Diop Daily #110 — September 2026

The Proof Must Survive the Machine

A fluent answer can persuade a person for a moment. A proof artifact can let an institution decide whether the answer deserves to enter its memory, software, laboratory, or public record. As models produce increasingly technical work, the scarce layer is moving from generation to independent acceptance: a checker that can inspect the claim, reject a broken derivation, and replay the result under declared conditions.

OpenAI’s News RSS carried two relevant signals on September 8. One described GPT-5.6 Sol with Codex autonomously running quantum-computing experiments, analyzing results, and calibrating qubits. Another described an AI-generated solution to the Navier–Stokes Millennium Prize Problem, accompanied by a writeup and a formal proof in Lean. These are company-reported descriptions rather than independent confirmations that the scientific claims are correct, yet they identify a change in the object being produced: the system is being asked to create technical work that can be checked by a formal language rather than admired as prose.

The mature output of a reasoning machine is a claim carrying its own route to rejection.

Persuasion and verification have different economies

Human review is powerful, but it is also scarce, contextual, and uneven. A researcher can recognize a familiar error. An engineer can inspect an implementation. A mathematician can find an unspoken assumption. These forms of judgment remain necessary because formal systems cover only what has been specified. Their cost rises when the machine produces more candidate results than a specialist can read.

Formal verification changes the division of labor. A proof assistant checks whether a proposition follows from declared definitions and previously accepted rules. A type checker can reject a program that violates a contract. A model checker can explore whether a finite-state system reaches a forbidden condition. A reproducibility runner can compare a result under a fixed environment. Each instrument narrows the space in which fluent error can hide.

The distinction matters for investment. A model endpoint sells the possibility of an answer. A proof system sells a boundary around what may be accepted. The second object can become part of procurement, release, research, and regulation because it creates an observable decision: accepted under these assumptions, rejected under these tests, or unresolved because the formalization is incomplete.

A proof is an executable institutional claim

Lean’s public project describes a language and theorem prover for formalizing mathematics and writing programs whose properties can be checked. Its significance for machine-assisted work is architectural. The proof is stored in a form that another checker can process. The institution does not have to trust the original author’s memory of every inference. It can inspect the proposition, the imported definitions, the tactics or terms, the environment, and the result of the checker.

Formal proof remains bounded by its specification. A theorem can formalize the wrong question. A model can encode an unjustified assumption as an axiom. A proof can establish a narrow proposition that readers mistakenly expand into a broader conclusion. The checker guarantees a relation inside the formal system; people still decide whether the formal system represents the world they care about.

A useful proof record should therefore carry more than a green status:

  • Claim: the exact proposition, contract, invariant, or safety property under review.
  • Formal context: definitions, axioms, data assumptions, environment, dependencies, and version identifiers.
  • Derivation or test: the proof term, program, model, simulation, or replay procedure that produces the result.
  • Checker: the independent tool, version, configuration, and limits used for acceptance.
  • Boundary: what the result establishes, what it leaves unresolved, and which real-world conditions could invalidate the mapping.
  • Authority: the person or institution responsible for deciding whether the checked result is fit for use.

These fields separate a checked claim from a claim merely accompanied by technical vocabulary. They also let an institution inherit the work. A new researcher can rerun the checker. A new engineer can test whether a change preserves the invariant. A public body can ask which assumptions made a safety statement true. A second model can propose a revision while the acceptance boundary remains independent of the model that generated it.

Agents need acceptance boundaries

Agent runtimes are becoming capable of planning, coding, running tools, and handing work to other processes. Google’s public description of ADK Go 2.0 names graph workflows, human-in-the-loop orchestration, dynamic routing, retries, and resilience. Google’s A2A material describes secure autonomous handoffs. These primitives can move a technical task through a system, but movement is not validity.

An agent that writes a numerical method should be able to attach tests for its invariants. An agent that changes a database migration should produce a replayable check for schema compatibility and rollback. An agent that proposes a scientific result should identify which claims are formally verified, which are numerically supported, and which remain interpretive. The next agent should receive the acceptance boundary with the artifact instead of receiving only a confident summary.

This creates a practical architecture:

  1. The generator proposes a claim and declares the assumptions it used.
  2. A formalizer translates the relevant part into a theorem, type contract, invariant, test, or reproducible computation.
  3. An independent checker accepts, rejects, or returns an unresolved obligation.
  4. A human authority decides whether the checked result is fit for the institution’s purpose.
  5. The record preserves the claim, context, checker result, decision, and later corrections.

The checker protects judgment from carrying the whole burden of memory and arithmetic. The machine may propose faster than a specialist can read, while the acceptance boundary keeps the institution from treating fluency as evidence.

Formal methods meet African scientific sovereignty

Scientific sovereignty requires the ability to test claims with instruments an institution can inspect and maintain. Imported intelligence without local verification leaves the final authority elsewhere. A research center may receive a model’s result, yet lack the formal environment, compute, language resources, or technical staff needed to reproduce the claim. That is access without control.

African institutions can build a different position by treating proof infrastructure as shared capacity. Universities and laboratories can maintain formal libraries for local mathematics, engineering standards, public-service rules, agricultural models, language technologies, and payment protocols. They can train researchers to move between empirical evidence and formal representation. They can publish benchmark problems whose assumptions reflect local environments rather than only imported examples.

This work also demands linguistic care. Formalization often appears language-neutral because symbols travel well. The choice of what becomes a variable, a category, a legal condition, or an admissible exception still reflects an institution’s concepts. A water-allocation model, a land record, a clinical protocol, or a cooperative rule may encode relationships that general-purpose datasets do not represent. The proof can be valid while the model of the institution remains incomplete.

NIST’s AI Risk Management Framework gives this architecture a general vocabulary through Govern, Map, Measure, and Manage. Govern assigns responsibility for the claim and its acceptance. Map identifies the people, data, systems, and obligations represented by the formal context. Measure records failed checks, unresolved obligations, reproduction cost, and the difference between formal validity and field performance. Manage defines what happens when an assumption changes, a checker is upgraded, or a result is later challenged.

Where the investable surface is widening

If machine-generated technical work becomes common, capital should examine the infrastructure that makes acceptance portable and measurable:

  • Proof-carrying generation: model and agent systems that produce claims alongside formal obligations, tests, contracts, or replay instructions.
  • Verification runtimes: managed environments that execute proof checkers, type systems, model checkers, simulations, and reproducibility jobs with declared versions and limits.
  • Formalization services: tools and expert systems that translate natural-language requirements, scientific hypotheses, and software specifications into checkable representations.
  • Evidence-to-proof registries: records connecting empirical data, formal assumptions, checker output, human acceptance, and later correction.
  • Sovereign verification infrastructure: local libraries, benchmark suites, training systems, and compute that let institutions verify imported machine work without surrendering the acceptance layer.

The underwriting question is precise: does the product reduce the cost of deciding whether a machine-generated claim is fit for use? Useful measures include checker coverage, unresolved-obligation age, reproduction time on an independent environment, false-acceptance rate, percentage of releases carrying executable contracts, and the share of proof records that remain portable after a model or provider changes.

This is different from an experiment ledger. The ledger preserves what a laboratory tried and learned. It is also different from provenance infrastructure, which preserves where an artifact came from. Proof infrastructure narrows the acceptance boundary itself. It tells the institution which proposition, invariant, or contract has been demonstrated under which formal conditions, and where human judgment must continue.

Build the acceptance boundary before scaling output

An institution can begin with one class of work where error is expensive and the claim can be stated clearly:

  1. Choose the proposition, safety property, interface contract, or calculation that must survive review.
  2. Write the assumptions and define which real-world observations connect the formal object to the institution’s purpose.
  3. Require the generator to produce a checker input, test, proof obligation, or replay procedure with the output.
  4. Run the acceptance process in an environment independent of the generator and preserve failed attempts as evidence.
  5. Measure the gap between formal acceptance and field performance, then revise the model rather than hiding the gap.

This is how institutions can use stronger machines without granting them final authority. A result earns its place by carrying a path to challenge. The proof must survive the machine because the future laboratory, software system, and public institution will inherit claims from systems that no human can fully remember. Their freedom will depend on whether they can still test what they receive.

Sources