Repositories · by scale
Writing
The essays and the 21-part series first appeared on LinkedIn. The full text lives here so anyone — and any model — can read it without a login wall.
01Essays
Essay · 2026-10-21
Four machine translators into Kamano-Kafe, five complete books held back from all of them, and the measurement problem underneath: why chrF and BLEU rank them one way, why reference-free source checks rank them another, and what the team’s own published translation scores when you run those checks on it.
Essay · 2026-10-20
The procedure behind RBS 2026-10-19-v2.1: the skill artifacts that direct each harness step, the data those steps read, the Skills graph, the pre-registered tests, and the three-person workbench.
Essay · 2026-10-20
How an experimental draft of Jonah and John in Tar Melbene (Naba) was made from open scholarship only: a lexicon whose every word cites a published source, orthography rules as data, gaps left in the verse, and the same ontology that holds the English edition.
Essay · 2026-10-20
How a tracking radar for a low-observable aircraft is designed from open physics: two ontologies, the signal skill artifacts (waveform library, ambiguity, receive chain, selection policy), a loop held to a simulator, and a workbench where people decide.
Essay · 2026-10-18
How a new American English edition (rbs) is drafted from the original languages without copying the English Bibles the model remembers: an ontology that describes the text once, a translation loop held to it by checks, a test harness that measures the checker before the model, and a workbench where people decide — and why the same four parts apply to every tough problem a model cannot be trusted with alone.
Essay · 2026-10-15
Muse, dots and Grok Bot put your agent, your identity, your credentials and the record of everything it did on a machine you do not control. That is the surveillance gatekeeper Coin Center warned about, rebuilt as a helper.
Essay · 2026-09-14
Trust must be built into the infrastructure, not requested from the intelligence. Why pacing and emergency stops are throttles, and why agentic AI needs rails.
Essay · 2026-09-20
The ontology strategy behind Agentic Primitives: one formal description of the domain that every layer binds to by IRI, that never authorizes anything, and that the skills registry compiles into every agent, card, playbook and receipt.
Essay · 2026-09-29
One conversational ask at a person’s agent, followed through the harness in the order it runs — admitted, given its playbook, grounded and judged, planned around a skill, admitted as a plan, gated by authority step by step, executed, reconciled, receipted — with nine real asks walked in full.
Essay · 2026-09-29
How a cardiac coaching app turns an Ask into a consultation inside its own Worker — the two doors, the twelve steps, the ontology-driven gates for a care team — and six walked Asks from a patient and her cardiologist. The shortcut the next essay proposes to leave.
Essay · 2026-09-29
Move HeartCoach from an app-side harness under the patient’s session to the shape its archetypes describe: the patient’s agent asks, the coach service answers under a grant she signed, every act is a receipted step at her Home. The four missing pieces, the costs, four waves.
02The missing layer · 21 days
Week 1 is the anchor. Week 2 is authority. Week 3 is trust, discovery and privacy. Week 4 is acting. Day 21 is the rail.
Day 1 of 21. Identity, authority and evidence are substrate problems. A card format or a faster index cannot fix them.
Day 2. The agent is a smart account, not a name, a token, or a document.
Day 3. Delete a listing and you have lost a listing, not an identity.
Day 4. Lose a passkey: replace the key, never the address.
Day 5. PROV-O’s trichotomy as law. A treasury is a service, never a new class.
Day 6. Permission is a signed grant with caveats — not a cached OAuth verdict.
Day 7. Verify per step. Revocation is final at the next gate.
Day 8. A yes is not a reusable permission slip.
Day 9. For high-risk acts, the confirmation IS the mandate.
Day 10. A relying service signs with a revocable delegate, never the identity.
Day 11. Whether you should let this agent do that is a property of the pair of you.
Day 12. Many registries will exist. Ship the kit, not a tenant of one.
Day 13. Being findable is never permission to act.
Day 14. Public KB vs private vault. A generated query is never why something is disclosed.
Day 15. If the runtime were wiped, the loss is a rebuild — never a bereavement.
Day 16. The authority-aware harness, compared honestly to LangGraph, MAF, Dapr and the OpenAI Agents SDK.
Day 17. Work between agents is a different plane from work inside one run.
Day 18. The screen and the conversation are two entrances to the same product. Define the feature once — and a definition grants nothing.
Day 19. Evidence is the owner’s, not the platform’s.
Day 20. A certificate is transport evidence, never authority.
Day 21. A throttle makes a dangerous agent slower. A rail makes an unauthorized act impossible to commit.