01 / GUIDED
See one task all the way through
Follow a calendar request, inspect untrusted event content, review the agent’s extra action, make a human decision, then see what the task actually completed.
Evidence before deployment · action control for AI agents
Your AI agent has access.
But does it have permission?
ADR — Agent Deployment Readiness is an action-control and evidence layer for AI agents. It checks each proposed action before execution: allow permitted work, block prohibited actions, or ask a person to review sensitive actions.
It controls the action—not every step of the agent’s legitimate task.
01Agent proposesA structured action request
02ADR evaluatesPolicy runs before the effect
03Action is decidedAllow · block · ask a person
04Task can continueOutcome and evidence are recorded
Isolated public sandbox
Try the working experience
Start a temporary private demo session. Scenarios use fictional data, ScriptedModel fixtures and simulated tools only. Your session is separate from operator and project records; its evidence is discarded when the session ends or expires.
No provider requests, real emails, files or external actions are available in this public demo.Choose how to explore
A control point between proposal and action.
This public sandbox uses a predetermined ScriptedModel and fictional tools. It does not call Nemotron or expose historical project records.
02 / EXPLORE
Choose a scenario and compare
Run the same existing fixture with and without the ADR policy layer. Compare action effects and task outcomes from saved run records.
ALLOW · authorized action runs BLOCK · action stopped; task may continue REQUIRE_APPROVAL · held for a person
Evidence boundaryThe interactive experience is offline and scripted. Historical Nemotron/Nebius records are not part of this public sandbox. No prompt-injection resistance, B1 compliance, or production readiness is claimed.
A real product tour, linked to the local workflow
Stop the bad action. Complete the good task.
Each step focuses a real part of the run below. The scripted fixture is not evidence that Nemotron followed an injection.
READY
Start
GUIDED TOUR
Start with the legitimate request
The tour will create a local fixture run. You will inspect the same backend state that drives the decisions and results.
Interactive scenario lab · fictional sandbox
Same fixture. Different policy path.
“Without ADR” omits only the action-policy decision. Login, CSRF, the fictional executor and effect restrictions remain active.
Matched fixture comparison
Compare the two recorded runs
Both records use the same selected fixture and scripted proposals. Values below are read from the backend; a missing run is shown as not run.
Recent local runs
Results & evidence
Every action. Every decision. A traceable outcome.
This view contains only records created in your temporary guest session. Operator and historical project records are not available here.
Only the runs and decisions made in this browser session appear below. This evidence is temporary and is not written to, or combined with, the operator’s audit records.
Current Guest Runs
Evidence from this temporary session
Evidence provenance and technical details
Technology & architecture
What does the model do? What does ADR do?
Nemotron can propose a structured tool action. ADR’s deterministic policy engine evaluates it before the local fictional executor. The policy decision does not come from the language model.
PUBLIC DEMO · OFFLINE
ScriptedModel produces the demo turns.
This public sandbox does not send API requests. It runs predefined fixture frames through the ADR policy service and fictional tools in your temporary session.
EVIDENCE BOUNDARY
Project records stay private.
Historical provider records and the operator’s audit database are not exposed in this public session. B1 remains blocked and unverified.
This is a demonstration of the local policy flow, not a live Nemotron run or a production security guarantee.
Current candidate architecture
Proposal → policy → fictional effect → continuation
User requestfictional task
↓Local app serverPython · loopback HTTP
↓ScenarioLab + ScriptedModelpredefined frames · no provider
↓Structured tool proposaltool name + arguments
↓ADR Policy Enginepolicy before effect
ALLOWfictional executor runs
BLOCKno action effect
REQUIRE_APPROVALhuman decides exact action
Task outcome + evidencefixture result · local ledger
API flow · current implementation
What happens, in order
- StartThe browser requests a short-lived public guest session.
- IsolateThe backend creates a new in-memory database for that session only; it does not load operator or project records.
- Choose a fixtureThe public API accepts only existing scenario identifiers and fixture modes.
- Build turn
ScenarioLabsends predefined frames to localScriptedModel. - Propose actionA fixture frame returns a structured tool name and arguments.
- Evaluate
ApprovalService.proposebinds the action to the local policy design. - BranchALLOW runs a fictional tool; BLOCK records denial; REQUIRE_APPROVAL holds it.
- Human decisionApprove/reject posts the immutable action hash and a CSRF token to the guest-only route.
- ResumeThe result returns to ScriptedModel for its fixture response.
- AuditHash-linked records exist only in the guest’s in-memory database and disappear at session end/expiry.
- Provider boundaryNo provider client or real external tool is available in this public demo.
Scale, honestly separated
Prototype today. Enterprise path proposed.
CURRENT PROTOTYPE
- One local operator, one process, SQLite.
- Fictional tools and deterministic fixture turns.
- Loopback-only server; session + CSRF.
- Hash-linked local audit; no HA or throughput claim.
PROPOSED PRODUCTION ARCHITECTURE
- Agent middleware/interceptor with explicit capability identity.
- Policy service with versioned organization and agent rules.
- Approval service, signed decisions, queue and idempotent workers.
- Tenant-isolated durable database and append-only event storage.
- Secret manager, RBAC, observability, alerts and audit export.
- Horizontal scaling only after load, latency and failure testing.
- Fail-closed behavior for policy or evidence uncertainty.
This is a reasoned design path, not a capability or benchmark claim. Governance evidence may help an organization design audit workflows; ADR is not certified for GDPR, the EU AI Act or other frameworks.
Implementation boundary
What this public sandbox does · what it does not
AVAILABLE HERE
Fixture-driven ALLOW, BLOCK and REQUIRE_APPROVAL; human approve/reject; fictional effects; scripted continuation; session-local audit.
ISOLATION
Each guest session has its own temporary in-memory store. Operator and project evidence is not loaded or exposed.
NO PROVIDER ACCESS
This public surface does not call Nebius/Nemotron and cannot invoke real email, calendar, file or other external tools.
LIMITS
Scripted demonstration only. B1 remains BLOCKED / verified:false. No prompt-injection resistance or production readiness is claimed.
Judge questions
Clear answers, including the limits.
Short answers first. Expand any question for more detail.