | Lane 05 | Named Cloudflare tunnel — permanent URL | PENDING PT-05-007 | Lane 05 | vault_ledger.json + write/confirm/sync | PENDING COMPLETED THIS CYCLE: PT-05-001 | Lane 05 | Input folder purge handler | COMPLETE PT-05-002 | Lane 05 | WL-02 Cyrillic detection wired | COMPLETE PT-05-005 | Lane 05 | Tailscale HTTP server port 8888 | COMPLETE CANCELLED: PT-05-004 | Google Drive sync — security concern — replaced by PT-05-005 SECTION 11 — SESSION PERSISTENCE (UPDATED — NOW LIVE) Architecture: /mnt/Vault/00_Overview/ served via vault-http systemd service (port 8888) Tailscale Funnel: persistent background service on gaming-ai-System Cloudflare Tunnel: active — provides public HTTPS access for Claude web_fetch STATUS: LIVE — confirmed working 2026-04-14 Claude fetched lane05_handoff_2026-04-14.md successfully via Cloudflare URL Full session context loaded autonomously — no copy/paste required CURRENT CLOUDFLARE URL (ephemeral — changes on cloudflared restart): https://roof-tribunal-federal-humidity.trycloudflare.com NEW THREAD STARTUP PROCEDURE: 1. Confirm cloudflared is running on gaming-ai-System 2. Get current Cloudflare URL (check terminal or restart cloudflared) 3. Paste URL into new Claude thread 4. Claude fetches: https://[url]/lane00_lossless_handoff_2026-04-14.md 5. Full context loaded — work begins immediately PERMANENT URL — NEXT SESSION: PT-05-006: Named Cloudflare tunnel Once complete: fixed URL, no restart needed, paste once and never again SECTION 12 — VAULT VERSIONING SYSTEM (DESIGNED — NOT YET BUILT) Architecture designed in Lane 05 session 2026-04-14: vault_ledger.json — tracks every handoff version, state, timestamp Version states: RECORDED / PENDING / MISSING / CURRENT Naming: lane00_handoff_v0001.md, v0002.md etc. Three write moments per thread: open / mid / close Sync script: reads ledger, merges PENDING → RECORDED, updates ledger Rule: new thread only loads RECORDED versions — never PENDING Build task: PT-05-007 — scheduled next Lane 05 session SECTION 13 — LOCKED DECISIONS 1. NAS: SMB 3.0 / /mnt/Vault canonical 2. Vault credentials: /etc/vault-credentials chmod 600 3. GPU 0 reserved: CUDA_VISIBLE_DEVICES="" for Docling 4. Embedding: all-MiniLM-L6-v2 384 dims locked 5. DB: ai_studio/vault_user/vault_secure_2026/localhost:5432 6. API: http://127.0.0.1:5000/v1 OpenAI-compatible 7. code-server: 127.0.0.1:8081 localhost only security design 8. input/ = canonical drop zone 9. Pipeline: batch→chunk→embed→query 10. Dual-GPU: gpu-layers=-1, tensor_split=1,1 11. HNSW index active on chunks table 12. Purge handler active in docling_batch.py 13. SUPPORTED extensions: .pdf .docx .pptx .html .xlsx .md .txt 14. Session persistence: Cloudflare Tunnel → vault-http → Vault (LIVE) 15. Lane 00 = Integration Layer 16. Prompt placement: Character→Context ONLY in text-gen-webui 17. Tailscale Funnel: background persistent on gaming-ai-System port 8888 18. Vault versioning: vault_ledger.json system (designed, build pending) SECTION 14 — SPRING DOMAIN KNOWLEDGE KEY APPLICATION LOOKUPS: Petroleum refinery valve spring → Inconel 718 (ASTM B637) max 1100F Marine application → Monel 400 (AMS 7233) or SS 316 Hydrofluoric acid → Monel 400 Sulfuric acid → Monel 400 or Hastelloy C276 Chloride environments → SS 316 or Hastelloy C276 or Elgiloy Non-magnetic high strength → Elgiloy (AMS 5833) max 850F Highest temperature → Inconel 718 or X750 (up to 1100F) Shock loads → Chrome Silicon (ASTM A401) or Chrome Vanadium (ASTM A231) Cryogenic → Monel 400 or Inconel 600 Electrical conductivity → Beryllium Copper or Phosphor Bronze Constant modulus instruments → NiSpan C (AMS 5225) MATERIALS INGESTED (691 chunks in pgvector): High Carbon Steels: Music Wire A228, Hard Drawn MB A227, Oil Tempered MB A229 Alloy Steels: Chrome Vanadium A231/A232, Chrome Silicon A401/A877 Stainless: SS 302/304 A313, SS 316 A313, SS 17-7 PH AMS 5678 Copper Alloys: Monel 400 AMS 7233, Monel K500 QQ-N-286, Phosphor Bronze B159, Beryllium Copper B197, Brass B134 Nickel Alloys: Inconel 600 B166, Inconel 718 B637, Inconel X750 AMS 5698/5699, Elgiloy AMS 5833, Hastelloy C276 B574, NiSpan C AMS 5225 SECTION 15 — FIRST ACTIONS FOR NEW THREAD 1. Confirm Cloudflare tunnel is running and paste URL into thread. Claude fetches: https://[url]/lane00_lossless_handoff_2026-04-14.md 2. Confirm studio_router.py working: python3 ~/apps/studio_router.py "What is the maximum operating temperature for Inconel 718?" --debug Expected: Auto-routes 72b, retrieves RAG chunks, answers 1100 degrees F, Unicode clean 3. Collect lane confirmations: Lane 03: PT-03-003 WL-04 detection strategy Lane 04: PT-04-006 batch 2 HD-72B-0100 through 0109 Lane 05: PT-05-003 unicode check wired Lane 05: PT-05-006 named Cloudflare tunnel Lane 05: PT-05-007 vault_ledger.json built 4. Write Vault state files: nano /mnt/Vault/00_Overview/studio_master_state.md nano /mnt/Vault/00_Overview/pending_tasks.md 5. Run four-query live test suite: "What is the maximum operating temperature for Inconel 718?" "What material should I use for a spring in hydrofluoric acid service?" "What is the ASTM specification for Chrome Silicon valve springs?" "Compare Inconel 718 and Inconel X750 for high temperature valve springs" SECTION 16 — HANDOFF RUNNER Location: ~/apps/lane00_handoff_runner.py Current passing: 12 checks Run: CUDA_VISIBLE_DEVICES="" python3 ~/apps/lane00_handoff_runner.py END OF LOSSLESS HANDOFF v2 Solutions-Based AI Studio — Lane 00 Integration Layer Updated: 2026-04-14 — Lane 05 PT-05-005 complete, session persistence live Pick up at Section 15 First Actions