Orcher · Component 03 of 07
Verifies the proposed action against systems of record before commit.
Layer 1 — sequential, gating
The Logic Scrubber runs in a separate process boundary from the agent that proposed the action. It checks the proposed write against the systems of record it will touch — the policy, the balance, the eligibility, the contract — and refuses anything that does not reconcile. This is the component that turns a model that will not misbehave into a system that cannot.
Without it, review is theater
A human approver facing hundreds of plausible-looking actions approves nearly all of them. Verification against a system of record does not fatigue.
What it emits
A verification result: pass with evidence, or halt with the reconciliation that failed.
Evidence
Figures drawn from the Enterprise Superintelligence Report, Vol. I, August 2026.
45.5%
human accuracy on cases where AI advice was wrong, versus 82% unaided
automation bias — Vol. I, p.6
2,992
alerts per day in the analogous security-operations case
37% investigated — Vol. I, p.7
30s
typical gateway approval timeout before an action auto-proceeds
Vol. I, p.7
Mechanism
Four stages, in order. Layer 1 components gate execution; Layer 2 components run continuously and never block.
01
Intercept the proposed write
The agent produces a proposed action, not a committed one. Nothing reaches a system of record until the Scrubber has cleared it.
02
Re-derive preconditions
Running in a separate process boundary, the Scrubber independently re-derives the facts the action depends on — the policy status, the balance, the eligibility, the contract term — from the systems of record themselves.
03
Reconcile
The re-derived facts are compared against the assumptions embedded in the proposed action. Any mismatch is a failure, regardless of how plausible the agent's output reads.
04
Pass or halt
A pass carries structured evidence into the ledger. A halt is total: nothing partial is written, and the reason is returned as a structured reconciliation failure rather than a generic error.
Operational contract
What it is not
It is not a second model reviewing the first
Model-checks-model inherits the same failure modes and the same blind spots. The Scrubber checks against systems of record, which do not hallucinate.
It is not a human approval queue
Approval gates that time out after thirty seconds and present hundreds of plausible actions produce rubber-stamping. Verification against a source of truth does not fatigue.
Failure behaviour
A Scrubber that cannot reach a system of record fails closed. An unverifiable action is treated identically to a failed one: the directive halts.
Questions
Where it shows up
Surfaced automatically from the components each solution engages and each industry relies on.
Solution
Turn agent activity into evidence a regulator will accept.
Use case
Clinical authority cannot be delegated to a process.
Use case
Every adjudication is a decision someone must own.
Use case
Supervised institutions need evidence, not dashboards.
The strategy behind the component
Binding a platform to one heavy frontier model wastes tokens and leaks proprietary data. Dipp AI's strategy — and the reason Orcher's seven components exist — is intelligent routing: decouple orchestration from any single provider, match each task to the cheapest model that can carry it, and keep prompts and data inside the enterprise's own tenant boundary.
DIPP AI / ORCHER
Seven components. One enterprise-owned control plane.
01 + 02 / AUTHORITY
Intent, named professional, permitted actions and systems of record.
06 / COST GOVERNANCE
Select the least costly sufficient model within the task’s stakes, latency target and spend ceiling.
Explore Cost Governance05 / DATA CONTROL GATEWAY
Region, redaction, retention and training-exclusion requirements determine which destinations qualify.
Models discussed in Dipp AI Research · September 6, 2026
Model families · exact versions and licences assessed per deployment
Hyperscalers, neoclouds and enterprise infrastructure
03 / Logic Scrubber
Verify proposed actions before commit.
04 / Immutable Audit Ledger
Preserve the decision and refusal evidence.
07 / Observability
Trace cost, latency and outcomes across routes.
Switch providers in a week, not a quarter.
An orchestration layer sits in front of every AI task rather than inside one vendor's SDK. Directives, roles, policy and evidence live in Orcher, so a provider change is a routing-table change — a week of work rather than a quarter of re-platforming — and no model vendor becomes the system of record for how the enterprise operates.
The protocol stackRoughly 90% of workloads never need the frontier.
Cost Governance routes by the stakes of the action rather than the habits of the developer. Standard, routine queries — roughly nine in ten workloads — go to cheaper open-weight or small in-house models; expensive frontier models are reserved for genuinely complex, high-reasoning tasks. Across 400-plus models and 70-plus providers the price spread is 4,500× (Vol. I, p.25), so the routing decision is the economics.
Cost GovernanceProviders never learn from your interaction data.
The Data Control Gateway forces data and prompt histories to stay inside the company's own cloud environment — a private tenant, in-region — and verifies training exclusion and the data boundary on every call before a token leaves. Using frontier capability never means donating enterprise interaction data to the firm that sells it back.
Data Control GatewayProof points
Dynamic model routing is only a strategy if it moves numbers a finance team and an auditor both recognise. These are the outcomes it is accountable to.
Cost
4,500×
price spread across 400+ models and 70+ providers
The spread between the cheapest sufficient model and the frontier default is the entire economic case for routing. A directive sent to the wrong class does not fail — it just costs orders of magnitude more than it had to.
Vol. I, p.25
Token waste
~90%
of workloads never need a frontier model
Routine classification, extraction, summarisation and drafting clear on in-house or open-weight classes. Reserving frontier capacity for genuinely complex work is what turns AI spend from a run-rate into a budget.
Dipp AI routing model · Vol. I, p.25
Budget
68%
of enterprise AI programmes run over budget
Overrun is a control failure, not a forecasting failure. Cost Governance applies a ceiling per role and per directive and halts loops at the step boundary rather than at the invoice.
Vol. I, p.9
Latency
10×
faster on routine classes than a frontier default
Cheaper classes are also materially faster. Routing the routine nine-tenths off the frontier shortens the median directive as a side effect of the economics.
Indicative class latencies, Orcher routing table
Auditability
28%
of enterprises can trace an agent action end to end today
Every routing decision Orcher makes — class chosen, alternatives considered, ceiling applied, boundary attested — is hashed into the same evidence chain as the committed action, so the economics are as auditable as the outcome.
Vol. I, p.6
Portability
1 week
to change providers, not a quarter of re-platforming
Because orchestration is decoupled from any vendor SDK, a provider change is a routing-table change. Directives, roles, policy and evidence stay in Orcher.
Dipp AI protocol stack contract
The other six
Layer 1
Plain language becomes the permanent record of what was asked.
Layer 1
Binds the directive to a human and a role, cryptographically and revocably.
Layer 1
Hashes the verified action permanently. Evidence, not logs.
Layer 2
Training exclusion and data boundaries verified before routing.
Layer 2
Routes by stakes and halts runaway loops.
Layer 2
Real-time cross-provider trace: which model, which role, what cost, what outcome.
Further reading · September 2026
Two long-form pieces per week through September 2026 — sourced, dated, and attributed. Each one links back to the components it argues about.
Jul 28, 2026
A new artificial intelligence research and product company, founded to close the gap between what enterprises can deploy and what they can prove.
By Odero Otieno
Aug 10, 2026
A 37-page examination of why enterprise agentic adoption has outrun the ability to account for it — and the seven-component architecture that closes the gap.
Aug 12, 2026
A contractual promise not to train on customer data is not a control. Dipp AI publishes the architectural commitment it holds itself to instead.
By Odero Otieno
Aug 14, 2026
Recommending a cheaper model is advice. Refusing to dispatch above a ceiling is governance. Orcher now enforces the second.
Orcher is deployed with named enterprises under the Human-in-the-Role model. Request a technical briefing with the founding team.
