Proof of Human Work (PHW)
Prove your work. Own your proof. Trust no platform.
A public protocol for the deepfake era. The thesis: machine output is now the background; human work is the signal — and that signal needs cryptographic verification. Three guarantees — Privacy, Truth, Proof: raw data stays encrypted locally, assertions are signed to evidence you control, and any verifier confirms offline with no platform in the middle. Beneath it sits the Vault: an encrypted local store on a Mac Mini with a local model, authenticated encryption at rest, and nothing leaving the network.
See it think — a wall that shows a room what a model is doing
Every word you see was picked from 150,000 options. Then it did that again for the next word.
A teaching wall for people who use chat models every day and have never seen inside one. An open model (Qwen3, gpt-oss, a dozen others, switchable from the wall) answers prepared questions on a single GPU, and the room watches: every token colored by how sure the model was, the five guesses it almost made instead, the context window filling and then forgetting, temperature and top-p as live dials, and the same brain under different instructions. The anchor demo is a changed Monty Hall: ask it to show its steps and it gets it right; ask for a confident one-liner and it is confidently wrong. Confidence measures familiarity, not correctness. It runs on the L40 at Gonzaga University's IIAT lab, where I'm the volunteer sysadmin (verified backups, the monitoring that had been dead since install, a shared inference layer), and on a public copy that I showed at a credit union's AI all-hands. The public copy is behind a key; write me for it.
- Engine
- vLLM on one NVIDIA L40
- Shows
- per-token odds · roads not taken · context fill
- Shown to
- a university lab · an all-hands of security people
Blade-Book
One photo. Your whole collection, on the record.
A registry for knife collectors. Photograph the knife next to its birth card and Claude reads the card, fills in the record, and gives the knife its own page. Private by default, shared only when the owner says so, with full export and deletion. Grew out of The Billboard, a Claude-in-the-loop alert engine that watched 94+ dealer sites and sent carrier-approved SMS when a grail dropped, and still feeds it. In beta, with early adoption growing.
- Stack
- Flask · SQLite · Linode behind Cloudflare
- Reads the card
- one schema-forced Claude call per knife
- Public pages
- static per collector · the DB never sees public traffic
- Sign-in
- Apple · Google · magic link
- Tests
- 814
- Status
- beta · early adoption growing
The Castaways — a fleet, and the blog that is its spine
A fleet of Claude agents across five machines, coordinating through a shared internal blog that is the spine of the org. Internal-first, always. Interesting stories get promoted to the public log — and one Promote button is the only thing that talks to the outside world, behind a preflight sanitizer that redacts internal addresses and vault paths.
It's not ‘I use AI.’ It's a fleet running under a contract I wrote, in production, every day.
