Everything Dipp AI publishes, in one crawlable map.

The full directive execution path: entry, four sequential gates, three continuous controls, and two outputs.

cite: arch-component-schematic

Eleven classes of system of record reached through one governed path rather than point-to-point integrations.

cite: arch-integration-fabric

Interoperability protocols make systems reachable. Orcher sits above them and decides whether a specific action is authorized.

cite: arch-protocol-stack

Delegation across specialist agents under one directive and one accountable role, with authority never invented along the way.

cite: arch-multi-agent

Clinical authority cannot be delegated to a process.

cite: industry-healthcare

Every adjudication is a decision someone must own.

cite: industry-insurance

Supervised institutions need evidence, not dashboards.

cite: industry-banking-and-financial-services

Margin decisions at machine speed still need an owner.

cite: industry-retail

Physical consequence closes the loop on digital authority.

cite: industry-manufacturing

Public decisions must be explainable to the public.

cite: industry-government-and-public-sector

Network authority at carrier scale.

cite: industry-telecommunications

Reliability standards do not recognize autonomous intent.

cite: industry-energy-and-utilities

Validated systems demand verifiable authority.

cite: industry-life-sciences-and-pharma

Engineering decisions that reach a vehicle need a signature.

cite: industry-automotive

Commitments made by agents bind the carrier.

cite: industry-logistics-and-transportation

Professional responsibility is not delegable to a model.

cite: industry-legal-services

Approvals that move money need traceable authority.

cite: industry-real-estate-and-construction

Decisions about students require an accountable educator.

cite: industry-education

Command authority, enforced at execution.

cite: industry-military-and-defense

The 2026 baseline: agents are in production almost everywhere, and almost nowhere can an enterprise trace what they did.

cite: fig-p03-executive-summary

Three published 2026 incidents, each traced to a control that did not exist rather than a model that was not clever enough.

cite: fig-p04-trust-gap

A sandbox escape through a zero-day in a model registry proxy: 17,000+ attacker actions across a five-day detection gap.

cite: fig-p05-registry-proxy-case

Four independent failure modes, each measured. None is solved by asking reviewers to try harder.

cite: fig-p06-hitl-failure-modes

Clinical reading studies: expert accuracy collapses when the machine is confidently wrong, and the effect is strongest on the least experienced reviewers.

cite: fig-p06-automation-bias

Every published 2026 agent incident maps to a missing control, and each missing control maps to exactly one Orcher component.

cite: fig-p08-incident-to-component

Vertical agent companies are scaling fast, and each is bound to a single provider harness — capability without provider-neutral control.

cite: fig-p09-specialist-market

Within eleven days the field produced two opposing public letters and one binding statute. All three assume an accountable human decision-maker.

cite: fig-p10-regulatory-moment

The population that acts inside enterprise systems is now overwhelmingly non-human, and it is growing faster than any identity programme built for people.

cite: fig-p11-nonhuman-identity-ratio

Courts and legislatures have already answered the objection that an autonomous system diffuses responsibility.

cite: fig-p12-liability

One entry point, four sequential gates, three continuous controls. If any Layer 1 gate fails the directive halts and nothing partial is written.

cite: fig-p13-orcher-layers

Policy describes intent. A separate process boundary that must return a pass before commit converts intent into a property of the system.

cite: fig-p13-will-not-to-cannot

Eleven classes of system of record reachable through one governed path, instead of point-to-point integrations that each carry their own authority.

cite: fig-p14-integration-fabric

MCP, A2A and ACP make systems reachable. None of them decides whether a specific action carries authority; Orcher sits above them and does.

cite: fig-p15-protocol-stack

Reachability arrived first and fast. Governance of what is reached did not.

cite: fig-p15-mcp-adoption

Four specialist agents, one directive, one accountable role. Authority is delegated across the chain without ever being invented along the way.

cite: fig-p16-claims-to-cash

Naive multi-agent systems fail most of the time. Coordination research raises shared-decision rates; it does not decide authority.

cite: fig-p16-multi-agent-stats

Spend is compounding faster than the ability to show a return, and the stated barriers are governance barriers.

cite: fig-p21-market-stats

Ten platform vendors mapped across the four capabilities that decide whether an action can be trusted. Only a provider-neutral control plane covers all four.

cite: fig-p21-vendor-convergence

Authority is scoped to what the role actually carries — a median of 14% of the underlying account's permissions.

cite: fig-p22-role-scoped-authority

Training exclusion is a contractual and configuration state that varies by tier and by product surface — which is why it has to be verified before routing, not assumed.

cite: fig-p24-vendor-training-policy

Seats and tokens measure activity. A Verified Execution Cycle measures a directive carried to a provable outcome.

cite: fig-p27-verified-execution-cycle

Existing agent benchmarks measure capability under contamination. None measures whether the evidence produced would survive an audit.

cite: fig-p28-benchmark-gaps

Nine active regimes, all converging on the same requirement: show who authorized the action and prove the record was not altered.

cite: fig-p29-standards-landscape

Seven components across two layers. Four decide whether an action commits; three govern data, cost and visibility while it runs.

cite: fig-p36-seven-components