Deterministic task and review signals.
Private beta pilot experience
Run a controlled AI software task from intake to approval, record, and analytics.
Evaluate ForgeLayer through one repeatable, deterministic workflow without live agents, repository access, external calls, or production data.
Preview onlyNon-persistentNo live executionGeneric Platform Guidance: Human approval required
Pilot purpose
Choose the control problem your company wants to evaluate
Each scenario reuses the existing Journey and module fixtures. Selection and approval state stay in this browser session only.
ScenarioSecure a payment webhook
Current stageHuman Approval
Progress6/10 stages
Approvalpending
Records6 previews
Expected outcome
Secure a payment webhook
Show how sensitive payment and session changes remain blocked behind evidence, policy, and explicit human approval.
- Evidence available
- 3
- Known risks
- 3
- Changed paths
- 2
- Data state
- deterministic preview
Human approval gate
A person decides whether the pilot advances
These controls update local preview state only.
Journey preview
One scenario across the ForgeLayer control plane
The pilot steps reuse the same stage status, ownership, records, and approval state as the connected Journey.
- 01WorkspaceWorkspacecompletedInspect
Task and private-beta boundaries are understandable.
- 02PromptForgePromptForgecompletedInspect
Prompt risks and instruction boundaries are visible.
- 03AgentOSAgentOScompletedInspect
Primary and supporting agent fit is understandable.
- 04Control FlowControl FlowcompletedInspect
Control stages and blocked actions are visible.
- 05Review RunsReview RuncompletedInspect
Evidence, policy, risk, and verification can be inspected.
- 06MergeGuardMergeGuardcompletedInspect
Merge recommendation remains separate from approval.
- 07Human controlHuman Approvalapproval requiredInspect
A person makes an explicit local-only decision.
- 08Agent Work RecordsWork RecordblockedInspect
Scope, risk, evidence, and approvals appear in a draft record.
- 09AI Work LedgerLedgerfutureInspect
Related records connect in a non-persistent Ledger preview.
- 10AI Work AnalyticsAnalyticsfutureInspect
Derived control activity is visible without production telemetry.
Evidence and records
What the pilot can demonstrate
Every record reports persisted: false.
Prompt, authority, review, merge, approval, and execution gates.
Human decision, never inferred.
PromptForgeprompt presetpersisted: false
AgentOSrouting planpersisted: false
Control Flowflow reviewpersisted: false
Review Runsreview runpersisted: false
Agent Work Recordsagent work recordpersisted: false
AI Work Ledgerledger recordpersisted: false
Completion checklist
What a reviewer should confirm
- Workspace
Task and private-beta boundaries are understandable.
- PromptForge
Prompt risks and instruction boundaries are visible.
- AgentOS
Primary and supporting agent fit is understandable.
- Control Flow
Control stages and blocked actions are visible.
- Review Run
Evidence, policy, risk, and verification can be inspected.
- MergeGuard
Merge recommendation remains separate from approval.
- Human Approval
A person makes an explicit local-only decision.
- Work Record
Scope, risk, evidence, and approvals appear in a draft record.
- Ledger
Related records connect in a non-persistent Ledger preview.
- Analytics
Derived control activity is visible without production telemetry.
Pilot evaluation
Evaluate product clarity, not customer performance
These statuses describe what the deterministic preview demonstrates. They are not customer scores, benchmarks, or production metrics.
Task understanding
Can a reviewer identify the task, scope, changed paths, and expected outcome?
Agent fit
Does the deterministic route explain primary and supporting agent fit without granting authority?
Control-risk visibility
Are sensitive work, blocked actions, and required oversight understandable?
Review evidence
Are observed evidence, missing checks, and simulated verification clearly separated?
Human approval clarity
Is the human decision visibly distinct from a system recommendation?
Traceability
Can a reviewer follow the task across modules and linked preview records?
Ledger completeness
Does the preview connect task, route, review, decision, evidence, and outcome?
Analytics visibility
Do derived metrics explain control activity without production telemetry?
Safety-boundary clarity
Are disabled execution, writes, posting, merge, and persistence explicit?
Production orchestration
Can the pilot execute agents or persist a production workflow?
Automatic actions
Can the pilot write to repositories, post to GitHub, or merge automatically?
Product proof points
What this pilot is designed to show
- A bounded task can move through one consistent control journey.
- Prompt and agent-routing risk remain visible before work is trusted.
- Review evidence, policy context, and merge guidance stay distinct from human approval.
- Work Record and Ledger previews connect decisions without claiming production persistence.
- Derived analytics explain control activity without customer telemetry or performance scoring.
- Every production side effect remains disabled by default.
Future requirements
What remains unresolved
- Production workspace identity and role boundaries
- Workspace-scoped persistence and RLS
- Durable approval, evidence, and audit records
- Explicit GitHub write permissions and idempotent workers
- Controlled execution permissions, audit events, and kill switches
- Production analytics, retention, consent, and privacy controls
Deterministic completion summary
Secure a payment webhook
9 modules demonstrated, 6 controls applied, 3 evidence signals, and 6preview records. Approval state: pending.
Preview onlyNon-persistentNo live execution