SHIRO & Co.
Visible World Audit
See what your algorithm chooses to make visible.
People choose from what systems choose to make visible. Visible World Audit records, compares, and preserves evidence of how ranked and personalized exposure changes across users, behaviors, channels, and time.
What it does
- Records ranked and personalized exposure as observed outputs.
- Compares visible worlds across personas, behaviors, channels, and time.
- Preserves a frozen evidence package and a read-only client report.
- Separates Known, Inferred, and Unknown claims.
What it does not do
- proprietary source-code inspection
- hidden model weights
- causal claims without evidence
- protected-trait profiling
- unauthorized scraping
- anti-bot circumvention
- It is not an AI bias detector, algorithm reverse-engineering tool, hidden-weight discovery product, or manipulation detector.
Evidence model
An audit stores observations, recorded interactions, refreshed observations, comparisons, and optional email events. Client delivery uses only a frozen snapshot. Live data after freeze does not change the delivered report.
In scope: ranked exposure · visible item order · before/after changes · persona differences · control differences · email-to-web sequence · structured semantic classification
Provenance
Claims are labeled Observed, Inferred, or Unknown. Causal relationships are not established unless separately demonstrated. Semantic classifications are inferences from observed text.
Visible World Divergence
Visible World Divergence is an experimental internal metric that compares observed exposure-set overlap, rank movement, and category distribution. It is not a validated industry standard. It is not a judgment that a system is good, bad, safe, unsafe, biased, or unbiased.
Authorized use
Observe only systems you own, control, or are authorized to test. Lab / demo is a controlled synthetic environment and is not client evidence.