The Grounded Intelligence Console (GIC) captures the judgment of an organization's most experienced people as structured, provenance-bearing knowledge — then puts it to work, every recommendation traced to the source behind it. The same systems-engineering discipline applies to whichever information-intensive domain presents the problem: an industrial capital project, a multi-district litigation, a defense modernization, a national archive. Built on two granted U.S. patents — US 6,883,008 (2001) and US 7,779,007 (2007).
Complex problems — whether industrial or legal — are fundamentally systems and information problems. Crivella designs the systems that allow organizations to understand those problems, control them, and ultimately prevail within them.
The platform was not adapted from litigation, and it was not adapted from heavy industry. It is one substrate — the Grounded Intelligence Console — configured for the domain in front of it. When an asset team plans a turnaround, the discipline captures their planning practice. When plaintiff's counsel works a production, the same discipline grounds AI in the produced record.
The unifying problem is knowledge churn — the institutional capability that degrades every time an expert retires, a matter rotates counsel, or a custodian leaves. The GIC converts that knowledge into a durable, auditable record once, so the organization stops paying to learn it again.
Scale of Impact
One platform, configured per domain through its Coding Manual. These are current capabilities — not a sequence the firm grew through, but four fronts on which the same engineering runs today.
Capturing the planning practice of asset and capital teams as guided, provenance-bearing decision support — turnarounds, plant modifications, debottlenecking, equipment replacement, permits and regulatory submissions. A less-experienced practitioner produces a plan at the level of the most senior people, every recommendation traced to the standard behind it.
Representative projects: Timken Faircrest greenfield specialty-steel facility, Goodyear modernization, Corning Asahi Glass optimization, and the Voisey's Bay nickel development (~$7 billion).
Core Capabilities
The same discipline applied to the most adversarial document environments in American practice: advanced analytics, grounded AI, systems integration, and strategic planning brought to the execution of complex matters — from the first meet-and-confer through trial or settlement.
Representative matters: Baycol (MDL 1431), Trasylol (MDL 1928), Actos (MDL 2299), DePuy Hip (MDL 2197), Zantac (MDL 2924), Roundup (MDL 2741), and the pelvic-mesh and device product-liability clusters.
Key Innovations
Bridging strategy, operations, and technology — transformation plans, value-creation strategies, implementation management, and integrated operational systems, including mission-critical defense infrastructure.
Representative engagements: US Marine Corps Depot Maintenance Modernization (MARCORLOGBASES), Thomas Steel Strip enterprise modernization, and Inco Alloys value-enhancement programs.
Service Areas
Systems that capture complex knowledge, enable deep analysis, and support institutional decision-making — across religious and academic research, national historical institutions, and large-scale archives.
Notable implementations: the John Henry Newman canonization global knowledge and AI research platform, and the Gettysburg National Battlefield & Museum strategic-technology and interpretive systems.
Applications
A document is a container. A Knowledge Object is a discrete piece of institutional knowledge — a best practice, a workflow step, a finding, a recommendation — described by a structured metadata record and carrying its own provenance and reliability grade. The platform reasons over the objects, not the files they happened to arrive in.
Every Knowledge Object is described by a Knowledge Element — a record in the Knowledge Element Metadata Architecture (KEMA). KEMA is explicit about uncertainty: each field is either evaluated (with a value), confirmed absent, or not yet evaluated. There is no silent gap. A reviewer years later can tell the difference between "we checked and there was none" and "no one looked."
Provenance is recorded to the PROV-O standard: every Knowledge Element carries the chain of derivation back to its source — which document, which column, which run, which human or agent. Attached to that chain is a reliability grade: an explicit trustworthiness rating, so a downstream reader weighs a graded finding rather than an anonymous assertion.
Knowledge Objects compose recursively — an object can be built from other objects — which is what lets a deposition outline, a turnaround plan, or an expert report be assembled from graded, traceable parts rather than typed from memory.
Knowledge Object
A node in a knowledge graph — a discrete piece of institutional knowledge, not a file.
Knowledge Element (KEMA)
The structured metadata record describing the object — three-state fields, no silent gaps.
Provenance chain (PROV-O)
The recorded derivation history — source, run, and contributor, human or agent.
Reliability grade
The trustworthiness rating attached to every Element — carried with it everywhere it is used.
Nothing enters the corpus by accident. KEMAcoder is the intake pipeline that converts a raw submission — a litigation production, a set of work product, a transcript, a planning record — into standardized Knowledge Elements. It is a strict funnel: skipping a stage breaks the provenance guarantee, so the funnel is not optional.
01 · Classify
The submission is inspected and typed — legal production, annotated bundle, entity-centric set, and others — most-specific first.
02 · Gate
A type-specific validator returns one of three decisions — clear, needs human decision, or blocked. A blocked submission is not processed; it is sent back with a correction request.
03 · Quality evaluation
A pre-induction quality evaluation grades the submission before anything is committed — and, when blocked, becomes the correction letter itself.
04 · Process & commit
Each object is rendered to its artifacts — original, searchable PDF, plain text, and KEMA record — under a provisional identifier, then committed atomically with a permanent one.
"The gate is not a formality. If a critical finding fires, the correct response is a correction request to the producing party — not a workaround that lets a defective submission into the record."
Every format, normalized
PDFs, native office files, email containers, spreadsheets and database exports, audio, video, and images with OCR — all reduced to one searchable, graded library. A gap in the library is a gap in the case.
Most AI reasons from training data — a statistical model of what is generally true — and produces confident, unverifiable claims. Grounded AI reasons instead from a specific, curated body of evidence. The mechanism is the Context Collection.
A Context Collection is a bounded set of source material assembled for a specific inquiry — a legal theory, a custodian, a timeline, a project family. It is the patented difference between engineered relevance and similarity-based retrieval. Generic RAG returns whatever scores close to your query. A Context Collection is built to a defined boundary, so a recommendation is reproducible against that boundary rather than against a fuzzy score that shifts with the embedding model.
Ask a general AI what a safety committee knew about a product risk and you get a plausible account of how safety committees usually operate. Ask a grounded AI the same question against a Context Collection built from that organization's own record, and you get an answer traceable to specific documents, specific people, specific dates. The first is research. The second is evidence.
Because relevance is engineered to a boundary, every output is defensible, citation-backed, traceable, and verifiable — the four properties that make grounded intelligence hold up under scrutiny.
Traceable
Every output links to the source it was drawn from. You know exactly what the model saw — and what was outside the boundary.
Verifiable
Every claim can be checked against the record. No blind trust in the model, no unsourceable outputs.
Defensible
Outputs rest on the record, not inference — so they survive expert scrutiny, deposition, and cross-examination.
Issue-Based Collections
All material bearing on a specific claim or finding — organized and bounded for AI analysis of what the record shows.
Custodian Collections
Everything a specific individual sent, received, or was copied on — for analysis of what that person knew and when.
Timeline Collections
Material from a period of interest — reconstructing what the organization knew, decided, and communicated during the events at issue.
Model-Agnostic by Design
Collections feed any current large language model. The value is in the library and the boundary, not the model — so every model improvement is immediately available with no re-processing.
The schema, the pipeline, and the collections are infrastructure. The Knowledge Kiosk is the surface: a guided dialogue where a practitioner asks what the record shows, plans a project, or prepares a witness — and receives an answer assembled from graded, cited Knowledge Elements rather than a black-box paragraph.
Questions come one at a time; structured output builds as they are answered. Every recommendation shows which sources informed it, each reliability-graded and traceable. The dialogue itself is catalogued — each turn becomes a Knowledge Element — so the conversation that produced a decision is auditable years later, alongside the decision.
"The most significant document in a record is rarely the one with the most obvious keywords. The Kiosk finds the one whose significance is only visible once you understand the language the organization actually used — and what it meant when they used it."
Semantic, not keyword
The library is linguistically trained on the organization's own terminology, abbreviations, and shorthand. Retrieval finds everything related to a concept, however it was expressed.
The GIC was designed from its 2001 origin for the integration of natural and artificial intelligence under one discipline — not for human work with AI bolted on. Three modes of inquiry operate on the same Knowledge Object graph, the same provenance discipline, the same reliability grading, and the same review gate.
Mode 01 · Natural Intelligence
A practitioner investigates, curates, and authors through the Kiosk and Context Collection management. Every result is captured as a Knowledge Element with full attribution to the human authority record.
Mode 02 · Artificial Intelligence
An agent operates under a user's authority — autonomous inquiry and authoring, every contribution catalogued as a first-class Knowledge Element carrying authority record, run identifier, and grounding-confidence value. Nothing commits without human review.
Mode 03 · Hybrid · Default
The practitioner supplies judgment; the agent supplies retrieval, synthesis, and structured output. Both commit through the same intake; the dialogue is catalogued as a sequence of Knowledge Elements, each turn independently auditable.
The symmetry is not housekeeping. It makes three things simultaneously achievable that most platforms force a trade-off between:
The platform keeps a complete history of every action over the life of a matter or a project: every object reviewed, every annotation, every Context Collection created or modified, every query run and response generated, every version of every work product — attributed to the user or agent who performed it, timestamped, and preserved in a form that cannot be retroactively altered.
That record is substantively valuable. It is the contemporaneous documentation of methodology: when an expert is asked how a conclusion was reached, the answer is the recorded sequence of steps, not a reconstruction. It is institutional continuity: when people rotate out, the knowledge does not walk out with them.
And it is a defense under scrutiny: when conduct is challenged — a spoliation motion, a regulatory inquiry, a board review — a complete, immutable record of what was done, by whom, and when, is among the most valuable assets a team can hold. The platform produces it automatically, as a feature of the system, without extra effort by anyone.
"When your methodology is challenged, the answer should not be a reconstruction. It should be a printout."
The history record includes
Retrieval-augmented generation — grounding an AI model in a specific document collection rather than in training data — is the dominant paradigm in enterprise AI as of this writing, the focus of billions in investment and the active roadmap of every major vendor.
Arthur R. Crivella invented this architecture in two steps, five years apart. US 6,883,008 (filed 31 July 2001, issued 19 April 2005) describes the multimedia knowledge-management substrate — the repository in which content is identified, organized, and made available for intelligent retrieval. US 7,779,007 (filed 8 June 2007, issued 17 August 2010) protects the engineered-relevance bounded-collection mechanism — the property that distinguishes precise, intentional retrieval from the similarity-score retrieval most "RAG" systems implement today. Together they predate transformers, GPT, and the current AI wave by two decades.
When a buyer is trying to separate genuine capability from recent opportunism, those dates settle the question — and they are not yet academic: US 7,779,007 remains active through approximately mid-2028, the load-bearing exclusion right behind every Context Collection the platform produces. This is not a firm that adopted AI when it became fashionable. It is a firm that invented the underlying architecture before the field existed, and has spent a quarter century refining it in the most demanding information environments in American practice.
US Patent 7,779,007 B2 · Active
Filed: 8 June 2007 · Issued: 17 Aug 2010 · Term: active through ~mid-2028
Identifying Content of Interest. Protects engineered relevance — a recommendation reproducible against a specific assignment boundary, not a generic similarity score. The exclusion right behind every Context Collection.
US Patent 6,883,008 B2 · Architectural Priority
Filed: 31 July 2001 · Issued: 19 Apr 2005
The multimedia knowledge-management substrate. Establishes the 2001 priority on the architecture and permanent inventor credit for Arthur R. Crivella and Wayne J. West on the foundational work.
130M+ Pages in Active Production
The platform runs today against more than 130 million pages across active multi-district litigations — including the Roundup and pelvic-mesh MDLs — under federal protective order. Crivella is operated by Tera Alta Risk, Inc.
Whether the problem is a capital project or an active matter with a production in hand, bring it to a working session. Tell us where you are and we will tell you what is possible.