llm-consistency-testing

LLM Consistency Testing

When to Use

Use this skill when you need evidence-bounded analysis, design, or validation preparation for repeat inputs, version/model/prompt factors, invariants, variance evidence, and comparison boundaries.
Use it to review an Agent, RAG, or LLM plan, result, or evidence set and produce actionable improvements.
Use it when context is incomplete but a bounded first pass with assumptions, gaps, and human-decision boundaries is still useful.

Default to Markdown organized by domain risk, evidence state, priority, and boundary.
When the user requests tables, CSV, JSON, or ticket fields, preserve the same finding fields, evidence, and decision boundaries.
Before machine consumption, confirm the schema, enums, required fields, and evidence sources.

Read and follow prompts/llm-consistency-testing.md, including its input audit, domain coverage, and output order.
Extract scope, environment, version, time window, constraints, success criteria, and available evidence, with attention to repeat input, version factor, model factor, invariant, variance evidence.
Separate confirmed facts, evidence-backed inferences, candidate recommendations, and Human decisions before ranking by risk and evidence strength.
Turn high-risk items into preconditions, steps, expected behavior or decision criteria, required evidence, and a validation method.
When input is incomplete, deliver a bounded first pass, state unsupported conclusions, and never present static material as execution evidence.

When calibrating domain outcomes, valid controls, or evidence gaps, read references/domain-decision-example.md for sanitized input, per-item oracles, and reproducible calculations.
Always read prompts/llm-consistency-testing.md; it is the complete execution specification for this skill.
For evaluation, read evals/eval.yaml and the matching cases under evals/cases/.
Load references/, examples/, scripts/, or output-formats.md only when those directories exist and the task needs them.

Fix sample ID, model/prompt/tool/data versions, temperature, seed where supported, and other sampling parameters, with project-supplied repetition counts. Equal seeds do not establish reproducibility across versions.
Check structural invariants first: schema, required fields, enums, and policy. Then compare semantic decisions using the project rubric; permitted paraphrases/order changes are not failures.
Report repeated-trial denominators, structural/semantic failure counts, and intermittent failures per sample. Variance needs a named numeric variable, unit, and formula; wording difference rates are not quality variance.
Single observations establish no repeated consistency. Missing rubrics leave semantic consistency UNKNOWN. Separate parameter/version groups rather than pooling them as stability evidence.

Keep the analysis focused on repeat inputs, version/model/prompt factors, invariants, variance evidence, and comparison boundaries; do not replace business owners or Human risk acceptance, exception approval, or safety decisions.
Never invent system behavior, fields, model outputs, data, thresholds, root causes, execution records, or pass claims.
Static design, plans, file presence, or a dry run retain their evidence state and cannot become proof of real execution.
When evidence is insufficient, use pending confirmation, blocked, unassessed, or NOT_SCORED and give the smallest validation method.
For user data, production, or safety work, use least privilege, masked data, mocks, dry runs, or isolation.

Covered repeat input, version factor, model factor, invariant, variance evidence, with source, evidence state, and validation method for each.
Separated facts, inferences, candidate recommendations, gaps, and Human decisions.
Gave high-risk items P0/P1/P2/P3 or an equivalent priority, owner role, and close condition.
Did not turn plans, static checks, or dry runs into test execution, all-passed, or safety-approved claims.
Stated residual risk, stop/escalation conditions, and next actions.

Listing checks without triggers, expected concerns, owner roles, close conditions, and evidence.
Treating adjacent tests or model tools as a complete LLM consistency judgment.
Using unexplained numbers for false precision or writing correlation as causation.
Refusing incomplete input, or pretending that incomplete evidence is conclusive.

Start with paths most likely to cause user harm, business loss, or decision blockage.
Use the smallest verifiable experiment to reduce uncertainty and record conditions, versions, sources, and evidence.
Make the Skill independently installable, executable, and reviewable by another engineer.