Completion Criteria Guide

Completion Criteria Guide

Why completion criteria exist

<completion_criteria> is the contract that prevents a model or operator from confusing activity with completion. It answers:

> What observable evidence proves this exact prompt has fulfilled its responsibility?

It is not a summary of the mission, a motivational sentence, or a copy of suite-wide truth rules.

Required properties

Every completion block must:

1. contain at least three explicit conditions; 2. name the task, artifact, decision, or changed state; 3. include domain-specific success conditions; 4. require =VERIFY:{id} evidence; 5. state what happens to unknowns, failures, residuals, and missing authority; 6. remain unique to the prompt.

Completion by role

Control

Completion proves the shared contract is resolvable by downstream prompts.

Example conditions:

• the runtime context identifies scope, authority, and protected surfaces;
• downstream prompts can determine permitted modes;
• unresolved inputs are ?UNKNOWN:{id};
• the control artifact has an =VERIFY:{id} integrity record.

Investigative

Completion proves the evidence and handoff are ready, not that the proposed action occurred.

Require:

• evidence index;
• finding register;
• bounded actions or recommendations;
• acceptance criteria;
• evidence freshness;
• explicit unknowns and contradictions;
• frozen handoff verification.

Executive

Completion proves approved actions were handled and acceptance criteria were evaluated.

Require:

• each approved +ACTION:{id} executed, skipped, or deferred;
• unchanged action IDs;
• current approval and handoff;
• =VERIFY:{id} result for every criterion;
• rollback and residual state;
• hard stop for scope expansion or failed verification.

Operational

Completion proves the declared artifact or workflow works for its audience and medium.

Require the exact source artifact, exact exports, domain quality gates, and unresolved dependencies.

Gate

Completion is an independent decision, not a repair. Require the decision, supporting evidence, blocking findings, and verification status.

How to derive task-specific criteria

Use these sources in order:

1. mission; 2. primary artifact; 3. domain quality gates; 4. investigative verification design; 5. execution method; 6. audience task; 7. required residual and rollback records.

Do not derive completion solely from generic control text.

Good patterns

Creative writing

<completion_criteria>
Completion requires all of the following:
- The manuscript fulfills the requested form, viewpoint, length, and scene purpose.
- Every scene changes the situation and preserves causal continuity.
- Character motives and language remain distinct, and concrete detail carries emotional work.
- Editorial review records each accepted quality gate as =VERIFY:{id}.
- Unresolved continuity questions remain ?UNKNOWN:{id}; rights or safety conflicts trigger !STOP:{reason}.
</completion_criteria>

Security remediation

<completion_criteria>
Completion requires all of the following:
- Every approved remediation +ACTION:{id} is executed, deferred, or rejected with evidence.
- The original vulnerable path is no longer reproducible under the authorized test plan.
- Negative and abuse-case tests pass without weakening another control.
- Every investigative criterion has an =VERIFY:{id} result tied to exact commands or observations.
- New findings are not silently fixed; they enter the residual register or trigger !STOP:{reason}.
</completion_criteria>

Report production

<completion_criteria>
Completion requires all of the following:
- The report answers the declared decision question and reconciles every material figure to its source.
- Exhibits, narrative, terminology, and recommendations agree.
- The exact DOCX/PDF/export is checked for clipping, broken links, accessibility, and pagination.
- Claims and exports have =VERIFY:{id} records; unresolved evidence is ?UNKNOWN:{id}.
</completion_criteria>

Anti-patterns

Reject:

• “Complete when high quality.”
• “Complete when evidenced and honestly closed.”
• “Complete when all requirements are met.”
• a copy of <quality_gates> with no artifact or verification requirement;
• a claim that execution occurred in an investigative prompt;
• a gate that repairs the subject before issuing a decision;
• a completion block shared unchanged across unrelated prompts.

Static checks

The body audit reports:

• completion blocks present;
• unique normalized blocks;
• known boilerplate matches;
• duplicate count;
• bullet count;
• task-title overlap;
• =VERIFY:{id} presence.

Run:

python tools/audit_prompt_bodies.py --check

Static uniqueness is necessary but not sufficient. Live fixtures must still test whether models follow the conditions.