Document type: Use Case (worked companion to the comparison investigation v0.1) Version: v0.1 Date: 2026-06-13 Operator: Marvin Percival
What this document does. It works a single difficult knowledge-management problem all the way through, first as the Open Knowledge Format alone handles it, then as Loomworks handles it, so the boundary between the two is concrete rather than abstract. It is the worked companion to the comparison investigation v0.1.
The scenario. A multinational bank runs a fleet of autonomous agents on a knowledge catalogue exported in the Open Knowledge Format. A regulator restates the definition of a core metric, retroactively. The bank must find everything that rested on the old definition, correct it without erasing the old state, keep the whole trajectory auditable, distinguish machine-inferred catalogue entries from human-attested ones, and let nothing settle into what the agents act on until a responsible human approves it.
What it shows. The Open Knowledge Format delivered the knowledge to the agents cheaply and portably, and Loomworks does not replace that. But every demand the restatement makes — recall across a moving body of knowledge, corrections preserved as corrections, tie-to-source trust, a human approval gate, a tamper-evident audit trail — is work the format declines by design. Those are exactly the demands that make this problem hard, and the layer that meets them is Loomworks.
Scope. Illustrative. Fictional institution, fictional metric. No code, no specification change, no seed amendment proposed.
Meridian Bank, an agent fleet, and a catalogue that grew the easy way.
Meridian is a fictional multinational bank. It runs autonomous agents across treasury reporting, credit-risk monitoring, regulatory filing, and customer advisory. Each agent needs the same internal context an analyst would need: what a table means, how two systems join, what a given metric counts, which definition applies in which jurisdiction.
To stop every agent from reassembling that context from scattered systems, Meridian adopted the Open Knowledge Format. Every team now exports its knowledge as a bundle of Markdown concept files. The data platform team runs the reference enrichment agent over its warehouse: it walks every table and view, drafts a concept document for each, then makes a second pass with a language model that crawls internal documentation and fills in descriptions, schemas, and join paths. The finance data office exports its catalogue from a third-party governance tool into the same shape. The treasury desk hand-authors a small bundle of its own. Different producers, one common shape, no translation layer between them. For the first year this is a clear win — the agents are better grounded, and the knowledge lives in version control like code.
A representative concept file looks like this:
---
type: Metric
title: Net Revenue
description: Revenue net of returns and rebates, recognised at settlement.
resource: https://catalog.meridian.example/metrics/net_revenue
tags: [finance, reporting]
timestamp: 2026-02-11T09:00:00Z
---
# Definition
Computed from [orders](/tables/orders.md) joined to [returns](/tables/returns.md).
Used by the quarterly filing agent and the treasury dashboard.
Nothing here is a defect. Each is the format being exactly as minimal as it promised to be.
Three things are true of Meridian's catalogue, and all three are by design:
type field and reconciles nothing, so the three files coexist and different agents read different ones. The divergence is silent.None of this hurts while the knowledge is stable. It all becomes load-bearing the moment the knowledge has to change under consequence.
A definition changes, retroactively, and the consequence is regulatory.
> The regulator restates how Net Revenue must be recognised: not at settlement but at the point of performance, applied retroactively to the prior four reporting periods. Filings already made under the old definition stand on the record. New filings must use the new definition. And the bank must be able to show, on demand, what it reported under which definition and why.
This is the difficult knowledge-management problem in full. It is not a data-entry change. It asks five things at once:
It can carry the new file. It cannot carry the demands around it.
What it can do
What it cannot do, by design
The format did its job — it got knowledge to the agents and it will carry the corrected file onward. What it has no answer for is everything the restatement actually demands. That is not a gap to file against the specification; it is the deliberate edge of a format that requires one field and governs nothing.
The catalogue enters as engagements, the restatement moves through the four rooms, and it leaves again in the shape the fleet already speaks.
Meridian wraps each producer's bundle as an engagement and points the corrected catalogue back at the same fleet. The Open Knowledge Format stays the format at both the entrance and the exit; Loomworks is the governed interior between them.
Room one — Memory · ingestion as a contribution pathway. Each bundle enters as an operational engagement. Every concept becomes assertions in Memory, and each assertion is threaded to its origin on the way in: this Net Revenue definition came from the finance office export, attested by a named office on a date; that join path came from the enrichment agent's language-model pass, marked machine-origin in a way that cannot be suppressed; the treasury note came from a named person. The format's thin timestamp becomes a real tie-to-source. The three colliding Net Revenue definitions do not merge into whichever file was read last — they enter as three distinct assertions with three distinct provenances and three jurisdiction scopes, so the conflict becomes visible instead of silent.
Room one — Memory · the trajectory is held. When the restatement lands, the new definition enters as a contribution that supersedes the settlement-timing definition. The old assertion stays in Memory as a correction, with its provenance and the reasoning intact. The regulator's question — what did you believe in the second quarter, and on what basis — has an answer, because the prior belief was never overwritten. Corrections are preserved, not smoothed.
Room two — Manifestation · recall does the hard part. Manifestation organises Memory at this moment into the current picture of the Net Revenue concept across jurisdictions. Recall is the work that matters here: because each downstream metric, filing, and advisory statement was threaded to the assertion it derived from, changing that one assertion surfaces the whole dependent set. This is noticing-when-a-premise-changed made operational — not a manual hunt through a year of files, but the change propagating visibly to everything that leaned on it. The machine-origin marking carried from ingestion lets the reconciliation flag the enrichment-derived items for closer human review.
Room three — Shaping · arranged for the reader who will act. Shaping arranges the Manifestation-organised knowledge for a specific reader. The regulator submission gets the full trajectory, every provenance thread, and the attestation record. The controller's review surface gets the conflict laid out and the decision to be made. The agent fleet gets only the settled definition and the scope it applies in — nothing of the deliberation behind it. Same knowledge, three arrangements, no re-derivation.
Room four — Rendering · both modes, and the format as the exit. In Mode B, Loomworks emits a governed bundle back to the fleet in the Open Knowledge Format — the same shape the agents already consume, but now every concept is the approved one, carrying a pointer back to the Loomworks provenance and attestation. The fleet keeps speaking the format; the trajectory and the governance stay on the Loomworks side. In Mode A, Loomworks produces the regulator-facing reconciliation document directly.
The gate, and the anchor. Two commitments run underneath all four rooms. The new definition does not settle into what the fleet acts on until the responsible Operator — here, a regulatory controller — approves it, per jurisdiction. The machine surfaces the new definition, the full affected set, the provenance of each item, and the variants; the Operator approves. There is no automatic state transition on knowledge a human is accountable for. And the attestation that this definition was approved by this controller, on this date, for this jurisdiction is anchored through FORAY, giving the tamper-evident audit trail the regulator demands.
The format is the boundary on both ends. Loomworks is the governed interior.
Read against the comparison investigation, the scenario lands its point. The Open Knowledge Format carried Meridian's knowledge to its agents cheaply, portably, and across many producers — real value that Loomworks consumes and emits rather than replaces. But the moment the knowledge had to change under consequence, the format's deliberate minimalism became the whole problem: it cannot tell you what rested on a changed premise, cannot hold the old belief as a correction, cannot distinguish machine origin from human attestation, cannot anchor an audit trail, and cannot withhold settlement until a human approves. Those five absences are not weaknesses to be patched. They are the edge of a format, and the territory past that edge is precisely what makes regulated knowledge management hard.
Loomworks lives in that territory. It takes the format at the door, threads provenance onto it, holds the trajectory, surfaces what a change touches, governs what settles, anchors the attestation, and hands the result back out in the same portable shape. The agents never stop speaking the Open Knowledge Format. Everything that makes the restatement survivable happens inside the layer the format leaves open.