Milestones

We are replacing trust with deterministic proof. This vector is entirely focused on building the hardware-attested semantic grounding required to isolate and measure AI execution risk. By shipping the trust root and capturing live telemetry on-chip, we transform abstract software behaviour into a mathematically auditable ledger. The terminal state of this track is a frozen, zero-entropy interface where capital can independently verify that our agent remains strictly within its defined operational bounds.

how to read it · Read this track as one chain with no shortcuts: a verdict has to be signed before a run is worth logging, a log has to be quiet before live work is worth measuring, and live work has to be measured before determinism is worth claiming. The A and B phases are already through — the signer exists, the receipts accumulate, the false-alarm rate is sealed with a confidence interval. What is genuinely open here is not code but WITNESS: every market check on this track fails, because nobody outside the building has recomputed a receipt. That single missing artifact is why a track at ninety-something per cent on specification and execution is not finished.

↓ the whole Technical Integration list · ↓ this rung in the sequence · reading written by claude-sonnet-5, after the counts and never inside them

Technical boundaries are useless to capital until they are translated into financial constraints. This vector maps the precise surface area of our Errors and Omissions liability. We are systematically drafting a narrow-channel posture specification that defines exact exclusions, prices expected losses, and hard-caps tail risk. The output is a standardised actuarial docket — a self-contained liability asset that allows a carrier-side syndicate to confidently price and bind the risk without requiring a single manual underwriting meeting.

how to read it · This track converts a measurement into a price, and the order is the argument. You cannot price an expected loss before you have said what is excluded, and you cannot cap a tail before you have priced the body. The exclusions and the wording are written and guarded — the language fence is a red test, so overclaiming vocabulary cannot ship. The break is at the D phase: everything up to pricing exists as computed artifact, and nothing has been handed to a counterparty who could disagree with it. Arming the underwriter is the hinge, and binding the first policy is the only line here that requires another party to sign.

↓ the whole Legal Standardization list · ↓ this rung in the sequence

Technical boundaries are useless to capital until someone outside the building can read the claim and check it without asking permission. This vector maps the surface area of the argument itself — the manuscript, the corpus, and the runnable demo — as one instrument rather than three marketing assets. The book states the physics; the blog corpus repeats it in one voice, five hundred times, indexed and interlinked; the npx command lets a stranger in an empty folder recompute the claim on their own machine and get the same answer. Every buyer meets the thesis here first, and every hostile reader attacks here first. The output is not persuasion. It is a falsifiable public record: claims listed, overclaims retracted in writing, and a command that either reproduces or does not.

how to read it · Read this track as the only one whose failures are public. The manuscript is frozen and built, the errata retracts real overclaims in writing, and the cold-run guard exists because a prospect actually hit the empty-folder crash — that is the shape of an argument being tested rather than promoted. Two lines carry the honest bad news. Voice consistency has a guard and a backlog the guard does not cover: the frontier posts pass the canonical spine and the great majority of the corpus does not, because the check warns rather than blocks. And the outside world has not repeated the claim yet — the registry holds only us, and no named hostile reviewer is on record. Reach is real; corroboration is not.

↓ the whole The Argument list · ↓ this rung in the sequence

A claim set is a bet nobody has scored yet. This vector runs the prosecution of US 19/637,714 — thirty-six claims, filed 2026-04-02 on Track One, carrying the benefit of seven provisionals back to April 2025 — as a docket rather than a memory. Every application, every priority date and every statutory bar lives in one file beside the document the date was read out of, and a passed deadline with nothing recorded as filed turns the build red. The most valuable output is the negative one: a chart naming which claims the shipped code does not practise, stated plainly before a diligence lawyer or a licensee asks. Prosecution ends when the monopoly is priced per agent and someone pays it.

how to read it · Read this track deadline-first, because unlike every other list here its clock runs whether or not anyone is looking. The A phase is genuinely done — the application is filed, the priority chain is intact, the as-filed papers are on disk. Then it goes red: the formalities reply due 2026-08-16 passed with nothing recorded as filed, and the docket guard is failing ON PURPOSE to say so. Counsel last said the papers would go that week; the docket still shows no filing receipt and no signed declaration. Everything downstream — the claim chart, the defensive publications, the per-agent licence — is real work sitting behind an administrative gate that closes on 2026-12-16.

↓ the whole Patent & IP list · ↓ this rung in the sequence · reading written by claude-sonnet-5, after the counts and never inside them

Every other list here produces something that is true whether or not a stranger cooperates; this is the only one that cannot be finished from inside the building. Market contact is a ratchet with three teeth and they turn in order: an outbound approach that reaches a named counterparty, a grassroots reply nobody solicited, and — where a public claim is false and the party will not answer it — a disclosed escalation ladder that ends at an institution rather than at an invoice. The three surfaces are already public and already point at each other. /playbook is the ladder, published in full with every rung's mitigation before the first rung can fire, so the party sees the whole thing in advance. /resources is the master key that ships with the advisory invoice, so a recipient can check every implication without asking us for anything. /cta is the lever a reader reaches for after finishing the book. The scoreboard is deliberately unflattering: the win is a false claim retracted, not a payment extracted.

how to read it · Read this as the one list the repo cannot mark its own homework on, which is why almost every check here fetches production or reads a ledger instead of testing that a file exists. The A phase is genuinely done and says so from outside: all three surfaces answer a cold fetch, the ladder carries a mitigation on every rung, and the invoice hands over the pack rather than a salesperson. Then it splits in two. Outbound is real — forty-one dispatches through the pipeline, a hundred and fifty logged sends, seven distinct carrier and broker domains in the send ledger. Reply is not: six recorded clicks, and no ledger of an unsolicited inbound anywhere. Everything from the C phase down waits on one missing artifact. The register that counts asked / answered / refuted is specified on the ladder's second rung and printed there as zero — honestly, because it does not exist — and four items point at data/register/ with every one of them red. No rung has fired, and none should until the count is real.

↓ the whole Market Contact list · ↓ this rung in the sequence