Johnny5b/portable-ai-runner

CORTEX Portable AI Drive - the complete map

This repo IS the drive. Download the whole tree (or let your agents walk it) and it runs offline on any machine with Python 3 - no setup step, no internet, weights included.

Total: %s across %d top-level entries. Generated %s from the live repo.

THE MAP - what lives where (measured, not estimated)

Path	Size	What it is
index.html	browser app: chat, model loader, mesh roles, symbol retrieval, trace meter
serve.py	server: files + COOP/COEP + POST /api/ledger + optional LAN mesh :8081
setup.py	one-time fetcher (NOW OPTIONAL - weights ship in the repo)
models/cores.json	the three pinned models + measured tok/s + amendment directives
models/weights-manifest.json	pinned repo ids + filenames + sha256 for every weight
models/models.json	what the Models tab loads (Ornith first)
models/knowledge_graph.json	the model's directional self-map
models/gguf/	THE WEIGHTS: Ornith-1.5-9B-Q4_K_M (5.8 GB) + Qwen3-0.6B-Q8_0 (639 MB)
vendor/wllama/	llama.cpp compiled to WASM - the engine the app runs on
symbols/symbol_index.json	the directory: every intent -> symbol -> file star (incl. reserved %%knowledge, %%proposals)
symbols/symbol_evolution_index.json	symbol -> module -> upstream commit -> schema -> manifest
symbols/schemas/	AST-extracted signatures/contracts: %d files (md + json + manifest per module)
modules/	%d mirrored skill repos (PYGGI, TinyverseGP, DEAP, EvoTorch, EVA, isanlp_rst, e2e-microtexts) with provenance
knowledge_vault/	%d real content files: computer_science, natural_sciences, healthcare_safety, cognitive_reasoning
memories/	identity.json, audit_ledger.jsonl (every trace), proposals/, questions/, learned/
scripts/	builders, verifiers, cortex_live_trial.py (the runnable three-phase trial)
stack/	docker-compose: code-server :8443 + runner :8080 + ollama
docs/	OWNER-INTENT-CANON, CORTEX-POINTER-ARCHITECTURE, USB-ONE-STEP, MEASUREMENTS, runbooks
CORTEX-ZIP-MANIFEST.json	(if present) sha256 of every file at zip build time

HOW IT EXECUTES - the actual flow, file by file

1. python serve.py - starts ThreadingHTTPServer on :8080 (app + weights + POST /api/ledger); LAN mesh on :8081 only if the optional websockets package is installed.
2. Chrome opens http://localhost:8080 -> index.html imports vendor/wllama/esm/index.js (drive, offline; CDN only a fallback), engine pill reads drive (offline).
3. Models tab -> Scan -> reads models/models.json -> Load -> wllama streams models/gguf/Ornith-1.5-9B-Q4_K_M.gguf FROM THE FOLDER (allowOffline, parallelDownloads 3).
4. Every message you send: awakeRetrieve() reads symbols/symbol_index.json, matches your text against intents, READS the mapped vault/schema files, injects them as system context (per request, never per token) -> [TRACE:AWAKE] symbols=... files_read=... lookup_ms=...
5. Generation streams through llama.cpp WASM; tokens, first-token latency and tok/s are counted -> [TRACE:GEN] tok=... tok_s=... -> POST /api/ledger -> appended to memories/audit_ledger.jsonl.
6. Optional LAN mesh: other devices pick Wi-Fi Client, the host runs the model, tokens relay over ws :8081 (Bluetooth LE via Nordic UART also supported).
7. stack/ docker-compose is the workshop variant: code-server :8443 + runner + ollama.

WHAT IT WORKS BEST IN - measured, not vibes

Runner	Use when	Numbers
This app (wllama WASM in Chrome)	zero install, USB drive, thin clients	browser tok/s UNMEASURED - the app measures it per reply and appends to the ledger
LM Studio (desktop)	towers/mini-PCs, GPU offload	Ornith-1.5-9B measured 76.7 tok/s on a10g GPU (job 6aad81ec)
llama.cpp server / stack/	GPU tier, headless	Ornith-9B 76.7 - Ornith-35B-A3B 122.7 - Qwen3-Coder-30B 146.3 tok/s (all measured)
Ollama	quick pull on machines WITH internet	works, but first pull online - less offline-clean

RAM honesty: Ornith-9B in the browser wants ~16 GB RAM; 8 GB machines use Qwen3-0.6B in-browser, or run Ornith on the host with any of the runners above.

TRACES - how you know it is working

Every reply shows its own measurement line in the chat and writes it to the ledger:

[TRACE:AWAKE] symbols=%%EVO_OPT:DEAP files_read=2 lookup_ms=3  |  [TRACE:GEN] tok=214 first_tok_ms=890 tok_s=11.4

Read the raw ledger: memories/audit_ledger.jsonl.

HONEST LIMITS

• Browser-tier tok/s: unmeasured until tonight's first run - never quote GPU numbers for it.
• The fine-tuned CORTEX model (trace-protocol-native patches) is designed but untrained: docs/CUSTOM-MODEL-BUILD-PLAN.md.
• Two historical measurement scripts still import the banned library - owner decision pending (port or documented exemption).