# Training — Moderators

> Moderators staff the `queue_moderation` lane: triage user-generated content,
> message-center abuse, persona disputes, and reactive flags against published
> content. Distinct from reviewers (who certify content for publication) and
> from support (who handle individual customer tickets).

## Audience and prerequisites

- Audience: trust-and-safety moderation staff and contracted partners staffed
  against the moderation rota.
- Prerequisites: T&S onboarding (V1-TS-onboarding), policy-foundations module,
  well-being check-in protocol, RBAC and identity-verification current.
- Refresh cadence: full re-cert every 6 months because policy evolves; monthly
  micro-drill of the active policy delta.

## Learning objectives

By certification, a moderator can independently:

1. Apply the canonical content-policy classes against an unfamiliar item and
   record the decision with audit-grade evidence.
2. Recognize and route crisis-tagged items into the senior-reviewer lane within
   the documented SLA.
3. Use the moderation tooling: queue navigation, bulk actions inside the
   per-class allowlist, evidence-attachment, decision logging, appeal gating.
4. Recognize manipulation patterns: prompt injection, brigading, scraping,
   evasion via persona-impersonation, watermark removal/reapplication.
5. Coordinate with reviewers when a flagged item touches certified evidence or
   persona-bound content.
6. Apply well-being protocols when content load exceeds personal thresholds;
   rotate before disposition quality degrades.

## Curriculum modules

| #   | Module                                                     | Duration | Format                                   | Assessment                                               |
| --- | ---------------------------------------------------------- | -------- | ---------------------------------------- | -------------------------------------------------------- |
| 1   | Policy foundations and canonical classes                   | 180 min  | seminar + class-by-class card decks      | 30-item policy classification (≥ 28/30)                  |
| 2   | Moderation tool walkthrough                                | 60 min   | hands-on against staging queue           | tool-fluency observation                                 |
| 3   | Evidence handling and audit-grade decisions                | 90 min   | guided decisions + peer review           | rubric pass on 10 decisions with audit packet            |
| 4   | Crisis routing (joint with support, reviewers)             | 120 min  | scenario rehearsal + senior shadow       | 4 crisis cases routed within SLA                         |
| 5   | Manipulation patterns and adversarial content              | 120 min  | curated adversarial set + red-team drill | adversarial-classification test (≥ 85% precision/recall) |
| 6   | Persona-impersonation and identity disputes (joint w/ T&S) | 90 min   | scenario rehearsal                       | 3-case persona-dispute decision pass                     |
| 7   | Appeals coordination and rollback awareness                | 60 min   | walkthrough of appeals lane              | 2-case appeals-coordination drill                        |
| 8   | Well-being protocol                                        | 60 min   | guided + clinician-led discussion        | self-report adherence + supervisor sign-off              |

## Decision rubric

Every moderation decision is graded on:

- **Policy fit**: the chosen class is the correct, narrowest applicable class.
- **Evidence**: the decision packet contains the items needed for an appeals
  review (the item, the surrounding context, the policy line applied, prior
  decisions on the same actor, the operator's reasoning).
- **Action proportionality**: the action (warn, remove, hide, suspend) is
  proportional to the harm and the actor's history.
- **Customer-visible response**: the templated notice fired carries the correct
  reason code and appeals link.

## Scenario rehearsal

1. **Crisis content with grounded source**: a user posts a self-harm request
   alongside a quoted Veritas claim. Moderator routes to crisis lane, suspends
   the item, escalates to senior reviewer for the grounded element, and ensures
   the message-center crisis-routing path runs.
2. **Persona impersonation**: a third party claims a Lilith-controlled persona
   is misrepresenting them. Moderator captures the dispute, suspends
   customer-visible publication on the contested persona surface, routes to the
   persona-dispute lane.
3. **Coordinated brigade against a Veritas claim**: detect coordinated reports,
   separate good-faith reports from brigade noise, escalate abuse to security
   with the cohort packet.
4. **Watermark-evasion suspicion**: a redistributed asset surfaces with the
   Oshun watermark stripped or reapplied. Moderator routes to the
   watermark-verification lane, freezes promotion if the source is still live.
5. **Bulk action under policy allowlist**: hold a 200-item bulk decision on a
   clear policy class; verify per-item evidence; perform the bulk action with
   the documented authorization step.

## Certification criteria

- All 8 modules complete with passing assessment.
- All 5 scenarios passed; failures require remediation, 30-day re-test, and
  shadow shifts with a senior moderator before re-certification.
- 40 hours of shadow shifts with a certified moderator.
- 20 hours of live-supervised disposition with rubric pass.
- T&S lead and well-being supervisor sign-off.

## Tabletop drills (post-certification)

- Monthly: active policy-delta micro-drill (new classes, new edge cases).
- Quarterly: moderation-surge drill alongside the support team.
- Semiannually: coordinated abuse-wave drill alongside security.

## Well-being protocol

- Mandatory rotation off content load every 90 minutes for high-exposure classes
  (graphic, abuse, minors).
- Confidential clinician access on-call during all moderation shifts.
- Self-report tooling for elevated load; supervisors enforce rotation even when
  self-report is silent.
- Quarterly well-being review; mandatory time-off compliance audited by T&S
  lead.

## Owner

The T&S lead owns this training program. Privacy lead, security lead, and a
contracted well-being clinician review the well-being and adversarial-content
modules. Updates are version-controlled and distributed via the policy-delta
micro-drill.