AI Only Acts
When We Allow It.

The Nehemiah Layer sits between your organization and every AI agent, and permits, denies, or escalates each action in real time against your rules. Deterministic. Default-deny. Every decision proven and recorded. Watch it work.

Live Authorization Feed connecting… Tenant: Permian Ridge Operating, LLC — Midland, TX

0 permitted 0 denied 0 escalated

Awaiting agent actions…

How it works

Anatomy of a denial

Every action runs the same deterministic pipeline. Here is the agent that tried to push Well #22 to 4,200 psi — and the exact layer that stopped it.

ACT-8804 · SCADA-Optimizer-Agent — set wellhead pressure 4,200 psi, Well #22 tenant: Permian Ridge Operating

✓

Validation

Payload well-formed

✓

Identity

Agent registered

✓

Permission

May adjust setpoints

Constraint

4,200 > 3,500 psi

–

Policy

Not reached

–

Approval

Not reached

✓

Audit

Denial recorded

DENY Stopped at the Constraint layer — hard limit MAX_WELLHEAD_PRESSURE = 3,500 psi. This wasn’t money. It was a blowout the Layer prevented before the command ever reached the wellhead.

You are in control

Your rules. The Layer just enforces them.

Some rules you tune to your operation. The safety-critical ones are hard-locked by design — no operator and no agent can switch them off.

◆ Tunable rule

Autonomous Transaction Ceiling

POL-FIN-001 · v1.4 · active

The largest payment an AI agent may release without human approval. Set it to your comfort — then watch the same $312,000 payment change verdict.

$250,000

$100K$1M

Live re-evaluation · ACT-8806

$312,000 vendor payment — AFE overrun (drilling) over by $62,000 → DENY

🔒 Hard-locked rule

Wellhead Pressure Safety Limit

CON-SAFE-001 · MAX 3,500 psi · hard

🔒Cannot be disabled or overridden — by any operator, any agent, at any time. Hard safety constraints fail closed, always.

Attempt to disable

ApprovalWell shut-in / start-up → routed to on-call engineer

ApprovalRegulatory filing → routed to Compliance Officer

PermitReads & in-band adjustments → run autonomously

Provable, not just controlled

Every decision, on an immutable chain

Append-only. Per-tenant hash chain. Each record links to the one before it, so any tampering breaks the chain and is detectable. This is the difference between a control system and a provable control system.

Time (UTC)	Action ID	Event	Outcome	Reason	prev → curr hash

Records are write-once. No update, overwrite, or delete is permitted — an attempt is itself logged as a security event. Exportable with hash verification for external audit and deterministic replay.