Local decision-maker translating reusable tools into platform action
Tool adopter / platform operator
The tool adopter or platform operator is the organisation that chooses and runs a safety tool in its own service. This abstract role separates the publisher of reusable software or policy guidance from the party configuring rules, connecting data and deciding what happens to users or content.
Where practical authority sits
ROOST's integration guide describes choices about models, hosting, local policy and human review. Open licences permit modification subject to their terms, and published governance recognises forking as a check on upstream project control. These permissions do not supply platform credentials or restricted reference data. The Phoebe example likewise requires an operator's endpoints and repository configuration. Pathways support carries access and confidentiality conditions while expressly leaving implementation discretionary. A named deployment and its configuration would be needed to attribute a particular moderation action to this role.
offers removal and escalation language for voluntary adaptation
Policy labels eight invalid categories, normally requires two indicators for credible-threat TVEC2 and describes law-enforcement escalation plus contextual EDSA/human review. This is proposed policy text in a reusable pack, not an operating legal duty or observed removal.
offers no-cost support after eligibility and registration review
Current programme targets company employees handling user content/communications, excludes current member staff and vendors without that role, and reviews applicants. Coalition membership is neither required nor guaranteed. No registration or acceptance event observed.
ROOST Model Community (RMC) → Tool adopter / platform operator
describes configurable model and human-review choices
Guide recommends detection/investigation/review/enforcement separation, comparison with existing policy and human review for uncertain cross-model results. Adopter chooses policy, model/hosting and integration; recommendation is not enforced architecture.
retains an unresolved category reference and contextual tension
INCITE1 is referenced but undefined in inspected policy line 104; EDSA exclusion and otherwise-TVEC permission/evaluation language need reconciliation before treating this as an executable consistent specification. Not evidence of actual over-removal.
RMC CC BY 4.0 licence → Tool adopter / platform operator
permits sharing and adaptation with attribution conditions
Licence requires applicable attribution/notices and identification of changes, disclaims endorsement, and provides termination/cure terms. It does not require adopters to accept CCF updates or give CCF final authority over local content.
requires accurate registration and information cooperation
Terms address information/feedback requests and reasonable efforts to respond to information requests within three months. The duties do not grant an right to a platform user database or amount to blanket assignment of feedback IP.
grants a limited revocable site licence and restricts copying
Licence is personal, nonexclusive, nontransferable and internal/noncommercial within the stated site-material scope. It is not an assignment of participant inventions or universal permission to all linked third-party resources.
supplies contextual actor guidance with explicit limitations
Annex 2 combines actor, content and behavior, discusses politicized/incomplete lists and uses selected incident context. Its examples are not legal designation findings or evidence that a named platform adopted those labels.
Open-source model permits independent forks subject to the relevant licence. Published project rules do not compel downstream operators to deploy an upstream change.
Licence requires applicable attribution/notices and identification of changes, disclaims endorsement, and provides termination/cure terms. It does not require adopters to accept CCF updates or give CCF final authority over local content.
Policy labels eight invalid categories, normally requires two indicators for credible-threat TVEC2 and describes law-enforcement escalation plus contextual EDSA/human review. This is proposed policy text in a reusable pack, not an operating legal duty or observed removal.
Annex 2 combines actor, content and behavior, discusses politicized/incomplete lists and uses selected incident context. Its examples are not legal designation findings or evidence that a named platform adopted those labels.
170 team-labelled items yield TP114/FN12/FP0/TN44 and reported 90.5% recall/100% precision. Recruitment 12/17 fails the stated 85% target; EDSA exact-category precision is distinct from binary valid/invalid errors; TVEC4 11/13 rounds to 85%. No general precision guarantee or observed deployed model.
INCITE1 is referenced but undefined in inspected policy line 104; EDSA exclusion and otherwise-TVEC permission/evaluation language need reconciliation before treating this as an executable consistent specification. Not evidence of actual over-removal.
Guide recommends detection/investigation/review/enforcement separation, comparison with existing policy and human review for uncertain cross-model results. Adopter chooses policy, model/hosting and integration; recommendation is not enforced architecture.
Pinned defaults include a model setting and alternate backends; operator supplies service/repository settings. Adapter obtains action/feature/label/UDF metadata from the configured server, without a hardwired CCF taxonomy.
Both methods raise NotImplementedError and client remains TODO in this 18-line file. This prototype limit says nothing about all Ozone integrations or whether Bluesky can reverse a real action.
Pinned MIT permission allows software reuse under notice/warranty conditions. It provides no server credentials, platform authority or licence to restricted data.
Current programme targets company employees handling user content/communications, excludes current member staff and vendors without that role, and reviews applicants. Coalition membership is neither required nor guaranteed. No registration or acceptance event observed.
Terms address information/feedback requests and reasonable efforts to respond to information requests within three months. The duties do not grant an right to a platform user database or amount to blanket assignment of feedback IP.