Naur's human-only conclusion needs more than the absence of explicit criteria

Type: kb/types/note.md · Tags: foundations, context-engineering

Peter Naur's Programming as Theory Building connects two claims: program theory cannot be expressed as rules or criteria, and it is bound to human beings. The second does not follow from the first alone. It also needs a reason why computation cannot perform the relevant judgments without their criteria having been explicitly supplied.

This note reconstructs one bridge between those claims. It is an interpretation of the essay, not a premise Naur states as a separate theorem. Removing that bridge reopens the computational question; it does not establish that a current model or automated house holds a program theory.

What Naur asks a theory-holder to do

Drawing on Ryle, Naur describes a capacity to relate program structure to the world, justify the program's parts, and incorporate new demands by recognizing relevant similarities. Following an explicit rule does not exhaust that capacity: applying a rule itself can require judgment. Naur uses this regress to argue against identifying intelligent performance with adherence to a formulated method.

That gives a demanding functional target. A modifier must do more than recall facts or produce a patch that passes today's tests. It must understand how a change relates to the program's purpose and organization, including implications not already stated as instructions.

Naur also locates theory in programmers' mental possession and stresses their knowledge of the world. These are part of his account, not consequences that his maintenance cases independently establish for every possible computational system.

The reconstructed bridge

The essay contrasts formal symbol manipulation with similarity judgments whose criteria cannot be formulated. One reading reconstructs the inference as:

  1. Program-theory judgments cannot be reduced to explicitly formulated criteria.
  2. A computer can make a judgment only by executing such criteria.
  3. Therefore a computer cannot perform the program-theory function.

The second premise is doing essential work. It identifies formal execution with the prior formulation of a judgment's criteria. Even granting the first premise does not establish the conclusion without it. Nor does excluding computers alone establish that humans are the only conceivable bearers.

The claim that this bridge captures Naur's reasoning must remain distinguishable from the passages supporting it. Another reading may treat his human-only position as an additional commitment rather than a conclusion from the rule-following argument. Either way, a human-only premise needs its own support.

Formal execution does not require a hand-written judgment rubric

A learned model is executed by defined computational operations. That fact does not mean its designers supplied a project-specific rubric for every judgment it can make. A fixed model can be asked to interpret a new explanation and relate it to a case without receiving a complete decision rule for that case. Whether it does so adequately is the empirical issue.

Two claims must stay separate: the criteria were not explicitly supplied, and the criteria cannot be formulated. The first does not prove the second. A model's numerical parameters and execution rules specify a computation, but that specification need not be a useful human-readable account of its judgment. Difficulty explaining the judgment is not a proof that such an account is impossible.

The computational proposal therefore need not defend inexpressibility or claim that machines were incapable of learned judgment when Naur wrote. Its narrower point is that absence of a supplied rubric does not by itself rule out computational performance. A working computational theory-holder would challenge the human-only thesis while leaving the expressibility question separate.

What the transfer evidence establishes

Naur's compiler case describes a motivated successor group with program text, annotations, extensive design discussion, and personal advice. It still proposed changes the original group regarded as damaging patches, where the original group could propose changes within the existing structure.

The negative result concerns that documentation-and-consumption system. It does not show that all possible records or interpreters must fail. It also does not show that the supplied records were sufficient and merely needed better retrieval. Missing premises, missing application skill, and failure to activate relevant understanding are different possible explanations.

Retained theory and holding a theory must therefore remain separate. Retained text is one addressable carrier. Holding a theory is a capacity of the system that consumes the available state. A document that is never used supplies no evidence of that capacity; successful reconstruction from other records need not require that document.

Tests for a computational bearer

Three tests follow from this functional reading:

  • Project-specific understanding. General programming competence is not the theory of this program. The system must relate the actual program to its purposes, justify consequential choices, and handle novel modifications. The understanding may be supplied in the seed or acquired during operation; its use and its acquisition are separate claims.
  • Access to necessary premises. A decision premise not recoverable from the implementation and general knowledge must reach the modifier through some other permitted path. Attempted recovery can expose that information gap without making one document format necessary.
  • Reliable continuation. One coherent extension may be lucky or familiar. The system must sustain coherent search, diagnosis, recovery, and revision across later demands, including evidence against earlier assumptions.

Holding program theory under delayed feedback develops the last test. Interventions on retained commitments or their consumption paths can test their causal contribution. Altering a written carrier alone does not isolate the whole capacity when other records can reconstruct it.

Scope

This is a criticism of an inference, not a proof of computational theory possession. The functional tests are an operational reading of Naur for a research program, not a claim that he would accept every system passing them. A successful computational witness would bear on human exclusivity; it would not settle every philosophical claim about knowledge or rule-following.

Relevant Notes: