// Reading the shelf: Redline Fieldbook Atrium v0.9 · 2026-04
Documentation / Atrium · Steward /Steward How-to · 11 min

The Silo Steward queue.

L2, L3, and L4 disclosure assets for documents classified above Internal route through a Silo Steward. Mandatory human review, SLA per classification, side-by-side editing, cascade-aware.

TypeHow-to
Reading11 min
Revisedv0.9 · 2026-04
Applies toAtrium · all tenants

/ 01What the steward sees

Each silo has at least one designated steward — typically a Knowledge Management lawyer or senior fee-earner. They see:

  • Items pending review, sorted by classification desc, ingest-age asc
  • SLA visibility per priority class (countdown to breach)
  • Side-by-side source extracted text + generated asset (facet / abstract / redacted view)
  • Edit-in-place for L3 and L4 assets
  • Approve · Reject · Escalate-to-author actions
  • Bulk approve for low-risk batches
  • Filter by author, classification, source DMS, age

/ 02SLA per classification

ClassificationTargetAction if missed
Privileged4 hoursAlert tenant admin; surface on dashboard
Restricted24 hoursAlert silo admin
Confidential72 hoursAlert silo admin (warning)
Internal7 days (5% sample)Audit-only
PublicNone

SLA windows count from ingest time. Pausing a queue item (e.g. awaiting author response) freezes the SLA clock; resuming restarts it.

/ 03Side-by-side review

The review UI shows three things at once:

  • Source — extracted text with span highlighting where the generated asset drew from
  • Asset — the L3 abstract or L4 redacted view, editable in place
  • Metadata — classification, jurisdiction, owning silo, author, AI prompt version, model version

Stewards approve as-is, edit and approve, reject with reason, or escalate-to-author. Edits are tracked: the original generated asset and the steward's revision are both preserved (see asset versioning).

/ 04Reclassification cascade

When a steward reclassifies a document, the consequences cascade through every related artefact and access. The cascade is the heaviest single operation in Atrium.

  1. Conflict pre-check

    If the reclassification would expose across an ethical wall, the operation is blocked and routed to group compliance.

  2. Asset regeneration

    L2, L3, L4 are regenerated against the new ceiling. L3/L4 above Internal queue for mandatory re-review.

  3. Index updates

    OpenSearch, pgvector, AGE, and Postgres are all updated to reflect the new classification.

  4. Prior-access scan

    The audit log is scanned for prior access to this resource. Accesses above the new ceiling become over-disclosure incidents.

  5. Grant revocation

    Existing access grants above the new ceiling are revoked. Grant holders notified.

  6. Claim cascade

    Claims sourced from this document inherit the new classification by default (per-claim overrides preserved).

Over-disclosure incidents at severity ≥ medium trigger a post-incident review. If personal data is involved, the GDPR Art. 33 72-hour clock starts automatically.

/ 05Bulk operations

Bulk reclassification (> 100 documents) requires group-compliance sign-off and a written reason. The bulk engine is throttled, resumable, and observable — the same machinery as backfill.

Per-item review still runs; the bulk operation merely schedules them in order. Cancel mid-flight: the in-progress batch completes, the rest is undone.

Read carefully. Then begin.

Request access Back to documentation