Roadmap & release notes

Multi-team tenancy: a team owns its own annotation schema and its own set of surfaces, and every page that reaches across organizations is admin-only.

• Which sites and AI chats a team may annotate on is a setting now, not a briefing: an admin ticks the 13 surfaces per organization and per team, so a team scoped to one research question — or an agreement that doesn't cover AI-chat transcripts — is enforced rather than explained.
• The layers intersect. An organization's list is a ceiling its teams narrow but never widen, and a surface the organization has turned off shows as unavailable on the team's row rather than as a box that can be ticked to no effect.
• A team with no list of its own inherits. Enforcement runs in the page (no Annotate button on a surface the team doesn't hold) and again when the submission arrives, since the first check runs on the annotator's own machine.
• Nothing is restricted out of the box — the shipped policy is empty, so every team keeps all 13 surfaces until an admin unticks something.

• Each team assigns its own annotation schema. Resolution runs team → organization → default, so re-pointing one team leaves its siblings untouched.
• An acting-team switcher: the team you are annotating for decides the schema, the destination and the meters, while what you can SEE still spans every team you belong to.
• Cross-tenant pages are admin-only. A lead's oversight covers every team inside their own organization — that is depth, not a reach into another organization's names, rosters or totals.
• The organizations page is one collapsible table of organizations and their teams, with drag-and-drop between them over the same confirmed move the buttons use.
• Every move confirms against a dry run that NAMES the people it re-files: existing annotations stay where they are, only future work moves, and a vague “are you sure?” never conveyed that.
• Moves now warn before they break. A member missing from the target organization, or a team schema that organization hasn't published, is reported up front instead of surfacing later as a failed submission.
• The teams page shows every team in your organization; a team you're not on renders read-only rather than as a form that would be refused on save.
• A platforms page sets which sites and AI chats each organization and team may annotate on. A team can narrow what its organization allows and never widen it, so an organization-level limit is a real ceiling rather than a suggestion.
• Platforms a team isn't cleared for simply don't show the annotation tools, and a submission from one is refused at the service too — the policy holds whether or not the browser honours it.

• Custom schemas match the built-in panel: re-opening an annotation now pre-fills the answers you already gave. Re-submitting overwrites the previous annotation, so an editor that opened blank was quietly replacing finished work — this was the release's most important fix.
• The flag pill, the thread-context box (on a chat: the opening request and the eliciting prompt) and the AI explanation all reached custom schemas, so a team on its own schema stops judging a response with no sight of the question.
• Annotation schemas are versioned, so an answer always renders against the definition it was asked under, however often the schema is edited afterwards.

The platform set nearly doubled, and the annotation panel became configurable per organization instead of fixed in the code.

Annotable after this release — 13 surfaces: X, YouTube, Rumble, TikTok, Reddit, Instagram, 4chan, and ChatGPT, Claude, Grok, Gemini, DeepSeek, Qwen.

• Rumble and TikTok — both the video and its comments.
• Three more assistants: Gemini, DeepSeek and Qwen.
• Reddit submissions themselves became annotable, with the submission carried as context for its comments.

• Flag and Annotate moved into the chat's own action row, so the buttons sit where every other control on that response sits — and stopped landing in the middle of a long answer.
• A highlighted passage now anchors to that passage. A partial highlight used to be filed against the whole response, which turned a precise piece of evidence into a vague one.
• The editor opens centered as a modal on extension pages, where there is no post behind it for a side drawer to leave visible.
• On Qwen, Annotate and Flag join the platform's own selection menu instead of fighting it for the same corner of the screen.

• Citations to watchlisted outlets are marked inline in an assistant's response, with the reason a click away.
• Anyone can report an unlisted outlet; leads triage reports into the watchlist every annotator's extension fetches, with no rebuild.
• Detection only — a flagged citation is never an automatic hate label.

• A schema builder with visual, JSON and layout tabs, and a live preview that renders through the same renderer the annotators use.
• Custom HTML and CSS layouts with field placeholders, validated the same way on the client and the server so a layout can't pass one and fail the other.
• Starting-point presets, so a new schema begins from a working example.

• AI use is metered per organization, team and person, with monthly limits and a Usage & limits page.
• Every account starts with 5 pre-annotations and 5 explanations a month — a default allowance rather than an unbounded one.

• A security review pass: hardened sign-in and queue authorization, HTTPS-only ingest, rate and cost limits, and baseline security headers on every response.
• The normalized annotation database became the authoritative store for every submission.
• Every network call carries a timeout. A connection that is accepted and then never answered used to hang forever, which could leave the extension looking dead on the page.

• Platform brand logos on every queue and history row.
• One shared token set behind the whole interface, so the popup, the panel and the workspace stop drifting apart.

Five more surfaces, and adding a colleague stopped requiring a shared password or a redeploy.

Annotable after this release — 8 surfaces: X, YouTube, Reddit, Instagram, 4chan, and ChatGPT, Claude, Grok.

• Grok conversations.
• Reddit and Instagram comments.
• 4chan threads.

• An admin can invite a colleague from the console: a passwordless account with its permission group, team and organization already set, claimed by signing in with Google.
• The accounts roster groups by team, permission group or role, collapses, and searches — a flat list had stopped working past about thirty accounts.

• Post identity is stable on YouTube, so the badge marking an already-annotated comment (and the flag pill beside it) no longer disappears and reappears as the page re-renders.
• A flag placed on a permalink is recognized wherever that post appears.

• The extension asks for no site-access permission beyond the ingest service itself. Every request it makes goes there and nowhere else, so the platform permissions it used to request bought nothing.

The first release: annotate in place on the platform itself, with a lead's approval standing between a submission and the corpus.

Annotable after this release — 3 surfaces: X, YouTube, and ChatGPT and Claude conversations.

• Hover any post or comment → Annotate (or Alt+A) → classify in a drawer beside the post, against the AddressHate label space. No copy-pasting into a separate tool, and the post stays on screen.
• The thread comes with it: reply chains, quoted posts, the parent comment, the video or conversation the post sits in, and its media.
• Keyboard navigation between posts, contextual co-text picking for the comments that make a post legible, and a five-second undo on every submit.
• ChatGPT and Claude conversations are annotable as well — an assistant turn, or a selected excerpt of one, with the prompt that elicited it captured as context.

• Submissions queue locally and deliver with retry, so annotating through a dead connection, a slept laptop or a browser restart keeps the work.
• The editor stays on screen until a submission is acknowledged, rather than closing hopefully.

• An annotation is archived immediately but stays pending until a lead approves it; approval is what commits it to the database and runs the AI second opinion.
• Leads get an approval inbox with the captured evidence, a conflicts view where annotators disagree on the same post, and coaching feedback that reaches the annotator on their own history rows.

• Flag a post for someone else to annotate, with the passage that prompted it highlighted.
• Flagging inside a private conversation is consent-gated, and the evidence is only ever readable in the workspace — nobody else can open someone's chat by its URL.

• Username/password or Google sign-in, with capabilities coming from permission groups — admin, lead, senior annotator, annotator, flagger — resolved live, so a change to a group applies immediately.
• Teams as the isolation unit: two teams do not see each other's flags or annotations.
• A one-time onboarding agreement before first use.

• A full-page Items dashboard over the shared work queue: directories, paste-a-list to create, assignment and filtering.
• History with gate status per row, date and project filters, and export to .xlsx.

• An optional model pre-annotation and a plain-language explanation beside the annotator's own answer, scored for agreement — a second opinion, never a substitute for the annotator's.

• The ingest service on Cloud Run with a durable archive of every submission envelope.
• These public pages — privacy, data deletion, terms and support.