Using Forge

Your first AI Release

Before you begin, gather the business outcome, intended users, important data or delivery constraints, and measurable success criteria. You do not need to choose a technical pattern first.

1. Describe the business intent. Open New system and explain the outcome, users, constraints, and success criteria. Submit the intake to generate an architecture plan.
2. Review the architecture. Check the recommended approach and its rationale. Choose Request edits when an assumption is wrong. When it is suitable, name and classify the system, then choose Accept & Continue.
3. Let Forge create the records. Accepting the recommendation registers the Enterprise AI System and starts its first AI Release with the intent and architecture attached. You do not create the first release separately.
4. Follow processing on the Release Record. Forge generates and seals the blueprint. If the Architect asks a question, answer it on the record so processing can continue.
5. Review checks and Delivery. Resolve failed checks or evaluation findings. Open Delivery, review the exact AI System Package, Engineering Specifications, and Engineering Plan, then attest to the displayed SHA-256 fingerprint.
6. Submit the release. When the Promotion Gate is enforced, resolve every blocker first. In advisory mode, review the findings and choose either Submit for approval or Submit with findings, as offered.
7. Have another authorized person decide. If separation of duties is enabled, the submitter cannot approve their own release. An owner, admin, or reviewer reviews the Delivery evidence fingerprint, attests to the exact artifacts, then chooses Approve & freeze or Reject.
8. Download the proof. Open the Deliverytab and download an internal or external Audit Package. The external package applies your organization's disclosure policy.

Find your way around

The main navigation is organized by the job you are doing. Use the Dashboard to resume work that needs attention; use the capability pages when you need a broader inventory or history.

Integrations is pinned separately near the account area. Open the account and workspace menu for Knowledge (org) and the Settings Control Center; they are not primary navigation destinations.

Press Ctrl+K on Windows or ⌘K on macOS to search and move directly to a system, release, or destination.

Use the Dashboard as your work queue

Start each session on the Dashboard. It summarizes releases and systems that need action. Open the suggested action instead of searching through every record.

• Continue a release that is waiting for clarification or processing.
• Open a release that is ready for review or awaiting a decision.
• Investigate open Runtime incidents or failed releases.
• Reconcile shadow cloud resources, missing provider bindings, or stale provider inventory.
• Return to an active system whose governance record is stale.

Manage Enterprise AI Systems

An Enterprise AI System is the long-lived portfolio record for an AI capability. It holds the current architecture, ownership, policy, evidence, delivery state, and the history of releases that changed it.

• Overview shows current status, registry details, policy, and accumulated evidence.
• Studio contains the system definition, delivery choices, and release preparation.
• Knowledgemanages the system's approved knowledge sources and retrieval readiness.
• Operate contains Runtime sessions, metrics, inference and agent signals, and delivery infrastructure.

Use Enterprise AI Systems to find a system. Open it and choose New release for a new model, prompt, retrieval, architecture, policy, or delivery change. Reviewers can inspect systems but cannot create releases.

Reconcile systems with cloud resources

Open Systems → Cloud resources to compare provider discovery with the Enterprise AI Systems registered in Forge. Run a scan after connecting a supported cloud account under Integrations.

• Shadow resources were discovered but are not governed by a registered system.
• Missing or unverified items need a fresh scan, corrected access, or an operator decision.
• Matched resources are linked to the system Forge expects.
• Register a discovered resource, bind it to an existing system, or record an exception only when that disposition is accurate.

Retire a system without losing history

Owners and admins can open Overview → Registry and choose Deprecate system. Enter a reason and type the system name exactly to confirm the change.

• The system leaves active use and cannot start new AI Releases.
• AI Releases, Evidence, Audit Packages, architecture records, and audit events remain available.
• Forge-managed resources are deprovisioned by default. Retaining them is an explicit exception selected during deprecation.
• Unfinished releases move to Archived unfinished releases and remain read-only.

Shape the system in Studio

Forge Studio turns business intent into an inspectable system definition. On a system's Studio tab, work from left to right:

1. Define — review the AI System Package, system map, specifications, and architecture rationale.
2. Choose delivery — assess providers and select how the system will be delivered.
3. Prepare release — confirm provider plans and the execution workflow before governing a change.

Record a model decision from measured evidence

An architecture recommendation can shortlist account-available models and show the evidence and unresolved constraints behind that shortlist. After the recommendation is accepted and linked to a system, use Model decision evidence to run a frozen-workload comparison when at least two eligible candidates are available.

1. Inspect hard constraints, capability coverage, benchmark provenance, and unknowns.
2. Choose Run comparison and wait for the candidate results.
3. Select a completed eligible result, record that no candidate is preferred, or reject the comparison. Add the measured rationale and choose Record decision.
4. When an AI Release carries the selected decision, confirm that its Release Record says the decision is bound. Stale, missing, or mismatched decision evidence appears in Checks & approval and can block promotion when policy is enforced.

Later AI Releases inherit the system's accepted business intent. Describe the governed change in the release's requested-change fields. If the underlying system intent or constraints have materially changed, reassess the architecture before treating it as the same system.

Ground the system with Knowledge

Use Knowledge (org)for reusable organization corpora. Use a system's Knowledgepage to choose the documents and index that ground that system's answers and release evaluations.

1. Documents — add and process the approved source material.
2. Index — use Forge storage or bind a supported customer-managed vector or cloud knowledge resource. Select from discovered resources when available; enter a resource manually only when discovery cannot list it.
3. Prove — run retrieval checks and review readiness before relying on the corpus.
4. Answers — inspect grounded answers and their traces for the selected system.

Keep organization corpora reusable, but make the system binding explicit. A corpus existing in the organization does not by itself make it release evidence.

Work from the Release Record

An AI Release is one governed change to an Enterprise AI System. Find releases under AI Releases. The Release Record is the source of truth for its processing state, checks, evidence, review, and delivery package.

During processing:

• Processing means Forge is generating the blueprint. The record updates automatically.
• Action required means the Architect needs clarification. Answer the question on the Overview tab.
• Blueprint sealed means the specification is complete and the release can move through evidence and approval.
• Processing failed means you should review the error and retry when offered, or start a new release if the underlying request must change.

Use the release inventory to find the next action

The AI Releases page can narrow the inventory to releases that need clarification, have a failed or stuck pipeline, await approval, or were archived unfinished. Archived unfinished releases are retained for history and cannot be resumed.

Prove and approve the change

The Promotion Gate combines configured policy with release evidence and evaluation results. A completed blueprint is not, by itself, approval to ship.

1. Open the Release Record's Checks & approval tab.
2. Review automated checks, evaluation results, model-decision evidence, and linked evidence.
3. Resolve every enforced blocker. Re-run an evaluation when the record offers that action. In advisory mode, findings remain visible but do not become decision authority.
4. Open Delivery and review the exact AI System Package, Engineering Specifications, and Engineering Plan. Attest to the displayed SHA-256 fingerprint before submitting.
5. Choose Submit for approval, or Submit with findings when the gate is advisory and the record offers that action.
6. An owner, admin, or reviewer opens the same tab, reviews and attests to the Delivery fingerprint, then chooses Approve & freeze or Reject. Approval freezes the current engineering plan and release evidence in the approval snapshot; it is not legal certification or deployment authorization. A rejection requires a reason so the release owner knows what to change.

Keep and share proof with Passport

Forge Passport connects release evidence, approval history, and portable Audit Packages. Use the Release Record for proof about one release and Passport for organization-wide evidence and history.

• Add external evidence links to the relevant release instead of keeping review context in chat or email.
• Use the internal Audit Package for complete organization records.
• Use the external Audit Package when sharing; it applies the configured disclosure and redaction policy.
• Verify package completeness and signature status before handing it to an approver, customer, or auditor.

Passport's Overview collects recent packages and evidence, while History traces actions back to their systems and releases. Owners and admins also have an Integrity workspace for advanced record verification and verification bundles.

Owners and admins configure disclosure defaults under Settings → Evidence defaults.

Operate workflow sessions with Runtime

Forge Runtime starts and monitors workflow sessions for governed systems. Use a dry-run first to validate the workflow without live cloud execution. Live execution requires a supported cloud connection and an eligible system workflow.

• Sessions — choose the system or workflow, start a dry-run or eligible live-cloud session, and monitor or control its execution.
• Triage — investigate stalled or failed sessions and recorded operational incidents.
• Runtime Data — inspect and prepare recorded data for supported retrieval or customization workflows.

Members can run simulations. Owners and admins can start and control eligible live-cloud work. Reviewers have read-only Runtime access.

Read portfolio trends in Analytics

Analytics summarizes technical signals recorded in Forge, including evaluation outcomes, Runtime session health, inference signals, and system counts.

Current values are recomputed from Forge records when the page loads. Trend charts use stored snapshot points, so they may not move until a new snapshot exists. Use Analytics for portfolio trends, not as a live estate-wide monitoring service.

Start with Portfolio pulse, then inspect Material changes. Check source availability and sample sizes before acting on a metric. Continuous Intelligence guidance is advisory and does not make a Promotion Gate decision.

Configure the organization

The Settings Control Center is role-aware. If you cannot change a setting, ask an owner or admin rather than trying to work around the control. Use its search and status filters to find configuration that needs attention.

• Members & invites — add people and assign organization roles.
• Integrations — connect cloud accounts and delivery infrastructure.
• Release governance — configure promotion and review policy.
• Evaluation rubrics and golden suites — define how releases are tested.
• Model benchmarks — maintain measured inputs used in architecture model decisions.
• Steering rules and inbound triggers — govern reusable policy and signed delivery ingress.
• Knowledge matching — configure organization engineering guidance used by Studio.
• Orchestration — configure agent and workflow contract policy.
• API reference and MCP access — connect approved developer and agent clients.
• Evidence defaults — control Audit Package disclosure and evidence expectations.
• Security and identity — manage login, provisioning, sessions, and organization controls.
• Vendor risk and exit & portability — review assurance and recovery readiness.
• Billing — manage the organization's plan and billing relationship.

Know what your role can do

Policy can further restrict an action. The Release Record explains the active blocker and the person who must act next.

When you cannot move forward

Glossary

Enterprise AI System

The long-lived portfolio record for an AI capability your organization owns and operates.

AI Release

One governed change to an Enterprise AI System.

Release Record

The source of truth for one release's processing, evidence, approval, and delivery history.

Promotion Gate

The policy result based on checks and evidence. It is advisory in shadow mode and has decision authority only when the applicable policy is enforced.

Model decision

The operator's measured selection, no-preference outcome, or rejection recorded from a frozen-workload candidate comparison and, when applicable, bound to an AI Release.

Delivery evidence fingerprint

The SHA-256 manifest that binds the reviewed AI System Package, Engineering Specifications, and Engineering Plan to a release submission or approval decision.

Evidence

Recorded facts, results, logs, and links used to support a release decision.

Audit Package

A portable, verifiable export of a Release Record and its permitted evidence.

Architecture baseline

The accepted system design against which later release changes are understood.

Dry-run

A Runtime execution that validates workflow behavior without starting live cloud work.

Forge capabilities

Forge Gate, Forge Portfolio, Forge Passport, Forge Studio, Forge Runtime, and Forge Analytics are connected platform capabilities within Forge.