Deterministic Code
Program logic that necessarily produces the same output given the same input. In harness approaches it takes on three roles: performing precise and consistent operations, gating and validation, and decomposing task steps. It contrasts with sampling-based language model generation.
Timeline (last 90 days)
In small-model applications, everything that must be exact (such as deadlines, amounts, reference numbers, issuing institutions) should be handled by deterministic code, and everything that needs to be understood (such as explaining the letter) should be left to the model.
The AI tutor system is split into three types of components: humans write and run the SQL; deterministic code handles the try-first rule, the hint/answer rules, the answer checker and the leak-prevention guard; and the probabilistic LLM only writes the hint text.
Deterministic code refers to programs that always produce the same output for the same input and can be exhaustively verified with unit tests, whereas the sampling mechanism of a probabilistic LLM gives its output inherent randomness.
Steps that can be backed by deterministic code (such as contrast validation, format validation, numeric range checks, and sensitive-word filtering) should not be handed to probabilistic models.
Diagnostic comparison table: when a rule must be strictly enforced, first check the approved rule description, inputs, edge cases and authorization, and the candidate intervention is to implement it in deterministic code and test it.
For a long time, the sole building material of software was deterministic code, written line by line by human engineers, predictable and reproducible.
Deterministic code and probabilistic AI material do not replace each other but coexist in a complementary manner.
All Facts (7)
In small-model applications, everything that must be exact (such as deadlines, amounts, reference numbers, issuing institutions) should be handled by deterministic code, and everything that needs to be understood (such as explaining the letter) should be left to the model.
50%UnverifiedThe AI tutor system is split into three types of components: humans write and run the SQL; deterministic code handles the try-first rule, the hint/answer rules, the answer checker and the leak-prevention guard; and the probabilistic LLM only writes the hint text.
50%UnverifiedDeterministic code refers to programs that always produce the same output for the same input and can be exhaustively verified with unit tests, whereas the sampling mechanism of a probabilistic LLM gives its output inherent randomness.
50%UnverifiedSteps that can be backed by deterministic code (such as contrast validation, format validation, numeric range checks, and sensitive-word filtering) should not be handed to probabilistic models.
50%UnverifiedDiagnostic comparison table: when a rule must be strictly enforced, first check the approved rule description, inputs, edge cases and authorization, and the candidate intervention is to implement it in deterministic code and test it.
50%UnverifiedFor a long time, the sole building material of software was deterministic code, written line by line by human engineers, predictable and reproducible.
50%UnverifiedDeterministic code and probabilistic AI material do not replace each other but coexist in a complementary manner.
50%