The Grounded DI platform
Grounded DI develops architectures for governing agent actions, tool use, and output release.
Specify the rules. Evaluate the next step. Carry the decision into execution and retain the evidence.
Documented architecture
A warning becomes useful when it changes what the system is permitted to do. The roadmap connects rule evaluation to a control that connected infrastructure can enforce.
Read the technical roadmapA missing source, changed input, or threshold breach.
The policy version and domain rules for this task.
The condition checked against a stated requirement.
Continue, correct, hold, restrict, or deny.
A machine-readable permission or restriction.
The connected controller applies that restriction.
The decision, action, and supporting record retained.
Conceptual control sequence. The roadmap describes architecture and development direction; it is not a test result or certification of an operational system.
Protocol A / Practical functions
Protocol A supplies the governing structure: task, sources, scope, rules, and authority. ELOC connects detected deviations to responses; AGDI governs continuation; Output-State Control governs whether a candidate may advance or be released.
Architecture explanations with illustrative operational examples
Evaluate a condition against identified rules, thresholds, and a policy snapshot. Preserve which version governed the decision.
ExampleA weather advisory uses the location, observation time, and review threshold specified for that task. A stale observation triggers the applicable rule.
ELOC links a qualifying deviation to a disposition: correction, review, sandboxing, restriction, or denial. The response depends on the policy; it is not a fixed escalation ladder.
ExampleA missing citation sends a draft back for correction; an unauthorized external action is denied.
AGDI binds continuation to governance state. Its architecture covers tool invocation, workflow routing, downstream agent calls, and role transitions.
ExampleAn agent can prepare a proposed message, while a connected tool gateway withholds permission to send it until approval is recorded.
A candidate output remains unvalidated until the required source, rule, logic, version, audit, and integrity conditions are satisfied. It can be held, revised, or rejected.
ExampleA legal draft with an unsupported claim stays in review instead of advancing to a release step. Passing a release gate does not itself establish legal correctness.
Retain the inputs, policy version, evaluated condition, disposition, and execution evidence needed to reconstruct the decision path.
ExampleA reviewer can inspect why a tool call was restricted and which rule applied, rather than seeing only the final response.
The multi-agent architecture describes shared state interfaces, mirrored constraint trees, local checks, and version-aware dispositions with state hashes and audit lineage.
ExampleWhen a research agent hands work to a drafting agent, the receiving agent checks the inherited constraints before continuing. This describes the architecture, not a deployed integration.
The enforcement boundary
Policy context, evaluated rules, selected dispositions, permission or restriction state, and the decision record. These give a controller a specific decision to apply.
A tool gateway, orchestrator, middleware service, or provider must consume that state and actually permit, hold, restrict, or deny the operation. Identity, network access, compute shutdown, storage, and rollback remain infrastructure responsibilities.
An enforceable design requires an integration that cannot bypass the relevant control. The exhibits do not establish that every described integration is deployed, or that probabilistic model outputs become universally deterministic.
Recorded implementation behavior
On 9 September 2026, the recorded BriefWise test reported an exact replay identity match and rejection of altered content. The record provides a bounded example of replay and integrity checking.
Legal correctness was not evaluated. Independent external verification remains unverified. This test does not validate the full control architecture described above.
Inspect the recorded testIllustrative workflows
Choose a legal, weather, or engineering example to follow a missing source or failed check into a hold for review. These examples illustrate the architecture; they run no live analysis and are separate from the recorded test above.
A draft claim and the sources cited for it
The cited material must support the claim.
The cited passage does not support the draft claim.
These examples illustrate the process. They don’t run live analysis or release a result.
Four ways to look at AI-assisted work
Finding information, getting tasks done, and working in familiar tools all matter. Grounded DI emphasizes the checks between a proposed answer and a result someone is allowed to use.
01 / Research
Find and examine the material behind the work.
02 / Workflow
Organize analysis, drafting, and repeatable tasks.
03 / Everyday tools
Bring research, review, and drafting into the tools people use.
04 / Control
Keep support, unresolved questions, changes, replay, and approval tied to the result.
Grounded DI’s focusThese are overlapping areas of emphasis, not exclusive capabilities or a ranking of products. Adapted from Grounded DI LLC’s four-lanes legal-AI framework; this broader framing is not an independently verified market taxonomy.
Finished does not mean approved. In the illustrative trace console, an unsupported claim stops at review instead of moving forward as if it passed.
Explore the held-for-review example See the legal versionTechnical sources
The seven-stage sequence, ELOC, AGDI, Output-State Control, and infrastructure boundary.
The Control Plane, pages 3–5 and 12; Selected Exhibits — Foundations, pages 3–5. Architectural examples retain their original qualifications.
Work with Grounded DI
Bring the actions, tools, and release decisions you need to govern. Discuss licensing, evaluation scope, and the infrastructure required for enforcement.