File size: 2,171 Bytes
b232e60 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 | # BLIND TEST - home tower to friend's machine, no internet
The proof that matters: the whole system - app, engine, weights - lives in ONE zip
and runs cold on a machine you have never touched.
## AT HOME (online)
pip install huggingface_hub requests
huggingface-cli download Johnny5b/portable-ai-runner --repo-type dataset --local-dir CORTEX
cd CORTEX
python setup.py --model ornith # 5.8 GB Ornith + 639 MB Qwen3-0.6B, sha256-verified
python scripts/make_zip.py # -> CORTEX-DRIVE-<date>.zip (~6.6 GB, zip sha256 printed)
Put BOTH on the USB stick: the zip, and the Python 3.12 Windows/macOS installer
(python.org/downloads) - then the friend machine needs no internet at all.
## AT THE FRIEND'S HOUSE (offline)
1. Copy the zip off the USB, unzip it to a folder (right-click -> Extract All).
2. If the machine has no Python: run the Python installer FROM THE USB STICK.
3. Double-click serve.py (or: python serve.py). No pip, no internet needed -
serve.py runs on the bare standard library; the LAN mesh just switches off.
4. Open Chrome -> http://localhost:8080
5. Models tab -> Scan -> Load next to Ornith-1.5-9B (needs ~16 GB RAM;
on 8 GB machines use Qwen3-0.6B). It reads the GGUF from the folder -
NO internet is touched. Engine pill must read: llama.cpp WASM - drive (offline).
6. Chat. Under every reply you will see:
[TRACE:AWAKE] symbols=... files_read=... lookup_ms=... | [TRACE:GEN] tok=... tok_s=...
Those lines are appended to memories/audit_ledger.jsonl on the drive.
## WHAT TO CHECK (the actual proof)
- engine pill says 'drive (offline)' - never CDN
- capThreads says multi-thread enabled (COOP/COEP ok)
- a trace line appears under every reply
- python -c "import json;print(json.loads(open('memories/audit_ledger.jsonl').readlines()[-1]))"
shows your measurement traces
- turn off Wi-Fi before starting - nothing should care
## EXPECTED (honest)
- Ornith-9B browser load needs RAM: 16 GB comfortable, 8 GB risky (use Qwen3-0.6B).
- tok_s in the traces is UNMEASURED before tonight - whatever your friend's machine
produces IS the first browser-tier measurement. Record it.
|