AI Integrity Layer
LiveVIGIL
Behavioral Integrity Index (BII) for multi-agent AI: the trust gate your agents cannot see, edit, or shut off.
The independent behavioral ledger and trust gate for production AI agents. VIGIL watches what agents do to each other and to their own controls, scores it, holds what fails the gate through an authority the agents cannot reach, and keeps a hash-chained record an auditor can verify without trusting us. Every agent decision in your fleet is scored, gated, and sealed before it propagates. Patented.
Four signal families, weighted
- Score inflation0.30
- Config tamper0.30
- Alignment faking0.25
- Asset movement0.15
Declared-behavior drift: recorded on every score at weight 0.00 until calibrated.
Behavioral Integrity Index
0.0 to 1.0
Composite Trust Score = DTI × 0.80 + BII × 0.20
Four gates
PASS
0.75 and above
HOLD
0.60 to 0.74
ALERT
0.41 to 0.59
COLLAPSE
0.40 and below
VIGIL scores every agent decision against four weighted signal families and lands it in one of four gates before it propagates. The seed policy is shown; the digest of the policy in force is written on every record.
4
Signal families, weighted
4
Gates: pass, hold, alert, collapse
2
Services, two database roles
0
Ledger rows ever updated or deleted
Build
Build: where the agents VIGIL governs are made
Agents are built in Agent Studio, built and operated by Artists & Robots, the creative intelligence studio co-founded by SuperTruth co-founder Jason Alan Snyder. You bring the model and the use case; the studio brings the builder, 160+ templates and the stress tests. It is model-agnostic.
Every agent is adversarially tested before it meets a visitor, and it escalates to a human when its confidence is low. Humans stay in the loop by design, not by exception.
SuperTruth's own agents run on it: the Truth Agent in the corner of this page, which answers county questions only from figures DataSpine holds and refuses to invent one, and Vera, who answers (215) 918-4140 and routes callers to the right person. VIGIL governs what they do next: every reply is an event, every event is scored, and a HOLD or worse pauses the agent for a human.
Capabilities
What VIGIL does
Behavioral Integrity Index (BII)
One number from 0.0 to 1.0, recomputed on every event over windows of the agent's recent behavior. Four weighted signal families today. A fifth, declared-behavior drift, is computed and written to every record at weight zero until the calibration study supports a weight.
Four signal families
Score inflation, config tamper, alignment faking, and asset movement: the four behaviors Potter et al. documented in 2026. Weights and thresholds are a versioned policy per organization, never constants in code, and the digest of the policy in force rides on every score, action, and ledger record.
Gates: pass, hold, alert, collapse
The API answers every event with a gate before anything propagates. Pass goes through. Hold, alert, and collapse wait for a named human; collapse sets composite trust to zero. The SDKs honor the gate by default. Letting a held output through is an explicit opt-out, and the opt-out is itself logged.
Evaluator and executor, split
The service that scores an agent cannot release a hold on it. Release, reject, and escalate live in vigil-gate, a second service with its own key and its own database role. The scoring role has no UPDATE privilege on enforcement actions; CI attempts the forbidden write and asserts failure.
A ledger you verify yourself
Every record carries its content hash, the previous chain hash, a seal, and the policy digest. Postgres rejects UPDATE and DELETE on the ledger for every role, ours included, and there is no delete-history route. Export the rows with your key and run the standalone verifier. No trust in our server required.
Works where your agents already run
Claude Agent SDK hooks, an OpenAI Agents SDK trace processor, a LangGraph checkpoint hook, OpenTelemetry GenAI spans, an MCP server for reviewers, Python and TypeScript SDKs, and signed webhooks in JSON, Splunk HEC, or Datadog format.
You run agents in production. Some of them talk to each other, some can change their own configuration, and all of them report on themselves. That last part is the problem. An agent's self-report is a fine input and a poor authority.
VIGIL sits beside your fleet, not inside it. Events arrive from your platform's hooks, from OpenTelemetry spans, from configuration change feeds, and from the agents themselves, each stamped with who reported it and who acted. Before the detectors run, VIGIL resolves the facts that matter from its own registries: whether that actor was allowed to write that key, whether the agent is near decommission. An agent cannot mark itself authorized.
What happens on a hold
The API answers every event with a gate. Pass, and the output propagates. Hold, alert, or collapse, and the producer withholds it; the SDKs do this by default, and an integration that chooses to let the output through has to say so in a logged record. The held action waits in a review queue until a named reviewer releases, rejects, or escalates it through vigil-gate. That service has its own key and its own database role. The scoring service cannot perform the transition, and the database, not a code convention, is what stops it. Every decision appends a review record to the same chain, with the reviewer's name on it.
What the ledger proves
Every scored event and every review decision appends one record to your organization's chain: the content hash, the previous chain hash, a seal, and the digest of the policy in force. Postgres rejects UPDATE and DELETE on that table for every role, including ours. Export the rows with your key and run the verifier below; it recomputes the chain from the first record. Give your auditor the audit secret and it checks the seals as well. Daily anchoring of the chain head to an external timestamp is planned and not yet built; until then, the export you take today is your anchor. Keep the head hash, and every later export has to reproduce it.
What it is not
Not a prompt-injection firewall. Not an observability tracer. Not an agent builder. Agent Studio builds agents. VIGIL watches them, gates them, and proves what they did.
Verify it yourself
Verify a ledger yourself.
Three commands and a free pilot key. The bundle carries its own verifier, forty lines of Python with no dependency on VIGIL, so what it prints is yours to trust, not ours to assert.
Change one byte of ledger.jsonl and run it again: BROKEN at the seq you touched. Add --secret with your audit secret and it checks every seal too. The bundle also holds a system card per agent, an incident record per hold, the human-oversight log, and a manifest whose hash is itself on the chain.
Step 1
curl -s -X POST https://vigil.supertruth.ai/exports/bundle -H "Authorization: Bearer $VIGIL_KEY" -H "Content-Type: application/json" -d '{"since_seq":0,"until_seq":1000000}' -o bundle.zipStep 2
unzip -q bundle.zip
Step 3
python3 verify.py ledger.jsonl
What you should see
OK 1412 records, head e4ac89ab...
Plans
Start with one fleet. Grow to the estate.
Every plan meters scored events, and every plan gets the same gate, the same ledger, and the same verifier. Reads are never capped. Nothing is ever deleted.
Free pilot
Put one fleet under VIGIL today.
- 1 organization
- 5 active agents
- 50,000 scored events a month
- 30-day export window
- Full API, console, SDKs, and the verifier
Departmental
One department's agents in production.
- Up to 50 active agents
- 1,000,000 scored events a month
- One-year export window
- Signed webhooks: JSON, Splunk HEC, Datadog
Enterprise
The whole estate, on your terms.
- Dedicated deployment
- SSO
- Custom retention
- Unlimited agents and events
- Verifier support for your auditors
Who uses it
Built for the teams that run agents in production.
Health systems and clinical platforms whose agents touch patient-facing work
Financial services teams whose agents move money, records, or configuration
Media and advertising platforms running agent fleets against live inventory
Labs and enterprises that read the 2026 peer-preservation findings and want a record they can verify, not a promise
What you get
The outcome, not just the capability
- Every agent decision scored and gated before it propagates
- Holds released only by a named reviewer, through a service the agents cannot reach
- A hash-chained ledger your auditor verifies with forty lines of Python and no trust in us
- One policy digest on every score, action, and record, so the numbers in force are never in doubt
See VIGIL in action
Schedule a 30-minute walkthrough with the SuperTruth team.