Open source ≠ transparent ≠ co-created
Dumping the code online isn't the same as transparency: ordinary people simply can't read it. You also need documentation, examples, explanations, and an appeals channel (which is exactly why this site exists). Opening up an API isn't the same as co-creation either — as the veteran user Denken once lamented, open source often ends up being just you, on your own.
“Open source never really led to co-creation… in the end I'd often find it was just me.”
Why Matters dares to open source
Big platforms don't dare open-source their systems, because the moment they do, people exploit the loopholes to slip past them. Matters' reasoning is a little counterintuitive: “We're already small enough that open-sourcing brings back more feedback than it costs — a bit more openness does no harm.”
Layered open source
The code is public, but sensitive parameters that could be directly abused are held back for six to twelve months, or abstracted first. A value like the spam-detection threshold (spamThreshold) is a sensitive operational asset. There's a certain resignation to it: keeping a censorship-resistant platform alive often means sacrificing some transparency for the sake of security.
“Keeping a censorship-resistant platform alive often means sacrificing transparency for security.”
Releasing the static spam dataset
matters-spam-dataset is a de-identified research dataset: identities are processed with HMAC A cryptographic method that turns identity information into an irreversible code, used to de-identify and protect personal data. and contact details are normalized. It's planned for Hugging Face (gated access) under a CC BY-NC-ND 4.0 license, with a first release of 425,689 records (406,850 spam / 18,839 legitimate). The dataset card openly discloses flaws in the early proxy labels, as well as a residual 5–15% mislabeling rate. ⚠️ It has not been officially released yet.
Why the whole moderation method should be opened up
“Transparent moderation,” put plainly, means laying out the entire moderation method rather than just dumping the code: the detection models, the enforcement workflow, and the training datasets, all released together. This isn't about sounding noble. The reality is that bots flood the platform with content while humans scrape by in the cracks; that's precisely why the whole algorithm and training data — along with the cases wrongly removed as false positives — all need to be laid bare.
The first thing transparency does is make moderation auditable. Many large platforms have drawn criticism lately precisely because the users they wrongly removed, blocked, or suspended (some of them paying customers) had no idea why it happened. Only when the models, how the thresholds decide, and the misjudged cases are all laid out can outsiders check whether “the line was drawn in the right place,” and can users know how to appeal. Ultimately, transparency is the very precondition for due process.
The second is letting other platforms use this method directly. The public cleanup records built up by the neighborhood-watch team and the de-identified spam dataset can both serve as front-line anti-fraud examples, so other small, local, and marginal communities can follow along instead of each platform figuring it out from scratch. This is exactly why opening the box is worth it — the payback far outweighs the risk of being gamed.
License
The code is licensed under MIT / AGPL, the documentation under CC BY 4.0, and the research data is licensed by sensitivity and de-identified. Every interactive element also comes with a “text-version data table” alongside it — partly to support accessible reading, and partly to serve as open data.
Further reading: layered disclosure and the open-dataset methodology