dec_01J9ZQ4M7K
- Sentinel
- S-0142 · Certified credit analyst
- Timestamp
2026-03-12 · 14:32:07 UTC- Reasoning
Stable repayment history over 36 months; debt ratio within policy.
- Audit
- AI context, decision, timestamp, identity — exportable
Arbitrium — the threshold where automation stops and human authority takes over. HumanLayer routes high-stakes requests to a qualified Sentinel, then returns an auditable, documented decision artifact.
Are you a licensed expert? Become a Sentinel
Orchestration · Decision · Evidence
Depending on the decision level
Verdict, reasoning and timestamp
Built for sectors where accountability matters
01The problem
Today, AI can automate 80–95% of a workflow… But as soon as you need to:
…everything goes back “out of system”, manually, without SLA or traceability.
The brake on enterprise adoption isn’t execution. It’s the lack of authority.
The context
Projections 2027–2028. Amounts in USD. Sources: IDC (AI software revenue, GenAI/AI spending); Microsoft citing IDC (AI agents).
02Arbitrium
Arbitrium is not a person, but a system threshold. Only high-stakes requests cross it: automation continues everywhere else, without breaking compliance. Pick a scenario and play with the parameters.
“€42 expense report — client lunch, compliant with internal policy. Reimburse?”
Automatic execution
Below the threshold: the agent continues its workflow. The action is logged.
{ "arbitrium": false, "action": "auto_execute" }
Illustrative simulation. In production, each organization defines its own risk thresholds, automatic rejection rules and mandatory escalations.
03The solution
HumanLayer orchestrates human decisioning exactly where it’s indispensable. A three-layer system, from the AI-prepared case to the audit-ready decision.
Detect Arbitrium and route to a qualified Sentinel. The AI agent submits a complete case with context and supporting documents; HumanLayer routes it to a qualified Sentinel based on the required criteria.
A qualified Sentinel issues an explicit verdict: a structured decision, with comments and justification. HumanLayer does not automate the decision; it orchestrates the intervention of an accountable human.
A timestamped, reasoned, audit-ready decision artifact. Back to the agent: the decision is logged, auditable and ready to execute.
Result: human authority triggered only at the right threshold.
From AI case to enforceable decision
Result: AI can complete the workflow end-to-end without breaking compliance.
Designed with Legal, Risk and Compliance teams04Who decides?
It orchestrates the intervention of an identified, qualified, accountable Sentinel. Every decision is:
Automating decisions without an identified human accountable for them.
dec_01J9ZQ4M7K
2026-03-12 · 14:32:07 UTCStable repayment history over 36 months; debt ratio within policy.
Illustrative example, fictitious data.
05Use cases
HumanLayer integrates into every sector where a decision commits someone. AI prepares, the Sentinel decides.
“AI analyzed 200 credit files. For this customer, the score is in a gray zone (480). Approve the €50,000 loan?”
Prepares the complete file with history and recommendation.
A certified credit analyst decides within 2 hours (custom SLA).
“Out of 500 contracts analyzed, 12 include a change-of-control clause. Is clause §4.2.b acceptable?”
Flags the critical points of each contract.
A qualified lawyer validates each risky clause, with comments.
“The algorithm detects an anomaly on patient #4521’s MRI. Tumor probability: 78%. Confirm the diagnosis?”
Pre-analyzes the imaging and flags the anomaly.
An authorized radiologist validates or adjusts the diagnosis within 30 minutes (custom SLA).
“3 candidates shortlisted by AI. Candidate B has a 2-year gap on their resume. Process exception?”
Pre-screens and prepares the files.
An HR director validates sensitive decisions, with justification.
“Batch #7823 shows a 0.3% deviation from spec. Below the critical threshold but above tolerance. Release the batch?”
Detects deviations on the production line.
An authorized quality manager decides on release.
“Auto claim: photos analyzed, statements cross-checked. Timing inconsistency detected. Fraud?”
Flags suspicious cases.
A mandated expert decides on compensation.
“SaaS purchase request: €45,000/year. Budget not planned. 3 quotes obtained. Budget exception?”
Validates the request’s compliance.
A CFO approves exceptions according to policy.
“Admin access request to the prod server. Justification: ‘urgent debug’. User: junior dev, 10:30 pm. Approve?”
Evaluates the risk and context of the request.
A CISO validates sensitive access under a priority SLA (Enterprise option).
06Developers
HumanLayer integrates like Stripe or Twilio — but for decisions. Arbitrium is exposed as an escalation state in the system.
/api/v1/decisionsSubmit a case or a second opinion/api/v1/decisions?id=Status, SLA and signed decision/api/v1/decisions?agent_id=Agent history/api/v1/decision-keyVerification public keyimport requests
API = "https://humanlayer.systems/api/v1/decisions"
HEADERS = {"x-api-key": "hl_live_..."}
# Arbitrium threshold crossed: submit the case
created = requests.post(API, headers=HEADERS, json={
"agent_id": "agent-credit",
"request_type": "compliance", # level 1, 30-min SLA
"domain": "finance",
"summary": "KYC: partial match. Validate?",
"context_json": {"customer_id": "c_4521"},
"callback_url": "https://your-app.example/hl",
}).json()
# Track it (or wait for the signed webhook)
res = requests.get(API, headers=HEADERS,
params={"id": created["request_id"]})
req = res.json()["request"]
if req["verdict"] == "approved":
continue_workflow(req["decision"]) # Ed25519-signed
import hmac, hashlib, time
def verify_webhook(headers, raw_body, secret):
# X-HumanLayer-Signature: t=1700000000,v1=<hmac>
parts = dict(p.split("=", 1)
for p in headers["X-HumanLayer-Signature"].split(","))
if abs(time.time() - int(parts["t"])) > 300:
return False # too old: replay
signed = parts["t"].encode() + b"." + raw_body
expected = hmac.new(secret.encode(), signed,
hashlib.sha256).hexdigest()
return hmac.compare_digest(expected, parts["v1"])
// POST to your callback_url (X-HumanLayer-Event: decision.completed)
{
"id": "evt_3c9e…",
"type": "decision.completed",
"data": {
"request_id": "8b1f2c4e-…",
"status": "decided",
"verdict": "approved",
"reasoning": "Compliant with internal policy.",
"sentinel_ref": "S-4F2A9C",
"signed_at": "2026-03-12T14:32:07Z",
"signature": "base64…",
"signed_payload": "{…}"
}
}
07Differentiation
What you buy is not human time: it’s verified authority, plugged into your agents, with the evidence to back it.
Qualified Sentinels, traceable accreditations. Qualifications and authorizations verified before any activation.
Endpoints, formats, agent integrations.
Evidence, logs, context, versioning.
Guaranteed timelines, escalations, fallback.
Built for Legal, Risk and Compliance teams.
Qualifications and authorizations verified.
Exportable, signed decisions (Ed25519).
SLA breaches, second opinions, reassignment.
Mandatory MFA for Sentinels, encrypted secrets (AES-256).
Collection limited to what is strictly necessary.
SOC 2 Type II certified hosting; processes aligned with GDPR and Law 25.
08Pricing
Pay for value, not time. Each level maps to a type of responsibility and expertise; up to 60% of every decision goes to the Sentinel who renders it.
Validation of rules and policies.
Value: speed + accountability + audit
When rules are not enough.
Value: you buy judgment, not time
Final act by a qualified person.
Value: assumed accountability + strong barrier to entry
Enterprise options: premium SLAs, dedicated pools, governance, integrations, reporting.
Get pricing09Join the network
Lawyer, physician, analyst, CISO… Monetize your decision authority by validating AI agents’ requests. Full flexibility, paid per decision.
Diligent review of your application.
Out of 500 contracts analyzed, 12 include this clause. The AI flagged the critical points and attached the full context.
Illustrative preview of the Sentinel portal. Click to decide.
HumanLayer makes autonomy auditable, accountable and enforceable.
As AI advances, HumanLayer defines when not to automate.