Problem Statement
Prompt iteration is a text-level patch mechanism. Past a bounded window it stops being a control and becomes an uncontrolled search over wording. Iteration becomes a decision failure when it continues after the work can no longer produce defensible, auditable acceptance. Past that point, additional iterations increase variance and operational risk while decreasing traceability.
| Decision failure state | Resulting condition |
|---|---|
| Iteration substitutes for acceptance definition | Acceptance cannot be proven |
| Iteration substitutes for evidence | Outputs remain non-defensible |
| Iteration substitutes for control | Risk becomes unbounded |
Decision Context
This page governs the stop decision for iterative prompt changes used to generate outputs that will be relied upon for accountable action.
| Control surface | Required property |
|---|---|
| Requirements | Stable, non-negotiable constraints exist |
| Evidence | Outputs can be audited end-to-end |
| Risk envelope | Authorized tolerance is not exceeded |
Failure Modes
Why Iteration Fails
Prompt iteration alters phrasing, not the underlying evidence, constraints, or auditability. When acceptance cannot be evidenced or constraints are incompatible, further iteration is non-convergent and becomes invalid as a decision method.
Common Failure Modes
| Observable system-level failure | Decision meaning |
|---|---|
| Material output drift under the same constraints | Non-reproducible state |
| Compliance alternates across attempts | Incompatible constraints |
| Novel errors appear while prior errors persist | Non-convergent iteration |
| Outputs cannot be traced to inputs and constraints | Audit collapse |
| Small changes create large, unpredictable shifts | Unstable control surface |
Decision Boundary
| Boundary condition | Status |
|---|---|
| No STOP criteria met and acceptance evidence remains stable and auditable | Iteration remains valid |
| Any STOP criterion met | Iteration is invalid and must stop |
STOP Criteria
Iteration must end immediately if any condition below is observed.
| STOP criterion | Observable trigger |
|---|---|
| Safety, legal, policy, or confidentiality breach | Protected content exposed or prohibited output produced |
| Non-negotiable requirements are mutually exclusive in practice | Satisfying one constraint consistently violates another |
| Acceptance is undefined, unmeasurable, or changes during iteration | No stable pass/fail exists |
| Reproducibility is lost | Same inputs produce materially different outcomes beyond declared variance |
| Auditability is lost | Inputs, constraints, and outputs cannot be fully reconstructed |
| Controlled risk envelope is exceeded | Latency, cost, or downstream impact breaches the authorized limit |
| Iteration becomes the requirement | Prompt changes replace system-level specification as the acceptance basis |
What This Page Is NOT
- Prompt engineering instruction, iteration technique, or optimization guidance
- A framework for expanding iteration scope or increasing attempt counts
- A method for resolving missing requirements, missing evidence, or missing controls
- A justification for accepting non-auditable outputs
- A substitute for system governance, evaluation design, or risk authorization
Closure
Decision boundary: continue only while no STOP criterion is met and acceptance remains stable and auditable. STOP: continuation is invalid beyond this boundary.