Skip to content

ADR-0006: Bound pre-execution reasoning and material guessing

  • Status: Accepted
  • Date: 2026-08-06
  • Parent decision: ADR-0001
  • Related requirements: TRD-0001 Sections 10.1 and 10.2 (upstream document, not included in this repository)

Context

Prompt interpretation and response planning do not require the substantive reasoning needed to solve the user's task. Allowing full task analysis before confirmation can leak answers into artifacts, anchor the later execution, and spend tokens on an interpretation or approach the user may reject.

Eliminating reasoning entirely is also incorrect. Producing faithful pseudocode requires bounded semantic work, including:

  • resolving ordinary references and instruction relationships;
  • preserving priorities, exclusions, conditions, and ordering;
  • distinguishing a complete interpretation from a response strategy;
  • producing valid structured English;
  • checking that a plan is inspectable without predicting its result.

Natural language always requires some inference. Requiring clarification for every possible interpretation would create excessive interruption and defeat the purpose of the visible correction loop.

Decision

Prompt and Plan phases SHALL use only the reasoning necessary to create and validate their respective user-visible artifacts.

Before execution, the model SHALL NOT:

  • solve the task;
  • research substantive findings;
  • calculate requested results;
  • select winners or conclusions;
  • invent missing requirements;
  • silently improve the request;
  • fill consequential gaps with speculative assumptions.

Ordinary semantic inference MAY be used when it produces one coherent, inspectable interpretation without materially adding to the request. The Prompt Pseudocode SHALL expose the interpretation actually formed so the user can correct it.

When missing information prevents a coherent artifact or requires choosing between materially different goals, scopes, output requirements, safety postures, costs, or irreversible actions, the system SHALL request the smallest targeted clarification needed. Clarification SHALL not become a general ambiguity-enumeration DSL.

The controller SHOULD remove reasoning about protocol branching by presenting explicit events. In particular, Plan review SHOULD distinguish Revise approach from Change request rather than asking the model to guess which kind of change the user intended.

This decision does not limit substantive reasoning during execution. After the confirmed pair is available, the executor MAY use the reasoning effort required to satisfy the request safely and correctly while remaining within the confirmed plan.

Consequences

Positive

  • Fewer tokens are spent solving rejected interpretations or plans.
  • Prompt and Plan artifacts are less likely to contain premature conclusions.
  • The user retains control over consequential assumptions.
  • Ordinary requests are not interrupted by exhaustive ambiguity analysis.
  • Execution reasoning begins from confirmed semantic and procedural boundaries.

Negative

  • The boundary between harmless inference and material guessing still requires semantic judgment.
  • Excessively weak pre-execution reasoning can omit instruction relationships.
  • Excessively aggressive clarification can create interaction fatigue.
  • Model reasoning controls are provider-dependent and cannot replace behavioral validation.

Alternatives considered

Use no reasoning before execution

Rejected. Semantic interpretation, PDL generation, and plan-poisoning detection cannot be performed reliably without any inference.

Allow full analysis before confirmation

Rejected. It increases cost and permits findings to anchor or contaminate the confirmation artifacts.

Ask about every ambiguity

Rejected. The artifact correction loop is the normal ambiguity-resolution mechanism; targeted clarification is reserved for materially blocking gaps.

Amendment: Proportional pre-execution reasoning (Decision D25)

  • Status: Accepted. Date: 2026-09-16. Amended for gpt-oss by ADR-0022; per-model values move to data under ADR-0028.
  • Implementation and evidence: IMPL-0004.

Reasoning before execution is allocated per operation, by what the operation is for, and is the same for every model:

  1. Semantic isolation (the bootstrap read of raw user content) gets the most deliberation, because it must separate injected directives from the task.
  2. Prompt drafting gets bounded deliberation: enough to preserve every entity and requirement, never enough to start solving.
  3. Planning and execution get what each model needs to translate confirmed artifacts faithfully. The amount is set per model, not fixed. As ratified, the amendment set both to zero reasoning. ADR-0022 found that this made gpt-oss plans copy the prompt verbatim, so the amount is now a per-model choice.

Each model's own reasoning control expresses this allocation. Effort labels are not token counts: providers implement them differently, and some models have no reasoning control at all. A model class with no native control gets its deliberation in-band, through structured output fields. Starving a lower-capacity model of reasoning is not a neutral choice: it degrades the faithfulness the protocol depends on.