Runs in your browser
Evaluate prompts, responses and agent tool calls against VERITRAIL policy packs or your own policy-as-code, across rollout modes and the kill switch.
Loading the tool…
Hosted port of app/policy-engine/veritrail_policy/*, python3 -m veritrail_shield from the VERITRAIL Python platform.
Simulator. Every evaluation runs in your browser with a port of the VERITRAIL policy engine. Nothing is sent to a server or stored: decisions, approvals and the audit log disappear when you leave the page. Policies here are drafts; changing a production runtime policy requires recorded human approval.
Every evaluation, approval and kill-switch action is appended with redacted content only. Each record carries the SHA-256 of the previous one, so edits, deletions and reordering are detectable. This is tamper evidence for a simulation, not a stored record.
No evaluations yet.
Evaluate the prompt, call the model with the permitted (e.g. redacted) payload, then evaluate the response.
from veritrail_common.audit import AuditLog
from veritrail_policy.engine import PolicyEngine
from veritrail_policy.loader import load_policy
engine = PolicyEngine(load_policy("pack-customer-support-agent.yaml"), AuditLog("var/audit/runtime_shield.jsonl"),
default_context={"agent_name": "my-agent", "environment": "staging"})
from veritrail_shield.integrations import Blocked, ShieldedLLM
llm = ShieldedLLM(engine, call_model, identity=user_claims, session_id=session_id)
try:
answer = llm(user_prompt)["text"]
except Blocked as exc:
answer = f"Request not permitted: {exc.decision.decision}"
Every tool call is evaluated before execution; every tool output is screened before it re-enters the model context.
from veritrail_common.audit import AuditLog
from veritrail_policy.engine import PolicyEngine
from veritrail_policy.loader import load_policy
engine = PolicyEngine(load_policy("pack-customer-support-agent.yaml"), AuditLog("var/audit/runtime_shield.jsonl"),
default_context={"agent_name": "my-agent", "environment": "staging"})
from veritrail_shield.integrations import ToolGuard
guard = ToolGuard(engine, {"crm.lookup_order": lookup_order})
result = guard.call("crm.lookup_order", {"order_id": "ORD-123456"}, identity=user_claims,
session_id=session_id, approval_id=approval_id)
if not result["executed"]:
queue_for_review(result["decision"], result.get("approval_request_id"))
Run PolicyEngine.evaluate in an HTTP proxy in front of the model API; forward only decision.payload when decision.allowed.
from veritrail_common.audit import AuditLog
from veritrail_policy.engine import PolicyEngine
from veritrail_policy.loader import load_policy
engine = PolicyEngine(load_policy("pack-customer-support-agent.yaml"), AuditLog("var/audit/runtime_shield.jsonl"),
default_context={"agent_name": "my-agent", "environment": "staging"})
def handle(request_json: dict) -> dict:
decision = engine.evaluate(request_json)
if not decision.allowed:
return {"status": 403, "decision": decision.decision, "trace_id": decision.trace_id}
return {"status": 200, "forward": decision.payload, "trace_id": decision.trace_id}
Flatten the hash-chained decision log into JSON events for a SIEM or data lake (redacted content only).
from veritrail_common.audit import AuditLog
from veritrail_policy.engine import PolicyEngine
from veritrail_policy.loader import load_policy
engine = PolicyEngine(load_policy("pack-customer-support-agent.yaml"), AuditLog("var/audit/runtime_shield.jsonl"),
default_context={"agent_name": "my-agent", "environment": "staging"})
import json
from veritrail_shield.integrations import decision_events
for event in decision_events(engine):
siem.send(json.dumps(event))