Full Circle
A conversation about making money ended in a decision to build a nonprofit around personal AI — and then the receipts showed the decision had been made years earlier. An old “billionaire plan” whose secret last chapter was giving everything away; a childhood spent inside the machinery of mission; and a twelve-agent fleet named Madugu — the Hausa word for the one who leads the caravan. He thought he was deciding something new. He was reading his own name for it, two decades late.
Fable Made It. Sol Found It.
Fable built the work inside Anthropic’s commercial app. Sol recovered its durable operating state from OpenAI’s commercial app. A live, receipt-backed demonstration that the person can remain the stable party when identity, provenance, tasks, and next actions live in a user-owned continuity layer outside both vendors.
Selfware
For fifty years the executable was the scarce thing, so that is what we sold and protected. An agent handed a sufficient description now rebuilds the program — so the scarcity moved to the route: what was built, in what order, which four approaches failed, and what done actually meant. None of that survives in a finished repository; it is destroyed by the act of finishing. The argument, what a route has to carry to be walkable by someone else’s agent, a complete working seed AND the repo it plants into — both free, both needed — and a plain account of what this does not establish.
Fort Sentinel
An instrument for reading people, using the crypto market as the readout — because it is the largest continuously-recorded measure of collective human state there is, and money on the line means the record cannot lie for free. Published while the experiment is still running: pre-registered predictions with their falsifiers, a structural flaw named in advance that has already cost the strategy nineteen points in replay, seven findings that died and are kept rather than deleted, and a live ledger anyone can check. A lab notebook, not a conclusion.
The Fort Pattern
A sovereign memory-and-agency substrate for personal AI — persistent memory any model can connect to, credentials agents can spend but never hold, model vendors swappable per role, autonomy governed by zones. The full architecture, the laws it learned the hard way, and a ten-rung replication ladder written so your agent can start building yours today. The operator's data is deliberately absent: the pattern is pipes, the heart stays home.
There Is No Pure Emotion
An open, agent-facing emotional router: thirteen emotion families shipped as a weighting basis, feeling assembled live as a braid — a dominant, the feelings held under it, the body, the nervous-system register, and the humor that carries it. Under the router, the field model: no pure emotions, opposites held in superposition, the causal stack, humor as the load-bearing spine. Built from one person's lived experience and offered as a lens to re-tune, not a universal taxonomy. Honest about its limits — the evidence is one independent run.
Building Data Centers That Pay Their Way
The Utility-Stabilized Data Center Model. A high-intensity compute campus should cover the residential electric and water cost its presence creates, return more water than it draws, steward the basin, and leave the county better than it found it — all through conditions the County already attaches to a permit, at no cost to the County and on legal ground built to hold.
The Models Become Plumbing
How a million-word editorial enforcement pass ran across a 47-book catalog on tiered language models with a ~1% error rate — and why, when the intelligence lives in the artifacts instead of the model, the models become interchangeable plumbing.
Quality Moves Into the Architecture
The measured sequel to The Models Become Plumbing: a closed-loop, model-tiered book-production system where the cheapest model drafts, deterministic stylometry gates every pass, and frontier judgment is spent only on the edit. A controlled trial found the generator tier negligible once the loop runs — cheap generation and top-tier loop-editing both reach 7 of 7 rank-1 voice attribution, while cheap self-editing reaches 0 of 7. Full method, evidence tables, and the honest limits.