Resources / Cognition and language
From Hypothesis to Decidability
Grammar, contextual prediction, and the distinct work of constructive examination.
In this article
- 1. Finite means and open-ended expression
- 2. What a transformer contributes
- 3. Discrete language, continuous representations
- 4. The examination has a different object
- 5. From expression to testimony
- 6. Construction is falsificationary
- 7. Two loops with different stopping conditions
- 8. Variation, selection, and retention
- 9. The institutional completion
Finite mechanisms can produce an open-ended range of language. That observation connects formal grammars, language models, and Runcible’s account of recursive disambiguation. The connection is useful only if we preserve what differs: the object processed, the mechanism, the test of success, and the purpose of the output.
Our broader thesis is that cognition repeatedly distinguishes a situation, proposes a continuation or explanation, and tests what the situation permits. Runcible applies that operational account to testimony and institutional judgment. It is a research framework, not a claim that every cognitive system has the same implementation.
1. Finite means and open-ended expression
A formal grammar can specify how a finite set of rules generates indefinitely many expressions. Recursive structures allow a phrase to contain another phrase of a related kind: “the manager who approved the contract that changed the schedule.”
The generative-grammar tradition associated with Chomsky investigates the structure and capacity of human language. Its central questions differ from whether a particular sentence is true, whether its proposed action is reciprocal, or whether the speaker has authority.
A grammatical expression can assert a falsehood. A recognizable sentence can conceal an unidentified actor. Linguistic generativity is a necessary contribution to communication, but it does not supply the entire examination of testimony.
2. What a transformer contributes
A transformer-based language model represents tokens as vectors, combines information through attention and other learned transformations, and produces a distribution used to generate subsequent output. Autoregressive generation then extends the token sequence and continues.
The distinction between that iterative generation loop and architectural recurrence matters. The original Transformer was explicitly introduced without recurrent sequence-processing layers. Calling every layer or token step “recursion” obscures this engineering distinction. Vaswani et al., Attention Is All You Need.
In this essay, recursive disambiguation names an operational interpretation: a partial context is repeatedly elaborated, differentiated, and used to propose what comes next. It is not a claim that a standard transformer is a recurrent neural network or directly implements a symbolic grammar.
3. Discrete language, continuous representations
The model processes discrete token identifiers through continuous-valued representations. Attention helps combine information relevant to a token’s context; feed-forward transformations, residual connections, normalization, and output projections also contribute. The useful result cannot be assigned to attention alone.
We interpret this process as forming a provisional context identity: what kind of situation is being described, which relations matter, and which continuation fits. “Context identity” is a functional description, not a claim that one particular node stores a complete situation model.
The model’s learned parameters, current context, and any external retrieval or memory are also different forms of retention. Adding a conversation to the context does not necessarily alter the model’s weights. An institutional record requires explicit persistence and access control beyond the generated sequence.
4. The examination has a different object
Prediction concerns a possible continuation. Examination takes the resulting claim as an object in its own right. It asks what the claim requires and what could defeat it.
| Function | Object | Relevant question |
|---|---|---|
| Grammatical generation | Linguistic structure | Can this expression be generated within the grammar? |
| Model generation | Context and candidate continuation | What output does the model produce under these inputs and controls? |
| Constructive falsification | Operational claim and necessary dependencies | What survives the required challenges, and what remains open? |
| Institutional authorization | Proposed action and responsible role | Who may direct this action under the current procedure and limits? |
These functions can use common computational components. Their responsibilities must still be separated. A model capable of generating criticism does not establish that the required criticism was performed adequately.
5. From expression to testimony
An expression becomes a claim when it asserts something about a state of affairs. It becomes testimony in the relevant institutional sense when another party is asked to rely upon it under a burden of truthfulness and responsibility.
That transition adds obligations. Which terms identify actual referents? What event or operation is asserted? What supports the causal relation? Which facts are observed, inferred, or assumed? What limits are material? Who accepts responsibility for the proposed reliance?
Runcible’s operational grammar reconstructs these obligations. A claim about a delivered service, for example, must distinguish the contract, the performed work, acceptance, an invoice, and an authorized payment. These can be closely related while remaining different events.
6. Construction is falsificationary
Constructing a claim from first principles does not mean assembling a more persuasive justification. It means specifying the objects, operations, resources, sequence, and constraints that the claim needs in order to hold.
A contradiction can defeat the construction. A missing prerequisite can prevent it from being established. A feasible construction can survive a possibility test while still lacking evidence that the asserted event occurred. A factual finding can survive while the proposed use fails reciprocity or authority.
The relevant outcomes therefore include survival within scope, defeat, repair, and an explicitly unresolved condition. “No counterexample was found” is not sufficient if the search omitted a material dependency.
7. Two loops with different stopping conditions
The generation loop extends a candidate. The examination loop reconstructs and challenges it. A repair returns to examination; it does not inherit the original proposal’s status.
Runcible’s contextual notion of decidability is the satisfaction of the required standard in the context at issue without an arbitrary substitution of discretion for a missing finding. Where the standard cannot be met, the system must preserve non-closure or identify the human adjudication required.
This differs from a universal algorithm that decides every instance of a formal problem. Bounded examination does not erase classical limits on computability. It identifies the conclusion available for the particular matter and the conditions on which that conclusion depends.
8. Variation, selection, and retention
The evolutionary analogy is useful because candidate generation can be productive even when many candidates fail. Selection becomes informative when the system records the particular failure and its cause. Retention becomes useful when later work can recover the distinction and test whether it still applies.
A rejected causal explanation may reveal a missing observation. A failed policy may reveal an uncompensated transfer. A denied execution may reveal a scope mismatch. Those are different learning signals; treating them all as “bad output” discards their value.
The analogy does not establish that neural inference, biological evolution, and legal adjudication are the same mechanism. Their variations, constraints, time scales, and retention processes must be specified separately. The proposed commonality is an operational pattern that guides investigation.
9. The institutional completion
The Decidability Record preserves the examined claim, evidence, tests, failures, repairs, limits, and unresolved matters. Oversing connects that record to institutional roles, commitments, resources, authorized work, and consequences.
The generating model cannot promote its own finding into a mandate. Oversing governs decision and authorization; a separate executor receives the permitted task. The outcome then becomes evidence about the earlier account.
That is the progression this framework seeks to make explicit: language can supply a hypothesis; constructive falsification can establish what reliance survives; institutional authority can permit a bounded action; and the continuing record can support correction and adaptation.
