# Training — Reviewers

> Reviewers certify content for publication: source binding, policy gates,
> approval, rollback, change-note authoring. Distinct from moderators (who
> triage published-or-shared user content) and from model operators (who certify
> model versions).

## Audience and prerequisites

- Audience: editorial and trust reviewers responsible for the studio release
  gate.
- Prerequisites: editorial-foundations module, T&S moderator awareness module,
  RBAC current, ADR-0006 audit-scope familiarity.
- Refresh cadence: 6 months; per-domain micro-drill monthly.

## Learning objectives

By certification, a reviewer can independently:

1. Bind a publishable item to its source set such that every claim has
   audit-grade evidence and the Sophia retrieval round-trip is verifiable.
2. Apply the canonical policy gates (accuracy, attribution, persona policy,
   rights, accessibility, locale, AI disclosure) and record a structured
   decision.
3. Use the studio publish gate, rollback hook, and public change-note authoring
   tools.
4. Coordinate with persona, model, T&S, rights, and accessibility leads when a
   review decision crosses domain boundaries.
5. Author a customer-visible change note that preserves trust during rollbacks
   and correction cascades.
6. Recognize the publish-freeze flags (provenance, watermark, moderation) and
   respect them without exception.

## Curriculum modules

| #   | Module                                                                 | Duration | Format                         | Assessment                                     |
| --- | ---------------------------------------------------------------------- | -------- | ------------------------------ | ---------------------------------------------- |
| 1   | Editorial foundations + studio walkthrough                             | 90 min   | guided demo                    | studio-fluency observation                     |
| 2   | Source binding and Sophia round-trip                                   | 120 min  | hands-on with fixture sources  | 10-item binding test (≥ 9/10 round-trip valid) |
| 3   | Policy gates and structured decisions                                  | 180 min  | seminar + class-by-class cards | 20-item gate-application test (rubric pass)    |
| 4   | Rollback and correction cascades                                       | 90 min   | scenario rehearsal             | 3-case rollback drill                          |
| 5   | Change-note authoring (customer-visible)                               | 90 min   | guided writing + peer review   | 5 change notes pass rubric                     |
| 6   | Cross-domain coordination (persona, model, T&S, rights, accessibility) | 90 min   | scenario rehearsal             | 3-scenario coordination drill                  |
| 7   | AI-disclosure compliance for published items                           | 60 min   | walkthrough + audit            | disclosure-render rubric pass                  |
| 8   | Locale and accessibility gates                                         | 90 min   | hands-on against fixtures      | RTL + long-string + transcript checks pass     |

## Source-binding protocol

Reviewers operate the Sophia source-binding tool as the canonical step:

1. Inspect every claim in the item; mark each with the bound source ID.
2. Run the Sophia retrieval round-trip — every claim must retrieve its bound
   source on a fresh query.
3. Verify the bound source's published status (no retracted or corrected sources
   without the explicit cascade handling).
4. Verify the source's rights status and the disclosure-required flag.
5. Verify localizations: each locale's content must bind to the appropriate
   locale source.
6. Sign the binding bundle and attach it to the publish request.

## Rollback authoring protocol

When a published item must be rolled back:

1. Identify the regressed claim and the failing gate.
2. Coordinate with the originating editor for the corrected version, or pull the
   prior certified version from the studio rollback hook.
3. Author the customer-visible change note using the canonical structure: what
   changed, why, what is now correct, where to read more.
4. Ensure the change note is bound to the audit packet for the incident; the
   public change note must not name internal versions.
5. Trigger the rollback via the studio rollback hook; confirm
   `studio.publish.rolled_back` audit event with reason code.
6. Verify cache invalidation across surfaces (Sophia retrieval cache,
   evidence-rail cache, search index).

## Scenario rehearsal

1. **Bound-source retraction cascade**: an evidence source is retracted
   upstream. Reviewer authors the rollback cascade for every published item
   bound to that source.
2. **Locale-divergent decision**: an item passes gates in `en-US` but fails
   accessibility in `ar` (RTL truncation). Reviewer holds publish on `ar` only,
   coordinates with localization for the fix.
3. **Persona-bound editorial**: a Tara teacher's-lineage piece is reviewed for
   both editorial accuracy and persona-policy alignment. Reviewer coordinates
   with persona ops.
4. **Rights-sensitive media**: a Nisaba edition uses a licensed manuscript scan;
   reviewer verifies rights status, watermark policy, and disclosure copy.
5. **AI-generated illustration**: an Isis-generated illustration is reviewed for
   provenance, watermark, accessibility, and disclosure compliance.

## Certification criteria

- All 8 modules complete with passing assessment.
- All 5 scenarios passed; failures require 30-day re-test plus shadow shifts.
- 40 hours of shadow review with a certified reviewer.
- 20 hours live-supervised review with rubric pass.
- Editorial lead, T&S lead, and rights lead sign-off.

## Tabletop drills

- Monthly: per-domain micro-drill against the latest content delta.
- Quarterly: cross-domain coordination drill (persona, model, T&S, rights,
  accessibility, localization).
- Semiannually: retraction-cascade tabletop with the grounding-failure runbook
  owner.

## Owner

The editorial lead owns this training. Updates require sign-off from T&S lead
(for policy gates), privacy lead (for DSAR-adjacent disclosure gates), and
rights lead (for licensed sources).