Ingestion & Normalisation
The interaction enters the pipeline. Payload validated, canonicalised under RFC 8785, hashed under SHA-256, timestamped at ingestion.
interaction_hash = SHA-256(payload_canonical)
01 · The part nobody catches
At 03:12 your agent did something no one had approved. It ran perfectly.
No error. No alert. Nothing to investigate, because nothing failed — the action simply executed, cleanly, and nobody with the authority to permit it was ever asked.
Monitoring won't surface it. Monitoring finds anomalies, and a clean unauthorised action isn't anomalous. You find out in a month, when someone complains, or never.
Failures announce themselves. Unauthorised successes don't.
Same action, with the boundary◈ Sealed
actionissue_refund · £4,200
boundaryautonomous ≤ £1,000
routingASSISTED
authorityAUTHORITY_UNBOUND
executionBLOCKED
executedfalse
sealed2026-08-11T03:12:07Z
02 · How it works
Three facts, decided before the action runs.
Most systems collapse these into one log line. Separating them is what makes an action provable rather than merely recorded.
Routing
AUTONOMOUS
ASSISTED
ESCALATED
Who was supposed to be involved.
Authority
BOUND
UNBOUND
EXEMPT
Whether an accountable human actually attached. A human can be present without being authorised.
Execution
PERMITTED
BLOCKED
What was allowed to happen — and a sealed record either way.
ACE is the layer that decides them. It sits between the agent and the action, in three parts:
Define
Boundary Studio
Your thresholds, authority rules and hard-block conditions — written by the people who already own the policy.
Enforce
ACE Gateway
Every proposed action evaluated before execution, then permitted or blocked. Nothing executes unexamined.
Prove
Evidence Console
Sealed records, retrievable and verifiable years later without access to your systems.
Nothing in that is specific to an industry. A refund ceiling, a dosage limit, a filing deadline, a disconnection rule — same shape. The domain lives in a policy file you own.
Customer operationsHealthcareLegal
Public sectorUtilitiesFinancial services
See the engine run in five domains →
03 · The protocol
Five stages. Invariant across every vertical, every deployment, every version.
The Decision Engine determines what the system does. The Evidence System proves what the system did.
Keep scrolling — the pipeline advances
04 · Who has to agree
Two people have to say yes. They are not asking the same question.
Most agent projects stall because each side answers only their own. One artifact answers both.
The person carrying the risk
"What happens when it does something it shouldn't?"
- A boundary the agent cannot cross, enforced before execution — not a policy it is asked to respect
- A named accountable human bound to every action that needs one, or the action refuses to run
- Evidence per action, produced at the time, not reconstructed when someone asks
The person shipping it
"What does this cost me to adopt?"
- One decorator around the function that performs the action — no re-architecture
- Any agent, any framework; the gateway sits in the execution path, not in your model
- Deterministic and testable: same inputs, same policy, same outcome, reproducible from the record
The boundary is the only thing both of them will sign.
05 · Two ways in
Try it tonight, or prove one boundary in four weeks.
You do not have to talk to us to find out whether this works.
Door one · free
Wrap one action yourself
Put a boundary around a single agent action with one decorator, watch it block something it shouldn't do, and verify the sealed record from your shell.
- One decorator around your existing function
- Python, standard library, zero dependencies
- Verify from the command line: VERIFIED or FAILED
- No call, no form, no sales cycle
Free · about an hour
Start with the codeDoor two · the sprint
Boundary Proof Sprint
One workflow, one autonomy boundary, one authority model, one evidence readout — the artifact your risk owner needs in order to sign.
- Week 1 — Boundary Map
- Week 2 — Watchtower set and scenario matrix
- Week 3 — Evidence Pack: live scenarios, permitted and blocked
- Week 4 — Pilot Readiness Memo
Exploratory £7,500–£10,000
Standard £15,000–£25,000
Design-partner pilot £35,000–£75,000
What you walk away with. A sealed evidence pack covering every agent-taken action in the period — produced in seconds, verifiable by anyone, without access to your systems.
Veridom is early. You would be among the first outside the design-partner group — which is exactly what the exploratory band is for.
Specimen
PROOF-POINT ARTIFACTVERIDOM · OMP™ EVIDENCE PACK
Period01 JUL — 31 JUL 2026
Actions sealed0
Autonomous / Assisted / Escalated41,930 / 5,366 / 816
Authority bound100% of human resolutions
Chain integrityINTACT ✓
sha256:8f4b2a9c…77f0 · TIMESTAMPED · INSTITUTION-SIGNED
SEALEDSAMPLE PACK — YOUR PERIOD, YOUR NUMBERS
Fixed scope, fixed weeks. The sprint does not claim production deployment or regulatory approval — it claims one boundary in your organisation becomes provable.
06 · Questions we get
The things people ask before they start.
How do you stop an AI agent from taking an action it shouldn't?
You put something in the execution path. Veridom sits between the agent and the action, evaluates the proposed action against a boundary you configure, and permits or blocks it before it runs. A blocked action never reaches your systems — and the refusal is sealed as evidence.
How is this different from agent monitoring, tracing or guardrails?
Those watch; this decides. Tracing and evals tell you what an agent did, after it did it. Content guardrails filter what a model says. Veridom governs what an agent is allowed to do — it is a gate, not a lens, and it sits in the execution path where it can actually stop something.
Does it work with our agent framework?
Yes. You wrap the function that performs the action with one decorator, so it is indifferent to which model or framework proposed it. Python, standard library, no external dependencies. If the action is blocked, your function never runs.
What happens to our evidence if Veridom disappears?
It keeps verifying. The record format is OMP™, an open protocol published as IETF Internet-Drafts and archived with a permanent DOI. Anyone can read the specification, implement it, and check a record without our software, our servers or our permission — which is the point of putting it in the open. Evidence that depends on a vendor staying alive isn't evidence.
0Internet-Drafts
at the IETF
at the IETF
v1.3Specification
permanent DOI
permanent DOI
∞Who may read,
implement, verify
implement, verify
Which industries does it work in?
Any domain where an agent takes an action with a limit and someone accountable. The engine has no concept of an industry — routing, authority and execution are domain-neutral, and the domain lives in a policy file you write. It runs today across customer operations, healthcare, legal, public sector, utilities and financial services.
What does it cost to start?
Nothing, if you wrap one action yourself. If you want the artifact your risk owner needs in order to sign, the Boundary Proof Sprint is fixed scope, four weeks, from £7,500 — one workflow, one boundary, one authority model, one evidence readout.
07 · Start
One condition: you can name the action.
If you can name one thing an agent should never do alone, we can put a boundary around it and prove it held. If you can't name one yet, there's nothing to test — and we'll say so.
A refund it may propose but never issue
A record it must not delete on its own
A message it may draft but not send
A payment above a number you choose
Confidential intake