{
"operation": "agent_change_gate",
"input": {
"subject": "[bounded revision metadata]",
"evidence": "[coarse evidence states]",
"policy": "[customer policy boundary]",
"context": "[scope and version]"
}
}DETERMINISTIC / FAIL-CLOSED / SCOPE-BOUND
GATE AI CHANGES
BEFORE MERGE.
For platform, security, and AI engineering teams, Verahelm evaluates bounded, caller-supplied evidence for an exact pull request or agent revision. It returns a required action—proceed within stated limits, block, or review—plus reason codes, required tests, and an optional signed authorization record. MCP admission and model migration gates use the same contract.
NO RAW WORKING DATA REQUIRED. No source code, prompts, traces, datasets, files, or URLs.
Start with restricted non-production Free access, or use paid hosted API access from $49/month. Create a Free key or view plan limits.
01 / INPUT TO DECISION
From proposed change to release decision.
Send bounded metadata and evidence states—not the underlying source material. Verahelm returns closed status codes and the next required action. The field map below is intentionally abbreviated; it is not a valid request, engine result, or decision rule.
{
"status": "unfavorable",
"result": {
"gate_state":
"additional_evidence_required",
"gate_codes": "[closed public reason codes]",
"required_action_codes":
"[closed public action codes]"
},
"decision_envelope": "[optional signed record]"
}Complete schemas, required fields, enums, and lifecycle endpoints are in the OpenAPI contract. Complete worked request/result pairs remain excluded to protect the private decision engine.
02 / PRIMARY WORKFLOWS
Three gates. One bounded decision contract.
Gate an agent or pull request
- Change
- New agent behavior, tools, permissions, or exact PR revision.
- Evidence
- Protected-policy source, exact-head tests, required reviews, and rollback readiness.
- Result
- Gate state, blocking codes, and required actions.
- Fits
- Required CI check before merge or release.
Admit an MCP server or tool change
- Change
- Publisher, release, capability, permission, or runtime-boundary change.
- Evidence
- Identity, supply chain, capabilities, authorization, operations, and rollback state.
- Result
- Admission state, restrictions, evidence gaps, and revisit triggers.
- Fits
- MCP gateway or internal tool registry.
Bound a model or provider migration
- Change
- Model, provider, configuration, or workload migration.
- Evidence
- Matched metrics, compatibility checks, failure cases, rollout plan, and rollback readiness.
- Result
- Readiness state, evidence gaps, required tests, and rollout actions.
- Fits
- Shadow, canary, or bounded-pilot promotion.
03 / INTEGRATION
Measure locally. Submit the minimum. Enforce the result.
- 01Produce evidence
Tests, evaluation systems, and authorized evidence sources produce structured results.
- 02Submit a boundary
CI, an MCP gateway, or an internal platform sends the minimum permitted summary.
- 03Receive an action
Verahelm returns a bounded status, reason codes, and the next required test.
- 04Enforce locally
Your pipeline or gateway applies customer policy and fails closed on uncertainty.
- 05Verify lifecycle
When requested, verify the signed, expiring Decision Envelope and its current status.
PRODUCTION INTEGRATION.
Authenticate evidence sources and submitters, protect signing and verification keys, validate subject binding and expiration, fail closed when verification is unavailable, and preserve audit records under your own policy. Verahelm evaluates submitted evidence; it does not prove that every unauthenticated caller assertion is true.
DIGEST-ONLY EXAMPLE MAPPINGS.
Public recipes show how Promptfoo, OPA, SARIF, Sigstore/in-toto, and SLSA outputs can be reduced to digests and coarse status references before they cross the service boundary. These are tested example mappings—not claims of native integration or independent evidence verification. Review the integration recipes.
04 / POLICY ALTERNATIVES
Why not just use OPA or custom CI rules?
Teams author and maintain the decision logic. This is often the right choice when requirements are known, stable, and already supported by mature policy infrastructure.
- Organization-specific logic
- Direct local enforcement
- Full control—and full maintenance responsibility
Verahelm converts bounded evidence into a standardized decision artifact with reason codes, required next tests, and optional signed lifecycle records.
- One contract across PR, agent, MCP, and migration gates
- Scope- and version-bound results
- Expiry, revocation, and supersession support
OPA, a GitHub Action, or an internal engine can enforce a Verahelm result. Evaluation and observability systems can supply evidence. Human review remains appropriate for high-risk decisions.
- Verahelm does not replace evidence producers
- Customer policy remains authoritative
- Custom rules may be simpler for narrow stable cases
05 / PRODUCT BOUNDARY
What Verahelm does—and does not do.
- Produce deterministic gate results from bounded structured evidence
- Return blocking codes and required next tests
- Support enforceable release and admission workflows
- Produce signed decision artifacts that integrations can verify
- Guarantee that an AI system is safe
- Replace model evaluation, observability, or qualified human review
- Certify general regulatory compliance or authorize production deployment
- Inspect source code or validate unauthenticated caller claims as true
06 / AVAILABLE NOW
Inspect, test, or purchase without guessing the next step.
Run live gate profiles
Developer and Professional plans use fixed monthly units, hard caps, and no automatic overage.
From $49/monthCreate a restricted Free key
After email, human, and Stripe payment-method verification, eligible accounts receive 30 units each calendar month, selected flagship scopes, low rate and concurrency limits, and a non-production marker. A payment method is required; no charge or subscription is created.
$0Verify a fictional envelope
Inspect the public schema, offline verifier, verification-only GitHub Action, threat model, and fictional conformance fixtures.
$0DESIGN-PARTNER PILOT / NON-PRODUCTION.
Use one non-production repository and an existing review workflow. Measure setup time, review time, binding failures, stale-decision reuse, integration burden, and willingness to pay. No source, prompts, traces, or repository content are required.
07 / USE BOUNDARY
Business evaluation support—not a sole basis for consequential decisions.
Bounded business use
United States business and professional users, age 18 or older, authorized to act, using non-confidential inputs for non-regulated internal evaluation support.
Consumer or high-consequence use
No medical, legal, credit, insurance, employment, housing, education, criminal-justice, immigration, child-directed, safety-critical, or similar consequential decisions.
Independent controls remain
Customers remain responsible for evidence authenticity, permitted use, final decisions, release controls, and qualified human review.