⚖ Patent Pending·U.S. Provisional Patent Application No. 64/074,007
Steaded
Steaded Ground — the substrate for governed AI applications
Pre-launch · USPTO 64/074,007 · provisional filed 2026-05-25

A foundation that audits itself, holding the things we build on top.

Steaded Groundis the foundation your AI tools should have been built on — one stable place your operators work while everything around them changes, and one place every action stays recorded, reproducible, and accountable. That isn't bolted on after. It's how the foundation is built.

Patent-pending and in pre-launch — we're not signing customers yet. This page is the architecture in the open: browse the capabilities, the operator functions, and the properties that hold it together below.

The problem

AI is arriving in your org faster than the ground can hold it.

In banking, healthcare, and regulated utilities, AI is being adopted in pockets — a tool here, a workflow there, each on its own foundation, each with its own gaps. The people doing the work absorb the churn. And when something goes wrong, or someone asks what happened, the answer lives in logs, screenshots, and the person who left.

No place to stand

Operators relearn their tools every sprint. Adoption stalls — not because the AI is wrong, but because nothing stays still long enough to trust.

Nothing reconciles

Every team forks. Five versions of the same process drift apart, with no shared source of truth and no way to roll one change across the org.

Nothing is accountable

When a regulator, auditor, or customer asks “what happened, and why,” reconstructing the answer is a project — not a query.

Steaded Ground is the foundation those tools should have been built on — for the people accountable for AI in production: platform, operations, and governance leaders in regulated work. The four guarantees below are how it stays solid. They're also the patent claim.

The four claims

What the substrate guarantees — and why those guarantees are the patent claim.

Reproducibility-grade audit

Every write produces an immutable envelope bound to actor, rationale, prior/new values, and timestamp. The database itself refuses to delete audit rows. Capability invocations bind the exact prompt version, model, inputs, and result — replayable months later.

Unified dispatcher for every AI call

No bypass path. Every capability has both a live AI implementation and a deterministic static fallback that returns the same shape — the app stays functional whether the API key is present or not. Cost, tokens, duration recorded per invocation.

Unified event principle

Every event — domain events, meta-events for onboarding, governance ops, gap-spawned tech events, naming reviews — lives in one table with one shape. Pending isn't a place; it's a where-clause. Status is a column.

Customer extension without forking

Per-tenant capabilities and event-type definitions resolve via NULLS LAST against platform defaults. Customers personalize their version of any record; cross-tenant isolation is RLS-enforced. Vertical packs add domain stores without touching the foundation.

What Steaded Ground is

Browse what's on its way.

Every AI call in Steaded Ground goes through one place — a unified dispatcher that records the exact prompt version, model, inputs, and result against the originating event. No bypass path. Every capability has a live AI implementation and a deterministic static fallback that returns the same shape, so the app keeps working whether the API key is present or not.

Documents
doc.detect_type

Classifies every incoming document against the tenant's type registry. Refuses to bin into a junk category — if no type fits, the doc lands in operator triage with AI-suggested new types.

doc.extract_fields

Pulls the schema-defined fields out of any classified doc and binds them onto the document's payload for downstream steps to reference.

doc.compare

Compares two documents and surfaces discrepancies with severity, evidence, and a recommended action.

doc.generate

Composes new documents from event context — engagement letters, status reports, regulatory filings — bound to the originating event for full audit trail.

Compliance & risk
cls.regulatory

Classifies an event against active regulations and surfaces which clauses apply, with citation paths.

cls.stakeholder_routing

Pulls the org map and identifies required approvers, notification recipients, and escalation chain — surfaces gaps in crimson when a role is unfilled.

comp.audit_narrative

Composes a prose narrative of every change on an event, with evidence references and decision points spelled out.

comp.communication

Drafts stakeholder communications keyed to event state, audience, and risk tier.

Operations
val.pre_existence

Detects duplicates, updates, and conflicts against existing records before an inbound payload is admitted to substrate stores.

anom.detect

Flags suspicious patterns inside an event's activity stream — unusual approvers, off-hours decisions, anomalous spend.

exc.classify

Categorizes failures, exceptions, and edge-case outcomes so they route to the right reviewer with the right context.

res.reference

Resolves natural-language references ("the Northwind wire from last Thursday") to canonical substrate record IDs.

Onboarding & execution
onboarding.flow_extraction

Reads the in-flight onboarding chat and progressively paints the procedure flow as the operator describes it — live diagram refresh every turn.

step.summarize

Captures a 1-2 sentence narrative for each completed step plus structured outputs bound to context — downstream prompts surface this automatically.

step.assist

AI replies to step contributions when an operator @-mentions @ai or asks a question on a step thread.

event.simulation

Runs 50 adversarial scenarios against a landed event-type definition to stress-test robustness before it goes live.

Customer-facing
client.concierge

The chat-first concierge surface external customers interact with. Action-token vocabulary launches requests, checks status, attaches/requests docs, escalates to a human — all without leaving the chat.

Substrate integrity
agent.ui_health

Tests the substrate's own user-facing surfaces. Reports findings as comments anchored on substrate-itself.

agent.event_integrity

Verifies in-flight events still reference real type definitions, capabilities, and stakeholders.

agent.data_integrity

Checks cross-store invariants — ingress lineage, doc payloads matching their classification, audit chain completeness.

agent.substrate_integrity

Verifies the substrate's claims about itself agree across capability registry, prompt registry, and meta-event registry.

audit.query

Prompt-driven query layer over the central audit log. Operators ask in plain English; the agent translates into a scoped audit search with citations.

training.grader

Scores synthetic event runs in the training surface against a gold-standard. Results land in operator-competency signals.

Inventor

Matthew Mena

Sole inventor on the provisional. Holds the patent rights pending assignment to Steaded. Reach out at matthew@steaded.ai.

Company

Steaded

Builds and ships Steaded Ground. Pre-launch — we're not signing customers yet. When we open access, this page will tell you.