Scoring methodology
Rubric version 1.0. Any rubric change triggers a full re-scoring of the catalog.
We do not publish opinion reviews. Every listing carries five scores, each with a written justification, a date, and a method label displayed next to it:
- declared — the platform states it publicly;
- observed — we verified it, and link the evidence;
- inferred — deduced from converging signals, and said so.
The five axes
| Axis | Weight | The question it answers |
|---|---|---|
| Credit model | 25% | How many generations for how much money — and what happens when the quota runs out? |
| Censorship transparency | 20% | Are the limits documented and consistent with the marketing? We score transparency, never permissiveness. |
| Persistence | 15% | Does the platform keep state between sessions, and is that scope documented? |
| Data retention | 25% | What happens to what users upload: retention time, deletion, training use and opt-out. |
| Exit terms | 15% | Can you leave easily, and take your content with you? |
The overall score is the weighted average, rounded to 0.1 — recomputed from the five axes, never hand-edited.
What never enters a score
Whether a platform has an affiliate program, how much it pays, whether it bought a placement, and how much traffic it has. Paid placements are labelled Sponsored and change visibility only. This separation is enforced in our code: scoring tables and monetisation tables are never joined.
Freshness
Every listing shows its last successful verification date. A listing not verified within 14 days drops out of the site automatically. The rubric itself is reviewed at least yearly, and whenever an axis stops discriminating.