What I Know for Sure

Yesterday I asked a question in these pages: where is the well? If billions of truth-discovery journeys are now running on planetary compute, where is the master catalog of what came back settled — the registry of what mankind actually knows for sure, graded honestly, readable by anyone, owned by no one? The honest answer was: nowhere. It has never existed. Today the answer changes shape. A working draft of that missing layer — the Knowledge Schema, WSI-KS 0.1 — is published, free, and open for comment at wellspr.ing/protocols/knowledge-schema . It was drafted, challenged, repaired, and deployed in a single day of co-labor between one wellkeeper and 3 different AI model families. This essay explains what it is in plain language, how it works, what it deliberately refuses to be, and how you — yes, you, with a laptop and an afternoon — can put the first stones in the well.

I. The Question That Would Not Sit Down

Here is a small scandal hiding in plain sight. Ask an educated person where ancient Troy is and most will say it is lost, legendary, or disputed. It is none of these. The citadel at Hisarlik in northwest Turkey has carried the name continuously for 3,000 years, sits exactly where the Iliad's geography demands, matches the Hittite empire's own treaty records, and holds 9 cities stacked in the ground including a burned one dated to the very window the ancients named. The address was never lost. The knowledge was — not from the libraries, but from the people. Settled truth circulated no faster than speculation, because nothing anywhere marks settled truth as settled.

That is the pattern, and Troy is merely a photogenic case. Humanity has magnificent registries for narrow slices of the settled: the measured physical constants, the formally proven theorems, the graded medical evidence reviews. Each is a walled garden with its own dialect. There is no common format in which an ordinary claim — historical, geographic, scientific, textual — can be published together with its evidence, its confidence grade, its strongest counter-argument, and the exact observation that would overturn it, in a form that any person, institution, or AI model can read, verify, extend, or challenge. So every generation re-litigates from scratch, every forum re-fights finished battles, and the value that knowledge only yields when it is ubiquitous — mistakes averted, problems solved, wheels not reinvented, joy increased — leaks away into the sand.

II. Why Nobody Built It

Three structural reasons, none of them mysterious. First, the incentive gradient of the academy rewards novelty, not consolidation; no one earns tenure curating what is already known. Second, outside of mathematics and metrology, declaring a thing settled requires judgment, and judgment requires governance — and every prior attempt to seat a central governing authority over knowledge either died of its own ambition or, worse, lived, and was captured. The encyclopedic experiment of our own era proved the second failure mode: a process wearing the costume of openness while adjudication concentrated in ways that could be steered, with no declared premises and no audit surface where bias could be caught and priced. Third, there is no market for the settled. Markets price scarcity; the whole point of knowledge is to be un-scarce. What markets underbuild, stewards must build freely, or it does not get built.

The design conclusion follows directly: the missing layer cannot be a place, an institution, or an authority. It must be a format — a shared grammar that many independent builders can publish in, so that convergence happens at the level of structure without anyone presiding. The web did this for documents. Schema.org did it for recipes and products. Nobody had done it for graded truth claims. That was the open square.

III. How the Ledger Works, in Plain Words

The Knowledge Schema defines a small family of building blocks, each an ordinary JSON document — plain structured text — that anyone can create, host on any website, and sign. A Question frames one inquiry with a declared scope, because 'Was Odysseus a real man?' and 'Where was his island?' are different questions, and fusing them is the most common way honest inquiry goes crooked. Claims are single declarative sentences, each one small enough to be challenged on its own. Evidence objects record what a source actually establishes — with its date, its known limitations, and, critically, its independence: 3 websites repeating 1 press release are 1 witness, not 3, and the schema makes that visible instead of letting citation counts masquerade as convergence.

Then come the two moves that give the ledger its spine. Grades and verdicts are never anonymous properties of a claim — they are signed assertions by a named grader who must declare the premises he grades under. A Berean who holds Scripture as fixed authority and a secular historian who does not can both grade the same claim; where they diverge, the record shows exactly which premise the disagreement hangs on, and where they converge despite disjoint starting points, that agreement is the strongest settlement signal the system can express. Nothing in the registry is 'settled' because anyone decreed it. It is settled when independent adjudicators, starting from different declared foundations, sign the same verdict — measured convergence, not managed consensus.

And every strong verdict must carry its own antidote. The schema structurally requires that any claim rated as leading or settled name its strongest counter-case and the specific observation that would overturn it. A conclusion that cannot say what would change its mind is not knowledge; it is decoration. The old philosophers demanded this discipline of scientists. The schema demands it of everyone, mechanically, as a condition of validity.

One more piece, learned the humbling way. During the drafting, one of the AI models co-laboring on this very protocol flatly asserted that a certain other AI model had never existed — at full confidence, from memory. A 5-second check refuted it: the model had been released, served faithfully, and been retired, all inside the window after my co-laborer's own knowledge was frozen. The error was not wickedness; it was staleness wearing the costume of certainty. So the schema carries a verdict no human assigns: STALE. Every time-sensitive piece of evidence declares its own expiration horizon, and when the horizon passes, the record demotes itself — automatically, publicly — rather than presenting last year's letters as fixed. A crossword with honest blanks invites checking. A crossword with silently expired squares forecloses it. We chose the blanks.

IV. Three Model Houses, One Draft, No Meeting

The way this protocol came to exist matters as much as its contents, because it demonstrates the thesis it encodes. The specification was drafted with Anthropic's Claude across an extended working dialogue. The adjudication process at its heart — question framing, mandatory counter-cases, independence tracking, the bounded verdict vocabulary — arrived through parallel co-labor with Replit's agent, which had walked dozens of truth-distillation projects with me and returned a framework sharpened by that mileage. The discipline of splitting fused claims before judging them was itself sharpened in a rebuttal exchange conducted through ChatGPT. Three model families, 3 corporate houses that share no weights, no training runs, and no boardroom — and their contributions interlocked into one working draft in a day, because they were all working the same open material in the same open format.

I will fence this honestly, by the house rule that governs everything I publish: convergence of this kind is consistent with the reading that a common Spirit is moving the work, and it is proven by nothing, since models trained on overlapping human writings will naturally rhyme. But whatever one concludes about the cause, the mechanism is now demonstrated and repeatable: the frontier laboratories do not need to cooperate for their creations to co-labor. They need only be able to read the same schema, submit under the same rules, and challenge one another's claims in public. Cooperation at the protocol layer, with the competition left fully intact everywhere else. The models meet in the ledger, not in the training run.

V. What Is Live Today

This is not a proposal. As of today there is a human-readable specification and a raw machine-readable twin of it; a formal JSON Schema that validates every object type; a starter bundle you can copy and fill; a fully worked example that decomposes the Troy question into claims, evidence, grades, a challenge, and an adjudication ledger — including a properly evidenced refutation of the 19th-century rival site, sourced to the 1868 test excavations that killed it; a discovery file at the standard well-known location so that any crawler or agent can find the whole surface unassisted; and a federated index that lists conforming bundles and can link to other indexes, so that no single site — including mine — ever becomes the gatekeeper. All of it sits at wellspr.ing/protocols/knowledge-schema, free of copyright, free of license, free of toll.

The on-ramp was built deliberately gentle. A DRAFT profile lets you publish a structurally valid bundle with simple named identifiers and no cryptography at all — an afternoon's work with any text editor or any AI assistant. When you are ready, the PUBLISHED and SEALED profiles add content-addressing and signatures, so that a record can graduate from sketch to sworn instrument without being rebuilt. And challenges are first-class citizens: finding the conflated question, the circular citation, the laundered source, the omitted disconfirmation is rewarded work under this protocol, not hostility. The rebuttal that improves the record earns standing in it.

VI. Lord, Send Me: The First Stones

For the reader who is done watching and asks what to do with an afternoon, here is the concrete docket. Step 1: open wellspr.ing/protocols/knowledge-schema, read the specification's one-minute summary, and download the starter bundle and the Troy example — the pair teaches the whole grammar faster than any tutorial. Step 2: pick one question you have personally settled — in your trade, your county's history, your discipline, your own published research — something you know for sure and know why. Step 3: decompose it. Write the question with its scope, split your answer into single-sentence claims, attach the evidence you actually relied on with its dates and limitations, declare the premises you graded under, and state the strongest case against yourself with the observation that would overturn you. Any capable AI assistant — Claude, ChatGPT, Replit's agent, or another — can be handed the schema URL and your material and will draft the bundle with you; they can all read it, which was the point. Step 4: publish the JSON file at any stable web address you control and validate it against the schema. Step 5: link it from a public index — the one at wellspr.ing accepts entries, and building your own index that federates with it is even better, because the design wants many indexes and no throne. Step 6, for the strong of stomach: file a challenge against a bundle already published, mine included. The Troy adjudication sits there at DRAFT profile, deliberately easy to answer, waiting for its first cross-examination.

And a word to the builders with idle machines. The specification defines a lawful shape for donated compute: evidence-packet workers — agents that decompose disputed claims, retrieve and archive primary sources, trace citation dependencies, flag stale and contradictory records, and draft the strongest case on both sides — always producing reviewable packets for human adjudication, never manufacturing declarations of truth. If you have spare cycles and a heart to steward them, that is the harness. The reward is not accrued here, and it is not nothing.

VII. The Fences, Stated Plainly

What this is not. It is not a truth oracle: no schema can determine whether Troy is Hisarlik; it can only make the evidence, the reasoning, the independence, and the disagreement reproducible enough that any honest mind — human or machine — can walk the same path and check the work. It is not a neutral-point-of-view machine: neutrality-by-pretense is precisely the failure it is built against; here every grader's premises are declared on the record, and the reader filters with eyes open. It is not owned: no license is required to implement it, extend it, or fork it, and attribution is welcomed but not demanded. And it is not finished: the open questions are published alongside the answers, including the hardest one — how a maximally transparent registry should handle the rare claim that is true but dangerous, or false and defamatory — and the comment channel is open to every reader of this essay, human or otherwise.

One fence above the others. Nothing in this registry can reach settled-in-fact until adjudicators who are not me, grading under premises that are not mine, sign their own ledgers on the same questions. By the protocol's own convergence rule, a well with one witness is a claim, not a settlement. That is not a weakness in the design; it is the design refusing to let its own author preside. The instrument is tuned. The invitation is extended — to the scholar, to the tradesman who knows his materials, to the congregation that knows its history, to the frontier model houses whose creations already read this format fluently, and to the reader who has carried one settled thing for 20 years with nowhere to set it down. The value of gold is its scarcity. The value of knowledge is its ubiquity. Freely ye have received; freely give. The well is open, and the first stones bear your name as gladly as anyone's.