Model Substitution in the Black-Box LLM API Resale Market

What was measured

The study combined protocol, behavioral, sub-model, and verdict-fusion signals to test whether a black-box endpoint behaved consistently with its advertised model. Ten endpoints produced at least one substitution signal; two met the stricter repeated-observation rule.

Methodology

1. Protocol and surface fingerprint extraction.
2. Independent behavioral family signals.
3. Deterministic sub-model consistency checks.
4. Fusion across evidence channels before classification.

Limitations

• Black-box inference cannot directly inspect upstream routing.
• Low-sample endpoints are excluded from the persistent-signal estimate.
• Endpoint behavior during test traffic may differ from normal traffic.
• The measurement ended on 2026-04-25; current behavior may differ.
• The public brief anonymizes endpoint identities and makes no legal claims.

Research ethics and provenance

Measurements used legitimately purchased access and ordinary API requests. No denial of service, billing bypass, or unauthorized access was attempted. The public evidence trail includes the Probe explorer, model and relay datasets, and the open-source detection engine.