Diop Daily #095 — August 2026

The Model Catalogue Is Part of the Institution

An institution that cannot explain why it chose one model over another cannot fully explain its own machine decisions. The choice may appear in a developer setting, a vendor dashboard, or a routing rule, yet the consequences reach the budget, the data boundary, the quality standard, the language of service, and the person accountable for the result. Model selection has become a procurement question with an execution trail.

Recent public signals make the shift measurable. OpenAI's August 24 RSS item on GPT-5.6 in Kiro describes better price-performance for software planning, building, review, and testing. Its August 13 builder's guide describes smarter model selection alongside new Responses API capabilities. A separate August 13 announcement previews an Ultrafast service tier for GPT-5.6 Sol, with up to fourteen times the speed and up to 750 output tokens per second. Google's direct description of ADK Go 2.0 supplies the runtime vocabulary around graph workflows, dynamic routing, human checkpoints, and resilience. The European Commission's General-Purpose AI Code of Practice supplies a regulatory reference for provider obligations under the AI Act.

A model catalogue is the institution's memory of what each machine may be trusted to do, under which conditions, and at what cost.

Model choice is a purchasing decision

A model is often introduced as if it were a replaceable component. The interface remains stable, the prompt remains similar, and the workflow appears to continue. The institution experiences a different reality. A model change can alter the evidence it cites, the languages it handles, the refusals it produces, the data it retains, the latency it imposes, the review burden it creates, and the commercial promises it can safely support.

This is why a model catalogue belongs beside procurement records and operating policies. It should answer the questions that a serious buyer asks before approving a supplier:

  • Capability: which tasks, languages, formats, tools, and reasoning demands have been tested.
  • Evidence: which evaluations, source sets, human reviews, and failure cases support the capability claim.
  • Cost: the full price of inference, storage, retrieval, integration, review, correction, and local connectivity.
  • Boundary: where data is processed, how long it is retained, which keys control access, and which jurisdiction governs the route.
  • Behaviour: how the model handles ambiguity, refusal, uncertainty, sensitive content, and requests outside its mandate.
  • Substitution: which alternative model can take the work, what evidence must be rerun, and which standards may change.
  • Owner: the person or institution responsible for approving the model, monitoring drift, and retiring the route.

The catalogue is useful because it preserves relationships among these fields. A fast model may require more human review. A cheap model may perform well in English and fail on a local language. A powerful model may process sensitive material under terms that the institution cannot accept. A local model may cost more per call while reducing cross-border transfer, translation repair, or connectivity dependence. Procurement becomes rational when the institution can compare the entire route.

The substitution problem is a continuity problem

Model providers change prices, service tiers, context limits, safety behaviour, and infrastructure partners. Institutions also change their own routes as budgets, laws, and data classifications change. Substitution is therefore part of ordinary operations. The question is whether a replacement preserves the standard that made the original route acceptable.

Consider a public research office that uses a model to summarize policy documents in French and Wolof before a human editor prepares a public brief. A substitute model may be faster and cheaper, yet it may flatten a legal distinction, lose a source citation, or move sensitive text through a different jurisdiction. The workflow can still complete successfully in a technical sense while the institution's standard has changed. A model catalogue makes that change visible before the new route becomes the default.

Substitution evidence should travel with the route. It should record the task set, the accepted outputs, the unacceptable failures, the reviewer decision, and the date after which the comparison expires. The record should also name the dimensions that cannot be traded away. A health service may accept slower inference to preserve local processing. A publisher may accept a higher cost to retain rights metadata. A financial cooperative may prioritize reconciliation accuracy over conversational fluency.

This record turns vendor mobility into institutional mobility. The organization can change a model without losing the reasoning that justified its earlier choice. It can negotiate with a provider from a position of knowledge. It can maintain a local fallback without pretending that every model is equivalent.

Runtime routing needs catalogue knowledge

Google's ADK Go 2.0 announcement describes graph-based workflows, human-in-the-loop orchestration, dynamic routing, retries, and built-in resilience. These primitives become more useful when the runtime can read a model catalogue instead of treating every route as an opaque endpoint.

A routing node can ask whether the model is approved for the data class, whether its evaluation is still current, whether the task requires a local language checkpoint, and whether the available budget supports the route. A retry can choose a documented fallback rather than repeating the same failure. A human checkpoint can see why the system selected the model and which evidence supports the substitution. A retirement rule can block a route whose privacy terms or evaluation window has expired.

The catalogue therefore sits between the institution's charter and its runtime. The charter answers what the machine is authorized to do. The catalogue answers which model is permitted to perform that work and under what proven conditions. The distinction matters because authority without model evidence is too broad, while model evidence without authority is directionless.

The European Commission's public description of the General-Purpose AI Code of Practice adds another reason to keep this record. Provider obligations and institutional obligations meet at deployment. A provider can document the characteristics of a general-purpose model, while a deployer must still know how that model entered a particular workflow, what data it encountered, what review surrounded it, and what local standard the institution applied. The catalogue is the bridge between general model information and a specific institutional decision.

African institutions need a catalogue written in their own conditions

Imported AI stacks often arrive with an imported procurement grammar. That grammar prices access in a foreign currency, concentrates evaluation in English, assumes stable connectivity, centralizes identity, treats data residency as a cloud setting, and designs payment and support channels around a market that may not resemble the one adopting the system.

A sovereignty-oriented catalogue records the operating conditions that determine whether a model is genuinely useful. It can preserve local language test sets, translation review, intermittent-connectivity behaviour, mobile-money or bank-settlement costs, regional hosting, key ownership, and the evidence required by a public authority or community institution. The catalogue gives local institutions a way to compare foreign and domestic routes on the same terms.

This is a constructive path to technological independence. African institutions can evaluate external models against locally authored fields, publish shared evaluation schemas, and maintain fallback routes that preserve language and jurisdiction. A model earns a place in the catalogue when it meets that standard, while vendor defaults remain inputs to the decision rather than the decision itself.

Where the investable surface is widening

If model choice becomes a continuing institutional decision, the capital-relevant layer sits between model providers and deployed workflows:

  • Model registries and procurement systems: catalogues that connect capability claims to evaluation evidence, pricing, data terms, owners, and expiration dates.
  • Substitution and regression harnesses: test systems that compare models on the institution's real tasks, languages, failure cases, review burden, and policy requirements.
  • Policy-aware model gateways: routing layers that enforce data, jurisdiction, cost, and approval rules before a model receives an institutional request.
  • Regional model assurance: independent evaluation services and local test infrastructure for African languages, public-sector records, financial workflows, and connectivity constraints.
  • Model continuity and exit services: tools that preserve prompts, evaluation sets, decisions, evidence, and fallback routes when a provider changes its terms or a customer changes its architecture.

The underwriting question is concrete: can this product show why a model was selected, which evidence supports the choice, what would invalidate it, and how the institution can move to another route without losing its standard? A system that answers those questions reduces supplier dependence and makes model spending legible to finance, operations, regulators, and local authorities.

The catalogue is a record of institutional judgment

Model capability will continue to improve, and model prices will continue to move. An institution that treats each change as a fresh improvisation will accumulate hidden risk. An institution that maintains a living catalogue can compare, approve, substitute, retire, and recover with a memory of why each decision was made.

The catalogue gives procurement a technical spine and gives infrastructure a political context. It says that a model is more than an endpoint, because every endpoint carries assumptions about evidence, language, jurisdiction, cost, and responsibility. When those assumptions are written down and tested against local work, an institution can use global intelligence without surrendering the right to define its own standard.

A model may be rented from elsewhere. The judgment that places it inside an institution must be authored, evidenced, and kept portable.

Sources