01
Moderation Benchmarks
Quarterly latency, throughput, and false-positive numbers from independent deployments — the closest thing the category has to a public scoreboard.
Comment Castle Research Desk · Vol. VI
The Resources Hub
A working library for community managers, editorial leads, and trust & safety teams who treat their comment sections as a publication, not a by-product.
Benchmark reports, integration field notes, toxicity research, and case studies — published by the same linguists and former community managers who built the platform.
Book a DemoEditorial Lanes
Six disciplined categories. Browse by job-to-be-done.
The library is organized around the work community managers actually do — not the funnel stage a marketing team would prefer they sit in.
02
Toxicity Research
Papers and field memos on multilingual toxicity, coded harassment, and the limits of lexicons — drawn from the 41-million-sample corpus behind Toxicity Score v3.
03
Integration Guides
Step-by-step deployment notes for WordPress, Ghost, Drupal, Discourse, and the documented REST/Webhook API used by 600+ engineering teams.
04
Compliance & Privacy
SOC 2 Type II and ISO 27001 posture, on-device GDPR pseudonymization, and editorial workflows for right-to-be-forgotten requests at scale.
05
Engagement Analytics
How leading newsrooms move past comment counts and toward reader retention, repeat commenters, and thread depth as first-class metrics.
06
Case Studies
Long-form deployments: the migration plan, the resistance inside the newsroom, the audit trail, and the moderation cost reduction measured at 90 days.
Featured Publication
The 2024 Comment Moderation Benchmark Report
An annual reference work drawn from 1.2 billion moderated comments across 3,400+ active publisher sites — the most cited independent dataset on moderation performance in the category.
- Pages
- 184
- Contributors
- 11 researchers, 4 linguists
- Methodology
- Double-blind audit of 280 customer migrations
- Published
- October 2024 · Ann Arbor Press
Inside: sub-180ms latency distributions, the 96.4% precision score of Toxicity Score v3 against brand-safe baselines, the 71% average moderation cost reduction measured across the audited cohort, and a chapter on failure modes that no vendor blog will publish.
Request the Print EditionMost Read This Quarter
Five pieces the research desk keeps getting asked about.
-
No. 01
Sub-180ms Moderation: How the Fastest Independent Stack Got That Way
A teardown of the inference path, the queueing layer, and the editorial decision to publish latency as a first-class metric rather than a marketing footnote.
-
No. 02
What 41 Million Labeled Comments Taught Us About Coded Harassment
Why lexicons fail on dog whistles, how the v3 training corpus was balanced across four languages, and what precision means when the cost of a false negative is a journalist's inbox.
-
No. 03
From Comment Counts to Reader Retention: An Analytics Refit
A working memo for editorial leads who want to stop optimizing for volume and start measuring the comment section the way a subscriptions team measures a homepage.
-
No. 04
One-Click Integrations: The Honest Engineering Notes
What "one-click" actually means for WordPress, Ghost, Drupal, and Discourse — and the four integration patterns our engineers wish more vendors documented.
-
No. 05
GDPR Pseudonymization at the Edge: A Production Postmortem
How the first moderation vendor to ship on-device pseudonymization in production earned its way through a German press council's audit — and what it cost in engineering time.
From the Editor
Filed from Austin, Texas · Autumn 2024
A short letter from the desk.
The Comment Castle Research Desk is staffed by seven full-time researchers — three linguists, two former community managers from daily newsrooms, and two engineers who would rather write than ship. We are not a marketing content team. We do not ghost-write for the sales pipeline, and we do not publish under embargo.
Our publishing cadence is deliberate: one long-form research report each autumn, four quarterly benchmark updates, and roughly two field memos a month — published when they are ready, not when the calendar asks. Most pieces are reviewed by at least one external practitioner before they go out. Many are quietly revised six months later when a publisher sends us a better measurement.
What you will find here is the work that informs the platform itself. The same models that score toxicity in production are documented here in enough detail for a community manager to challenge. The same integration patterns our platform engineers ship are written up so a CMS lead can replicate them. If a paper disagrees with our marketing claims, the paper wins — and we revise the marketing.
Reader mail reaches us at [email protected]. We answer every note, usually within a week, occasionally with a footnote in the next issue.
— The Editors
The Next Step
Bring the research into your comment section.
A 20-minute live walkthrough with a member of our solutions team. We will run your last 30 days of comments through the platform on the call, in front of you, and show you what the moderation desk would have done differently.
Book a DemoNo slides. No sales script. No follow-up sequence you didn't ask for.