The jobs that share one word
Knowledge in an AI programme is several jobs that share one word. Capture brings what exists into reach with its provenance: the wordings, the procedures, the lists, the reading of the law, each stamped with where it came from. Structure turns that material into claims and typed relations, so the clause is joined to the contract and the contract to the customer. Govern runs across the others: it scores, tiers, finds and types contradictions, applies decay, and enforces who may assert what. And serve assembles the bounded package the agent receives for this case, with the record of having assembled it.
The division that matters is what each pair produces. Capture and structure give the firm knowledge it can use. Govern and serve give it knowledge it can act on and defend. Most programmes we have seen buy the first pair, a connector and a graph, and call it done. But the check at w4 reads only what the second pair produces, so a firm with the first pair alone has a capable system it cannot put into consequential work.
The origins of what the agent knows
What the agent may rely on comes from the law, the procedure, and the tacit. The law, as the firm reads it: which articles apply to this process, what they require, what the firm has decided they mean, written by compliance and legal, dated, and owned. Then the procedure: the steps, the thresholds, who may decide what, which system holds the truth. Much of it is explicit and never written, because a specialist could say it if asked and nobody asks, or it is written where no index looks, in defect logs and post-mortems.
The third is the tacit: how a handler treats the unusual case, why the system breaks at month end, what the supervisor looks at first. It was never explicit, so it cannot be extracted. It spreads through shared practice and the traces practice leaves on the record, and it reaches an agent through those traces, through the people who take its referrals, and through the loop at w6 that reads what they decided. So the people who take the referrals are part of the second essential, and a firm that staffs the agent and not the referrals has starved the third origin.
Part three sorts the same essential by what the firm holds: the law as read, policy and procedure, the contract's terms, and the case. The origins sort it by where it comes from. The contract and the case are explicit material that arrives through the first two origins; the tacit is the one part three's kinds cannot hold, because nobody has written it.
The claim, and what it carries
A document is a container. The unit that can be checked, dated, and owned is the single assertion inside it: a claim. The slide takes one claim from the case chosen above: the clause, the list entry, the set of accounts, the notice, the flag, or the version of the guidance that the case's decision rests on, as one assertion with one date from which it holds.
The things around it are what the check needs to know. Its provenance: where it came from and when it entered, stamped at intake. Its source authority: the level of the thing that asserted it, a signed wording, a public authority's act, an issuer's notice, the firm's own record, as read by whoever may read it for the firm. Its evidential warrant: how far the evidence supports relying on it, validated, current, uncontradicted, acted on without a finding. Its validity window: the date or the hour from which it holds, and what held before. Its contradiction status: which older version or stale copy still says otherwise, and whether that is open or resolved. Its named authority: who may change it, and who may not. And its governance status: what it may be used for today, and what it may not. The take under the slide says what the case's decision is worth on this claim, and what a similarity search returns instead. A similarity search returns the most similar text, and in every case but the chain that is the older copy, while in the chain it is a figure a model derived; the check needs the claim that applies, with these things on it.
Standing has two axes
Standing is what the check reads about a claim, and it is two things kept apart. Evidential warrant is how far the evidence supports relying on the claim, assessed at curation, by rule where it can be and by a person where it cannot, and reassessed as evidence and time move. The labels the set prints, validated, current, and uncontradicted, are bands of a warrant that the reference architecture computes on several dimensions, in The Knowledge Layer. Authority is who asserted the claim, under what mandate, in what scope: a fact about provenance and permission, never a score.
The corners show why the two stay apart, each filled from the case chosen above. High on both: the basis the case's decision rests on, current and uncontradicted, and the check permits or holds as the case requires. High on authority and low on warrant: the same authority in a stale copy, the agent's index, feed, or nightly file, behind the source by hours or by a year, so currency fails and the check refers. High on warrant and low on authority: a practitioner's note borne out across forty cases, which nobody with a mandate has made a rule, so it is true and not yet usable for an act. And low on both: the agent's own earlier output read back as a source, refused and traced to its first entry. A claim from an impeccable source can be poorly evidenced, and a well-corroborated claim from a junior source is no less true. So the firm keeps the axes apart, sees how a claim is weak, and fixes the right thing.
The second rule: standing scales with consequence
The set has one hard rule on the agent: autonomy may not rise above evidence. An agent may be operated at a level only when the evidence for that level is met for that agent in that deployment, and the rule widens an agent over quarters, at the loop and the investment decision, and narrows it at once on a rule written in advance; the board, the CRO, internal audit, and the supervisor read it. It needs a second rule on the basis, at the same rank. Standing scales with consequence: an act of a given consequence tier may commit only on claims whose standing meets the bar set for that tier, under the authority the tier requires. This one is decided per act, at the check, against the ceilings and the control standard, and read by the engine that executes the act and then the record.
The check reads the weakest premise the act rests on, never an average, and refuses or refers when the act would exceed the authority of the claims it rests on, when the standing of those claims is below the bar for the tier, or when an unresolved contradiction sits among them. The reads at the foot of the slide come from the case chosen above: first the act on the basis that applies, current and uncontradicted, where the bar is met and the check permits; then a second act of the same kind, later that day or in the days after, on the stale copy the agent declared, where currency or supersession fails on the weakest premise and the check refers, so the money, the election, the message, or the notice waits for a person. Rule one sets how far the agent may go this quarter. Rule two decides whether this act may commit now. A firm needs both, and none we found sells either.
The tier itself comes from outside the knowledge layer. The paper requires the mapping from an act to a tier, from what the act is, its money, its effect on a customer or a third party, its reversibility, the border it crosses, and what it counts toward, and it refuses to own it. The agent that proposes an act must not set the class it is gated by, and the gate applies a classification the firm governs apart from it. So the firm declares the consequence table at the control standard, beside the ceilings, under a named authority and in a version, and the knowledge layer applies it. Part four gives the table for the case chosen here, the tiers with the bar each sets on both axes, and the record line carries the tier and the version that classified the act.
How knowledge is prepared
What the agent relies on decays at the speeds the papers name, and the set adds one of its own. Domain mechanics, how a claim is adjudicated or how settlement works, decay in years, with a process redesign or a migration. Regulatory claims, the firm's reading of the law, decay in months, with a regulatory change or an enforcement action. Operational reality, the procedures and thresholds and client configurations, decays in weeks, with a reorganisation, a staffing change, or a leadership change. And live lists and prices, the set's own class, decay in hours, which is why the payments case turns on a copy refreshed at 05:30.
Someone has to keep all of it fit to act on. That is a function of the firm, the knowledge function, running the stages with the source owners, compliance for the reading of the law, and the platform team for p2: ingest, consolidate, curate, expand, apply. Curation is the continuous stage, the quality engine where quality is assessed and decay is applied, and it costs in the order of half to one person per domain once it runs, a domain being one product line's wordings, one list service, or one legal register, after a first load the pilot measures; that is the set's own illustration to budget against, and no benchmark. Rules keep the estate fit. Contradictions are kept and shown, typed and routed to the authority that can resolve them, never settled by deleting the inconvenient claim. A claim leaves by governed disposition, demoted, archived, or sealed with its trail, never silently gone. And an agent's own output never re-enters as an independent source: it comes back labelled and provisional, because corroboration needs an independent origin and ten copies of one claim are one claim. The second essential at the enterprise altitude, e2, is this function's charter, and most firms we have seen have not written it.
The ways the basis fails
Governed Intelligence Architecture names the ways a knowledge estate fails. The defences are the lifecycle's, and the slide before states them as stages and rules; this slide gives each its name. Pollution: a claim that was never true enters, an extraction read as fact or the agent's own letter read back as a source; the perimeter at ingest holds it provisional, a provisional premise is below the bar, and the check refers. Staleness: true once, superseded since, wording v13 after the renewal onto v14; the validity window is the defence, currency fails on the weakest premise, and the check refers. Fragmentation: two words for one thing, excess here and deductible in the acquired portfolio, and the clause is never joined to the contract; drift detection on the estate is the defence, and the check sees only the effect, a package complete and uncontradicted and a settlement without the excess. Amnesia: superseded without the chain kept, a migration that carries every claim and drops the links between versions, and nobody can say what applied to a contract on 4 September; governed disposition is the defence, and the control plane sees only the effect, a record that names a version the estate can no longer explain. Cascade: one bad claim and the ten that depend on it, each healthy on its own; dependency-aware warrant is the defence, and the check sees it at the act where warrant follows the dependency to the origin. Structural distortion: a hub everything touches, and the package for one act pulls in what does not belong; topology is the defence, inside the knowledge plane, and the check sees only a declared basis whose window closed.
Three of them reach the check at the act, on a field of the claim. The other three show only in their effect: a settlement without the excess, a version the estate cannot explain, a package too wide. Those live inside the knowledge plane, where the control plane sees only what they did, in the part of the second essential the check never sees, which the knowledge function keeps fit with the source owners and the platform team. Every dossier runs staleness; situation cards 8 to 12 run the other five.
What you govern when you govern claims
One test tells a firm which side of this chapter it stands on. For any act of the last quarter, name the claims it relied on and their standing at the moment it committed. A firm that passes has knowledge in the full sense: claims with provenance, standing, validity, and an owner. On that content sits the firm's intelligence, the handlers and the agents and the loop reasoning over it, seeing what matters in a case and choosing an act; a property of the firm and never of a model. And governed intelligence is what the firm can defend: capable agents acting on governed knowledge, with every consequential act admissible at the moment it commits, which is what the supervisor, the customer, and the board see.
Then the empty column of 2026. Management planes inventory the agents. Control planes, as sold, gate the tool call. Knowledge planes hold what the agent knows. But no product we found reads the standing of the basis at the moment of commit, which is the input of w4 that joins the second essential to the fourth essential. The set fixes that condition. Producing standing a firm can trust is the harder half, and a research question the Hub answers by implementation, at clients in different places, as many as it can run. A firm that governs its documents knows where its documents are. A firm that governs its claims governs its own intelligence, and can show, for any act, what it rested on. That joint is the set's argument, and the thinking behind it is on the last page of this site.