Public release status
We publish the evidence
behind product decisions.
Evaluation protocolSource separation, calibration design and untouched holdout.Last reviewed 6 Sep 2026
Development data, independent calibration data and untouched holdout data are kept separate. Results are reviewed by source family and transformation slice so an aggregate score cannot conceal a weak subgroup. The untouched holdout is not used to choose a model or tune a threshold.
Release criteriaSeparate Beta and production evidence gates.Cascade Beta controlled
The controlled Beta can publish a directional AI-generation assessment when independent model channels agree, while a single channel is labelled as lower confidence. A formal accuracy claim still requires a frozen candidate, transformation robustness and an untouched holdout that passes the stricter production gate.
Known limitationsGenerative edits, screenshots, compression and unseen generators.Open limitations
Screenshots, repeated recompression, subtle generative edits and generator families absent from evaluation can weaken signals. The product reports inconclusive when the available evidence cannot support a defensible conclusion.
Change logDetector, provider and result-language changes affecting reports.Release history
As reviewed on 6 September 2026, pipeline version hybrid-provenance-provider-beta-2026.08.6 combines the self-hosted model ensemble with Hive. Earlier commercial candidates remain excluded after their recorded evaluations. Deployment controls remain fail-closed and each report records its pipeline version.
Evidence publication status
What can be independently checked today?
This table distinguishes implemented product facts from evidence that is deliberately withheld until its release gate passes.
| Item | Current public status | Interpretation |
|---|---|---|
| Processing scope | Published | Supported image types, selected-evidence boundaries and provider handling are documented in Methodology and Privacy. |
| Release path | Controlled cascade Beta | Independent model channels remain visible; one-channel output is labelled as lower confidence. |
| Formal accuracy percentage | Not published | No production accuracy claim is made before a frozen candidate passes calibration, robustness and untouched-holdout gates. |
| Benchmark CSV/JSON | Not yet released | A downloadable result set would be misleading until sample definitions, exclusions, thresholds and subgroup errors pass review together. |
| Known limitations | Published | Screenshots, recompression, generative edits and unseen generator families can weaken or change model signals. |
Current answer
What is available in the public service now?
What Made It? validates supported image files, inspects file structure and metadata, checks available Content Credentials and produces an evidence report. When the controlled detector Beta is enabled, the self-hosted model ensemble and Hive provide separate model channels. Two agreeing independent channels can produce a strong Beta result; one channel produces a lower-confidence review indicator.
This does not turn a model score into proof. Screenshots, repeated compression, generative edits and generators absent from evaluation remain known limitations. A low-score result is phrased as “no strong AI signal,” not “authentic.” What Made It? does not publish a formal accuracy percentage until a frozen candidate passes the production release process and the supporting sample definitions and error metrics can be published without overstating performance.
Product and methodology content is maintained by the What Made It? product and methodology team. Processing behavior is described in Privacy, and result interpretation is explained in Methodology.
Release status reviewed by the What Made It? product and methodology team · Last updated