Diop Daily #135 — October 2026

The Front Door Must Be Legible to the Machine

An institution used to publish for people first. Its rules lived in documents, its expertise in staff, and its public face in a website, office, catalogue, or archive. Agents change the order. A machine can now search, retrieve, compare, summarize, and act on institutional material before a person opens the page. The front door of an institution is becoming a machine-readable knowledge interface.

Google’s October 7 release of the Developer Knowledge API makes this movement unusually concrete. The company describes an official, programmatic source for its developer documentation, with semantic and keyword search, document chunking, grounded answers, metadata retrieval, citation tracking, and frequent indexing. The product claim is narrow: better access to Google’s own documentation. The architectural lesson is wider. An institution that wants agents to represent it must publish an authoritative surface designed for retrieval, verification, and correction rather than assuming that a model will infer the right meaning from a web page.

When a machine meets an institution through its documents, publication becomes part of the institution’s operating system.

A page is a room; an interface is a governed address

Human readers can tolerate ambiguity. They can follow a navigation trail, notice a date, ask a colleague, interpret a footnote, and recognize when a page has become stale. An agent needs those signals represented more explicitly. It must know which document is authoritative, what version it is reading, who owns the claim, which jurisdiction applies, what the text permits, and where a correction will appear.

This is why “make the website searchable” is an incomplete institutional strategy. Search locates words. A knowledge interface states relationships. It can identify a policy, its scope, its effective date, its exceptions, its source authority, its language variants, and the actions that may follow from it. It can distinguish a current price from an archived one, a binding rule from an example, and a public description from an internal procedure.

A useful machine-readable front door should expose at least:

  • Identity: a stable name, identifier, owner, and contact for each document, dataset, policy, product, or service.
  • Time: publication, effective, review, expiry, and last-verified dates, with a clear rule for superseded material.
  • Authority: the person, institution, mandate, contract, or source that gives the material its standing.
  • Scope: the people, places, languages, products, cases, and jurisdictions to which the statement applies.
  • Evidence: citations, calculations, attachments, provenance records, and the distinction between observed fact and institutional interpretation.
  • Correction: a visible route for amendment, appeal, withdrawal, and reconstruction of what an agent saw at an earlier time.

These fields determine whether an automated system can decide that a document belongs in a context window, a recommendation, a procurement packet, or an external action. Without them, the model may retrieve a plausible sentence while losing the institution that made the sentence meaningful.

Freshness is a responsibility, not a timestamp

Google’s API announcement emphasizes frequent indexing and low lag between upstream documentation changes and agent access. That is an important operational property, but freshness alone does not establish authority. A newly indexed error is still an error. A current document can still be written for the wrong audience, omit a local language, or preserve a policy whose legal status has changed.

NIST’s AI Agent Standards Initiative places standards, open protocols, agent identity infrastructure, and security evaluations in one public frame. Its work is still an initiative, not a completed certification scheme. The useful implication for institutional publishing is that retrieval must be joined to identity and evaluation. An agent should be able to ask not only “what does this document say?” but also “which institution published it, under what authority, for which use, and how has the system tested my interpretation?”

The initial public draft of NIST’s TEVV-Athlon Framework gives the evaluation problem a more disciplined shape. It proposes a flexible, organization-specific approach that articulates objectives, constructs assessments, applies measurement, and synthesizes results. A knowledge interface should borrow this logic. Its tests should cover retrieval accuracy, version selection, scope recognition, citation completeness, refusal when evidence is missing, and behavior when a correction arrives after an agent has already used the material.

The result is a publication system that can explain its own uncertainty. It can say that a document is current but not authoritative for this jurisdiction, authoritative but awaiting review, translated but missing a local legal term, or technically retrievable but unsuitable for an external decision. A machine that can state the limits of its context is easier to govern than one that merely returns fluent text.

African sovereignty begins at the publishing boundary

The African Development Bank’s 2026 policy brief on data governance frames the central development problem as the conversion of data into institutional value. It points to fragmentation, limited data sharing, capacity constraints, and trust deficits as barriers. The same barriers appear when agents encounter African institutions. A ministry may have a policy in French and a service description in English. A cooperative may hold authority in relationships that do not fit a single account field. A local archive may know which community can authorize reuse even when a global catalogue sees only an unstructured file.

Machine-readable publication can either preserve these categories or erase them. The choice belongs to the institution that publishes. An African university should be able to expose its research catalogue with local names, multilingual abstracts, rights, custodians, and correction routes. A public office should be able to state which language a resident may use for a consequential service. A publisher should be able to distinguish reading, translation, adaptation, excerpting, and redistribution. These are knowledge interfaces with political content.

Participation in global agent networks therefore requires more than translating an imported schema. It requires publishing the institution’s own vocabulary and making the mapping to wider standards explicit. A local term can have a global equivalent for interoperability while retaining a note that marks where the equivalence ends. That boundary is evidence of intellectual agency.

Where the investable surface is widening

The commercial object is the authoritative knowledge interface: a publishing and retrieval layer that makes institutional material fresh, attributable, scoped, multilingual, testable, and correctable for agents.

  • Institutional knowledge gateways: APIs and retrieval systems that expose official documents, records, policies, and service descriptions with stable identifiers, metadata, citations, and version control.
  • Evidence-aware publishing systems: tools that bind claims to sources, distinguish fact from interpretation, and preserve the material an agent used at the time of a decision.
  • Knowledge conformance testing: evaluation services that test retrieval, scope, freshness, language coverage, citation fidelity, refusal, and correction behavior across agents and channels.
  • Local-language adapters: maintained vocabularies, terminology packs, translation memories, and domain-specific retrieval for African languages and institutional concepts.
  • Change and recourse infrastructure: systems that notify dependent agents when a rule changes and let a person challenge, amend, withdraw, or reconstruct a published claim.

Underwriting can begin with measures that a buyer can inspect: time from policy change to agent availability; percentage of retrieved claims carrying authority and effective dates; citation completeness; stale-answer rate; language-specific retrieval quality; successful correction propagation; and the number of external decisions that can be reconstructed from the published record. These metrics describe institutional capacity more honestly than a search-volume chart.

Publish the institution’s address

A practical programme can start with one consequential domain. Choose a public service, a research archive, a commerce catalogue, or a rights repository. Assign an owner. Define the identifiers, authority fields, effective dates, languages, evidence requirements, and correction routes. Expose a machine-readable endpoint beside the human interface. Then test it with agents that ask for current information, historical information, information outside their jurisdiction, and information whose source has been withdrawn.

Make the system refuse well. If the record is stale, say so. If the translation changes a legal condition, escalate it. If two authorities disagree, show the disagreement. If a claim cannot be reconstructed, mark the gap before an automated workflow turns it into a decision. The institution earns trust through the quality of its boundaries, including the boundaries around what it cannot yet answer.

The market will keep improving models that can speak about institutions. The deeper work is helping institutions speak for themselves in a form machines can inspect without flattening their authority, history, language, or right to correct the record. The front door is already an interface. Build it as one deliberately, and the institution can enter the agent economy with its categories intact.

Sources