reproducibility-validate

namespace	aiwg
name	reproducibility-validate
platforms	["all"]
description	Run a workflow multiple times and compare outputs to produce a similarity score and pass/fail verdict

Reproducibility Validate

You run a workflow multiple times and compare outputs to produce a similarity score and pass/fail verdict, confirming that the workflow produces consistent results across executions.

Triggers

Alternate expressions and non-obvious activations (primary phrases are matched automatically from the skill description):

• "is this workflow stable" → run reproducibility validation with defaults
• "check if results are consistent" → run reproducibility validation
• "does this run the same way every time" → run reproducibility validation
• "test determinism" → run reproducibility validation
• "compare workflow outputs" → run reproducibility validation

Trigger Patterns Reference

Pattern	Example	Action
Default validation	"validate reproducibility of onboarding-flow"	Run aiwg reproducibility-validate onboarding-flow
Custom run count	"validate with 5 runs"	Run aiwg reproducibility-validate <id> --runs 5
Custom threshold	"validate with 99% threshold"	Run aiwg reproducibility-validate <id> --threshold 0.99
Full options	"3 runs, 90% threshold"	Run aiwg reproducibility-validate <id> --runs 3 --threshold 0.90

Behavior

When triggered:

1. Extract intent:

   • What is the workflow ID or name to validate?
   • How many runs? (default: 3)
   • What similarity threshold must be met to pass? (default: 0.95)

2. Run the command:

   # Default: 3 runs, 0.95 threshold
   aiwg reproducibility-validate <workflow-id>
   
   # Custom run count
   aiwg reproducibility-validate <workflow-id> --runs 5
   
   # Custom threshold
   aiwg reproducibility-validate <workflow-id> --threshold 0.99
   
   # Both
   aiwg reproducibility-validate <workflow-id> --runs 5 --threshold 0.99
   

3. Scoring method:

   • Structured outputs (JSON, YAML): exact match required — score is 0 or 1 per artifact
   • Text outputs (Markdown, prose): semantic similarity score (0.0–1.0) computed across runs
   • Overall score: weighted average across all artifacts in the workflow

4. Pass/fail verdict: If all per-artifact scores meet or exceed the threshold, the validation passes. Any artifact below the threshold is flagged.

5. Report the result — provide the overall verdict, overall similarity score, and a per-artifact breakdown showing which artifacts passed or failed.

Examples

Example 1: Default validation

User: "Validate reproducibility of the risk-management-cycle workflow"

Extraction: Workflow ID risk-management-cycle, default 3 runs, default 0.95 threshold

Action:

aiwg reproducibility-validate risk-management-cycle

Response:

Reproducibility Validation: risk-management-cycle
Runs: 3 | Threshold: 0.95 | Result: PASS

Artifact                                  Score    Status
.aiwg/risks/risk-register.md             0.97     PASS
.aiwg/risks/mitigation-plan.md           0.96     PASS
.aiwg/working/risk-analysis-draft.md     1.00     PASS (exact match)

Overall score: 0.977 — all 3 artifacts passed.

Example 2: Stricter threshold

User: "Validate the SAD generation workflow with 99% threshold and 5 runs"

Extraction: Workflow ID sad-generation, 5 runs, 0.99 threshold

Action:

aiwg reproducibility-validate sad-generation --runs 5 --threshold 0.99

Response:

Reproducibility Validation: sad-generation
Runs: 5 | Threshold: 0.99 | Result: FAIL

Artifact                                       Score    Status
.aiwg/architecture/software-architecture.md   0.98     FAIL
.aiwg/architecture/component-diagram.mmd      1.00     PASS (exact match)
.aiwg/architecture/adr-001.md                 0.99     PASS

Overall score: 0.990 — 1 of 3 artifacts failed the 0.99 threshold.
Recommendation: Review non-deterministic sections in software-architecture.md,
or lower the threshold to 0.95 if semantic variation is acceptable.

Example 3: High-confidence validation before release

User: "Run 5 iterations of the deployment-planning workflow and check for consistency"

Extraction: Workflow ID deployment-planning, 5 runs, default 0.95 threshold

Action:

aiwg reproducibility-validate deployment-planning --runs 5

Response:

Reproducibility Validation: deployment-planning
Runs: 5 | Threshold: 0.95 | Result: PASS

Artifact                                  Score    Status
.aiwg/deployment/deployment-plan.md      0.96     PASS
.aiwg/deployment/runbook.md              0.97     PASS
.aiwg/deployment/rollback-plan.md        0.95     PASS

Overall score: 0.960 — all 3 artifacts passed across 5 runs.

Example 4: Validating a failing workflow

User: "Test whether the test-strategy workflow is deterministic"

Extraction: Workflow ID test-strategy-execution, default 3 runs, default 0.95 threshold

Action:

aiwg reproducibility-validate test-strategy-execution

Response:

Reproducibility Validation: test-strategy-execution
Runs: 3 | Threshold: 0.95 | Result: FAIL

Artifact                               Score    Status
.aiwg/testing/test-plan.md            0.94     FAIL
.aiwg/testing/test-strategy.md        0.97     PASS
.aiwg/testing/coverage-targets.json   1.00     PASS (exact match)

Overall score: 0.970 — 1 of 3 artifacts failed the 0.95 threshold.
Recommendation: Enable `strict` execution mode (`aiwg execution-mode strict`)
to reduce variance in test-plan.md, then re-validate.

Clarification Prompts

If the user's intent is ambiguous:

• "Which workflow should I validate? (e.g., risk-management-cycle, sad-generation)"
• "How many runs would you like? Default is 3; more runs give higher confidence but take longer."
• "What similarity threshold should I apply? Default is 0.95. Use 0.99 for near-exact determinism requirements."

References

• @$AIWG_ROOT/src/cli/handlers/subcommands.ts — Reproducibility validate command handler
• @$AIWG_ROOT/docs/cli-reference.md — CLI reference
• @$AIWG_ROOT/agentic/code/addons/aiwg-utils/skills/execution-mode/SKILL.md — Set execution mode to reduce variance before validating
• @$AIWG_ROOT/agentic/code/addons/aiwg-utils/skills/snapshot/SKILL.md — Capture state before running validation