T&S Workbench

Free, private tools that help Trust & Safety and product teams find risks before launch, rehearse incidents, measure what matters, choose vendors and write better policy.

Open the live site · Explore a demo company, with every tool filled in for a fictional app · no sign-up, and nothing leaves your browser

The tools

Tool	What it helps you do	Open content
Abuse pre-mortem	Profile a product and see how it will be misused before launch: 58 risks, 104 safeguards, 33 legal obligations in 7 jurisdictions.	abuse-premortem
Incident tabletop	Rehearse a crisis: 43 scenarios across 8 company types, with a lesson and the law behind every call.	incident-tabletop
Incident post-mortem	After a real incident, a blameless post-mortem in fifteen minutes: a one-page review in the handbook's shape, a grade for the response (never the incident), and three actions with owners and dates for Jira, Asana, Linear or GitHub.	ts-workbench
Metrics framework	A reference for learning and building your own measurement: 36 metrics with formulas, measurement steps and SQL, plus an optional worksheet for your numbers. Outside the assessment on purpose.	ts-metrics-framework
Vendor scorecard	Choose a moderation vendor on evidence: 8 weighted criteria, rubrics, minimums and RFP questions.	moderation-vendor-scorecard
Coverage radar	Rate five layers of defense for 8 kinds of harm and see them against your products' risk.	harm-coverage-radar
Program maturity	Rate a T&S program in 8 areas against the targets for its stage, and get a phased roadmap for the biggest gaps.	ts-maturity-model
Classifier eval	Build a labeled test set of hard cases from a rule, run a moderation classifier against it, and see precision, recall and where it fails, with what to change. Works with your own model or an open one like gpt-oss-safeguard (runner).	ts-workbench
DSA readiness	Find which duties under the EU Digital Services Act apply to a service, article by article, and turn the gaps into a plan, four drafts and a memo for Legal.	ts-workbench
Grooming pathway review	See how an adult could find, reach, bond with, isolate and exploit a child through a product's features, and what's in the way at each of six stages, with a product brief, a response ladder and starter rules. Nothing describes how grooming is done, only what a platform can see.	ts-workbench
Child-safety reporting	Check the first hour after you find child sexual exploitation against the duties in the US, UK, EU, Canada and Australia: remove, preserve, report, handle the account, protect the people, with four drafts for Legal and Ops.	ts-workbench
COPPA readiness	Find out whether the US COPPA Rule applies, map the children's data you handle, and close the gaps against the amended Rule with drafts ready to edit.	ts-workbench
Policy stress-tester	Paste a rule and find the words reviewers would read differently, the cases it forgets, and how it holds up against real edge cases. Instant checks anywhere; a deeper review on the user's own Claude account.	ts-ai-assistants
Enforcement notice writer	Turn a decision into a notice a person can understand and act on: what happened, which rule, what they can do next. Runs on the user's own Claude account.	ts-ai-assistants
Appeal reviewer	Review an appeal against the original decision and the rule, with a recommendation framed for a human to make the call. Runs on the user's own Claude account.	ts-ai-assistants
Transparency report	Build a DSA transparency report from your numbers, with a completeness check against the articles and the EU templates.	ts-ai-assistants
Red team studio	Test your own AI feature for harm the way real red teams do, one calm step at a time: a target card, tries graded beside an expert rubric, findings an engineer can act on, a week plan, and a one-page summary for the customer who asked whether you red team.	ts-workbench
Learn: red teaming guides	Two guides, for LLMs and for world models (video, image-to-video and interactive worlds): the method one step at a time, drills to run with your team, worksheets and sources.	ts-workbench
Glossary and practice	48 red-teaming terms in plain words, with flashcards and a quiz.	ts-workbench
T&S glossary	103 everyday Trust and Safety terms, from strike to statement of reasons, in plain words.	ts-workbench

Works with ROOST (work in progress)

ROOST is a nonprofit that makes free, open-source trust and safety tools, such as the Coop review console and the Osprey rules engine. It also runs a community for open safety models such as OpenAI's gpt-oss-safeguard. The workbench doesn't use these tools. One part connects: the classifier eval can score labels from an open model you run yourself with the runner script. Two reference files map the metrics to Coop's data and the risk areas to ROOST's harm taxonomy format. It's text only and an early start. See the ROOST page. The workbench is independent and not affiliated with ROOST.

Why it exists

Trust & Safety knowledge mostly lives in people's heads and in slide decks that don't travel between companies. Every new team rebuilds the same thinking from scratch, usually after something has already gone wrong. This project packages that thinking into tools a product manager, a founder or a new T&S hire can use today, and publishes the underlying knowledge as open content.

Design principles

• Private by default. Everything is stored in the visitor's own browser, so teams can describe unreleased products freely. A workspace file carries it to any device: save it to a synced folder and open it anywhere.
• Works with your tracker. Launch plans and maturity roadmaps go to Jira, Asana or Linear as a CSV, or open as pre-filled Jira, Linear or GitHub issues, with no logins or tokens.
• Plain language first. Written for product managers and founders, with jargon explained on hover.
• Every number explains itself. Each risk score lists the answers that drove it; each metric names its guardrail.
• Summary before detail. Reports open with the handful of actions that matter most.
• AI where it helps, with a person in charge. The AI tools run on the visitor's own account, only on a click, and frame every result for human review.
• Works everywhere. Light and dark themes, keyboard navigation and layouts that work at phone width.

Screenshots

How it's built

• One HTML file. No framework, no backend, no accounts. Source parts in src/ are assembled by src/build-ws.js into docs/index.html, which GitHub Pages serves.
• Hash routing gives every tool, tab and metric its own link, and a command palette (⌘K) reaches any tool, metric, scenario or saved result.
• Automated checks render every page for every platform, stage and regulation setting, play every tabletop scenario in every tailored version, verify scoring and status logic, and run each AI assistant against a stubbed model, including declined permission, rate limits and malformed output.

npm test      # builds, then runs every check
npm run build # rebuilds docs/index.html

License and credit

Code is MIT. The knowledge content (risks, safeguards, scenarios, metrics, rubrics and prompts) is CC BY 4.0. Law notes are general information, not legal advice.

Built by Steven Macchia, Trust & Safety leader, with AI-assisted development (Claude).