Declared, Not Filled: Context You Can Audit

There is an essay making the rounds — Jessica Talisman’s Context Is a Property, Not an Object (Intentional Arrangement, July 2026) — that cuts the year’s favourite buzzword in half. The AI industry, she observes, talks about context as a container: a window to fill, a token budget to spend, a string assembled at the last moment and discarded when inference ends. Library and information science, codified in ANSI/NISO Z39.19, holds the opposite view: context is a property — declared at design time, attached to concepts and relationships, tested, displayed, maintained. Qualifiers that fix which Mercury you mean. Scope notes that travel with the term. Warrant that records on whose authority a category exists at all.

We read that essay and recognised our own design bet — stated from a tradition a century older than the problem ontoref was built to solve. Two verbs carry the whole difference: a window is filled; a property is declared. And we declare precisely to make the context visible and explicit — so nothing is completed by invention. The declared can be verified; the filled cannot. Which turns the year’s buzzword into an old, sharper question: on whose authority do we believe a result — a signature and an exit code, or fluency?

The question the window never asks

Talisman’s sharpest instrument is warrant, named by E. Wyndham Hulme in 1911: a term enters a vocabulary because evidence put it there — literature, community, operation — never because a designer preferred it. Her diagnosis of the language model follows in one line: it generates taxonomies whose authority is borrowed from fluency. A window filled at inference time carries whatever retrieval returned, and the reason those tokens existed dies with them. No record, no authority, no audit trail.

ontoref is built on the same refusal, enforced. A Live value-prop MUST cite Accepted ADRs (evidence_adrs, audited for drift). A template earns promotion to the protocol only on deposited emergence.log evidence — use warrant, operationally. And since the governed-delivery executor (ADR-066), a gate step’s status derives from the declared check’s real exit code: a self-report that contradicts the check is refused. Authority borrowed from fluency is precisely what the witness seam exists to reject.

Convergence, field for field

The uncomfortable, satisfying detail: ontoref never cited Z39.19. The glossary term schema was designed from the protocol’s own needs — and it converged on the standard almost field for field.

Two disciplines — one cataloguing meaning, one governing agent work — arriving at the same record shape is not imitation. It is the strongest kind of validation prior art can give: independent rediscovery.

From the term to the act

Here is where the twist stops being someone else’s essay and becomes ontoref’s own move. Library science declares context on terms. ontoref extends the identical move to work. A Statement of Work is one set of terms — declared before execution and owned outside the worker — plus its Work Orders, plus the signed witness each one deposits on completion. That witness is what Talisman calls operational warrant: the audit trail that answers “on whose authority does this result exist?” with a signature and an exit code instead of with fluency. Where the object approach spends its tokens and the reason they existed evaporates, the governed act leaves a record.

Her production evidence points the same way. The remedy that worked for LLM cataloguing was not a bigger window — it was a server that made the governed structure traversable at the moment of the act. ontoref’s daemon and MCP tools are that remedy, generalised from the catalogue to the architecture.

The window becomes a projection

“The window still exists in the property world, but it becomes a projection of a described state rather than a construction site.” That sentence is the reveal architecture (ADR-057) verbatim: what you see is generated from a declared source, and a drift is fixed by re-running the projection, never by hand-editing the served copy. What the agent — or the visitor — sees is never assembled improvisation. It is a declared state, projected. The window did not disappear; it stopped being where the truth is decided.

What the lineage buys — and what it does not

It buys a name and a floor. The category ontoref occupies now has a title older than the industry that needs it — context as property — and a century of bibliographic prior art underneath: Z39.19, warrant, authority control, SKOS. That floor qualifies rather than widens: it will not seduce the recipe-seeker, and it will recognise the knowledge-organization reader who has lived the property view all along.

It does not certify us finished, and honesty is part of the record too. Warrant is typed on glossary terms and value-props but still prose inside the ontology’s node descriptions. Z39.19’s reciprocity rule — every relation A→B must carry B→A, “for all types of relationships” — is stricter than our current validators. And emergence remains claim-only, a spectrum asserted but not yet measured. Prior art is a mirror, not a diploma — and each of these gaps now has a date on it.

If you would rather traverse the claim than read it argued: the graph is the declared state, browsable — every node carries its own warrant, every edge its kind. Or step through the door that fits the work you do: the developer door for governing an agent that cannot promise to follow your rules, the authoring door for keeping a body of knowledge coherent across everyone who touches it. A sister essay, The Amnesiac Scholar, records the same thesis arriving from the opposite shore — machine learning’s own admission that the agent needs a memory it can carry, governed and legible, between sessions.