Authenticity AI: Methodology
This page explains exactly how Authenticity AI builds a verdict, so you can audit its conclusions. Nothing here is hidden.
1. Input requirements
The model is shown up to six images, labelled positionally as Front, Back, Label/Tag, then Detail 1, 2, 3…. The label/tag photo is the highest-value input: it carries the manufacturer label, wash label, jock tag, and (often) the product code.
2. What the model checks
Each verdict is built from a fixed set of canonical markers, including product_code_verified, jock_tag_present, three_stripe_quality, and ~40 others. Each marker has a result (pass / warning / fail / not_checked) and a severity:
- Critical: strong fake indicators (product code mismatch, missing jock tag on a kit that should have one, wrong manufacturer logo, broken hologram). A single critical fail can flip a positive verdict to suspect.
- Minor: soft signals (wear consistency, packaging, secondary labels). On their own these don't reverse a verdict.
3. Confidence caps
The confidence number is bounded by the evidence you provide, not by how convincing the shirt looks:
| Condition | Max confidence |
|---|---|
| Only one photo supplied | 40 |
| No label/tag photo | 65 |
| No product code recorded | 75 |
| All evidence present | 95 (hard ceiling) |
We never return 100. There is always residual uncertainty in a photo-only check.
4. Verdict ladder
- verified: all three core photos, every checked marker passes, product code confirmed.
- likely_authentic: multiple passing markers, no confirmed fails, one or two checks inconclusive.
- uncertain (default): insufficient evidence to form a view.
- suspect: at least one fail, or a single critical fail.
- counterfeit: four or more fails, two critical fails, or a known-fake indicator confirmed.
5. Known limitations
- Photo quality matters. Out-of-focus close-ups produce more uncertain markers.
- Vintage kits (pre-2000) have thinner training data. Verdicts skew conservative.
- Player Issue and Match Worn variants share most authenticity markers with retail; AI cannot prove provenance, only authenticity of construction.
- A verified verdict is not insurance. It is a strong photo-based signal, not a chain-of-custody certificate.
6. Effect on Prestige
Only verdicts of verified or likely_authentic at ≥ 80% confidence contribute +0.5 prestige points to qualifying shirts (Match Worn / Match Issued / Player Issue / Signed). A suspect or counterfeit verdict zeroes out a shirt's prestige regardless of version. How Prestige is calculated