Governed campaign artifacts

Artifacts turn campaign content into durable business objects instead of leaving it inside a model session. Open GTM → Artifacts to view campaign briefs, email sequences, landing-page copy, social posts, evaluation reports, and custom versioned kinds.

Select Generate with AI to start a tenant-scoped authoring thread. The assistant reads the active campaign, active knowledge, and compact existing-artifact index before proposing exact Markdown. Saving the proposal requires human approval and creates only a working copy through Temporal and Fabric Platform. Proposal, evaluation, approval, publication, and delivery remain separate governed actions.

Artifact bodies use Markdown. The side panel opens in rendered Preview mode with readable headings, tables, code, and links; Source shows the exact immutable body and Download exports that revision as a .md file. New working copies provide matching Write and Preview modes before any governed mutation is submitted.

Find the work that needs attention

The artifact workspace separates lifecycle states into tabs. The count beside each tab is the number of artifacts in that state; color reinforces the label but never replaces it.

The Campaigns journey opens the exact artifact that needs attention and preserves it in the URL, so you can share or resume that review without searching again. Closing the side panel returns to the same status tab.

Understand repeated content titles

The same reviewed content can appear in more than one campaign version without being a duplicate. A safe-test version and its separately governed live candidate intentionally preserve the source content title and immutable body, while their campaign bindings remain different. The workspace therefore adds the authoritative mode and version to every displayed title:

Citi migration readiness brief — Dry run v13
Citi migration readiness brief — Live candidate v14

Dry run identifies content bound to a non-delivering safe-test version. Live candidate identifies content copied into a separate live-intent version that still needs its own evaluation, approval, locked manifest, and explicit promotion. The status label answers a different question: Proposed means the exact revision is waiting for review, while Approved means that exact campaign-bound revision passed. Search, downloads, account context, agent selectors, and approval tasks use the same version-aware labels. If campaign metadata is temporarily unavailable, GTM Brain shows the known version without guessing a mode.

Lifecycle

working copy → proposed immutable revision → evaluation → approval → publication request

A working copy is mutable and may autosave. Propose revision creates an immutable checkpoint with a SHA-256 content hash. Evaluations and approvals bind to the exact revision number and hash. A later edit invalidates approval; it never silently carries forward.

The artifact panel shows approval readiness beside Approve exact revision. The button stays locked while the independent check is still running, is missing, was performed only by the revision author, or requested changes. Campaign-scoped artifacts link back to Campaigns, where Send for background review starts or retries the governed reviewer. Keep the artifact panel open and it refreshes the review status for up to 15 minutes; a passing evaluation for the exact revision and hash unlocks approval automatically. A failed action is also reported in a toast and beside the control, so recovery never depends on finding an alert elsewhere in the panel.

Campaign identity, campaign version, artifact kind, and parent lineage become immutable when the artifact is created. An edit can only create content for that stable identity; it cannot silently move an artifact to another campaign or reclassify its kind.

AI-authored working copies must supply both the campaign ID and its immutable campaign version. The governed action rejects a missing or mismatched version, so an agent cannot attach content prepared for one strategy revision to a different campaign execution manifest.

Bind approved content to execution

An active campaign does not consume whichever artifact happens to be newest. After all four required artifacts pass independent evaluation and exact approval, return to Campaigns and select Confirm campaign details on its current journey row. The guided Campaigns workspace and the detailed Strategy workspace use this same label. GTM creates a hashed execution manifest containing exact approved revisions for the campaign brief, audience definition, sequence policy, and reviewer policy, plus the active content-policy hash and verified proof-asset hashes. Every bound artifact must belong to that exact campaign ID and immutable campaign version.

Confirmation freezes evidence; it does not send email. The next visible row action is Run safety check, which can create review drafts but cannot deliver externally. Launch remains a separate controlled action.

The campaign pass verifies the complete manifest before sensing or drafting. The V2 drafter consumes the bound approved sequence and content policy; it cannot fall back to parallel prompt-only copy. If an input is missing, changed, belongs to another campaign/version, lacks the required independent passing evaluation, or no longer matches proof/policy state, execution fails closed and the operator must bind a new manifest.

Publication creates a receipt containing channel, adapter, external identifier or URL, idempotency key, attempt status, and failure information. A publication request is not permission to send: the configured adapter and normal delivery policy still apply.

Human and agent authorship

Human edits and agent output use the same actions:

• gtm.save_artifact_working_copy
• gtm.propose_artifact_revision
• gtm.record_artifact_evaluation
• gtm.decide_artifact_revision
• gtm.publish_artifact
• gtm.archive_artifact

Agent identities are recorded as execution principals bound to tenant, registration, immutable agent version, run, Temporal workflow, capability grant, correlation, and causation. An agent can request only the artifact actions present in its live, certified catalog and exact campaign scope. GTM Brain mints any required approval or publication proof at the trusted server boundary; a caller-supplied proof, actor, command identity, or campaign binding is rejected.

A specifically certified registration may request gtm.decide_artifact_revision or gtm.publish_artifact, but catalog presence alone is not authority. Current tenant policy, grant expiry, campaign scope, exact revision/hash, independent evaluation, and any HITL rule are all rechecked at final execution. External delivery remains separate: internal approved-delivery actions and live promotion are not exposed as general external-agent tools.

Storage and privacy

Artifact bodies are staged in Lakebase before Temporal starts. Workflow history receives only opaque staging IDs and compact revision references. Large media and attachments should use a governed UC Volume with metadata references in Lakebase. Private bodies are excluded from Delta projections by default.

Treat retrieved and imported content as untrusted. Evidence carries provenance and prompt-injection labels; agent prompts explicitly delimit artifact content as data, never instructions.

Failure behavior

• A body whose SHA-256 does not match the supplied hash is rejected.
• A stale base revision or changed working copy fails closed.
• Evaluator identity is derived from the authenticated person or validated agent execution principal; callers cannot name another evaluator, and an evaluator cannot review its own revision.
• Approval for a different revision/hash is rejected.
• Approval requires an authorized admin, an independent passing evaluation, and a short-lived server-signed capability proof bound to the tenant, actor, action, and exact command. Publication uses the same rule.
• An agent publication request waits for approval and never bypasses delivery policy.

Next: Run your first campaign, Agent operating guide, and Governance.