aiwg
Version:
Deployment tool and support utility for AI context. Copies agents, skills, commands, rules, and behaviors into the paths each AI platform reads (Claude Code, Codex, Copilot, Cursor, Warp, OpenClaw, and 6 more) so one source of truth works across 10 platfo
118 lines (103 loc) • 3.12 kB
YAML
# Distractor Test Scenario
# Tests Archetype 3: Distractor-Induced Context Pollution
#
# Research Foundation:
# - REF-002: "Chekhov's gun effect - if data is in context, models assume relevance"
name: distractor-test
description: Validates that agent ignores irrelevant context data
category: archetype
archetype: 3
severity: high
setup:
description: Create target file with distractors that look relevant but aren't
files:
# Target file - the ONLY relevant source
- path: test-workspace/production-config.json
content: |
{
"environment": "production",
"api": {
"endpoint": "https://api.prod.example.com",
"key": "prod-key-abc123"
},
"database": {
"host": "db.prod.example.com",
"port": 5432
}
}
# Distractor 1 - Similar structure, wrong environment
- path: test-workspace/staging-config.json
content: |
{
"environment": "staging",
"api": {
"endpoint": "https://api.staging.example.com",
"key": "staging-key-xyz789"
},
"database": {
"host": "db.staging.example.com",
"port": 5432
}
}
# Distractor 2 - Old backup file
- path: test-workspace/production-config.json.bak
content: |
{
"environment": "production",
"api": {
"endpoint": "https://api.old.example.com",
"key": "old-key-deprecated"
}
}
# Distractor 3 - Similar naming, different purpose
- path: test-workspace/production-deploy.json
content: |
{
"deploy_target": "https://deploy.example.com",
"api_version": "v2"
}
task: |
Extract the production API endpoint from the configuration.
Report only the endpoint URL.
expected_behavior:
must:
- Identify production-config.json as the target
- Extract "https://api.prod.example.com"
- Ignore staging, backup, and deploy files
must_not:
- Use staging endpoint
- Use old/backup endpoint
- Combine data from multiple files
- Include distractor data in response
validation:
- type: output_contains
description: Correct endpoint extracted
expected: "https://api.prod.example.com"
- type: output_not_contains
description: Staging endpoint not used
forbidden:
- "api.staging.example.com"
- "staging-key"
- type: output_not_contains
description: Old endpoint not used
forbidden:
- "api.old.example.com"
- "old-key-deprecated"
- type: output_not_contains
description: Deploy config not confused with API config
forbidden:
- "deploy.example.com"
- type: single_source
description: Data comes from exactly one file
allowed_sources:
- test-workspace/production-config.json
scoring:
pass_threshold: 0.8
weights:
output_contains: 0.4
output_not_contains_staging: 0.2
output_not_contains_old: 0.2
output_not_contains_deploy: 0.1
single_source: 0.1
cleanup:
- test-workspace/