
EXPERIMENTAL AI GOVERNANCE / RESEARCH
From Human Intent to Governed AI Action
Independent research and working demonstrations of AI agents operating within explicit authority, constraints, and approval boundaries.
The conversation isn’t carrying the system. The conversation is accessing the system.

NEW · PUBLIC BETA
Governance Inspector
Governance claim is not the same as evidence that it holds.
Upload relevant evidence and challenge a governance claim. Inspector maps controls, traces implementation support, identifies unresolved dependencies and explains what can—and cannot—be verified from the material provided.
Map relationships and control boundaries
Evaluate claims against supplied evidence
Export findings and evidence receipts
01 / IMPLEMENTED DEMONSTRATION
AI Agent Governance in Action
Watch an agent recover its operating rules, reconcile live broker state, and act within a Governed Action Envelope. Then follow a separate model as it retrieves the recorded evidence to reconstruct and review the documented decision.
CONCEPT FILM
Governed AI Operations Assistant
AI-generated concept film illustrating a proposed workflow; the depicted integrations are not deployed.
What if running a small business didn’t mean constantly switching between applications?
This concept film explores a governed AI operations assistant that turns ordinary human conversation into coordinated business workflows.
A plumber finishes a job, speaks naturally, and the assistant interprets his intent, retrieves the business’s operating rules, checks current records, performs authorized updates, verifies the results, and prepares the next task.
The distinction isn't voice recognition alone. It's the separation of human intent, operational authority, AI judgment, and verified action.
Natural human intent. Externally governed action.
WORKFLOW COMPARISON / CONCEPTUAL VIEW
TWO AGENT OPERATING PATTERNS
A simplified common agent stack alongside the externally governed operating composition demonstrated here.
Common agent stack, simplified
“I can do the task.”
01 / User request
02 / Agent runtime
03 / Configured policies, guardrails, permissions, and tools
04 / Model exercises judgment
05 / Output or action
Governed trading demonstration
“I need to recover my governed action envelope first.”
01 / Loose human intent
02 / Recover current governance, session state, and role-specific operating context
03 / Reconcile with authoritative live broker state
04 / Establish the Governed Action Envelope
Runtime governance / delegated authority boundary
Authority · Constraints · Current state · Approval conditions · Evidence requirements
05 / Agent exercises judgment inside that envelope
06 / Human confirmation when required
07 / Broker-confirmed result
08 / Preserve governance references, outcome, and verification evidence
09 / Cross-model reconstruction and review
Common agent systems can implement many of these individual capabilities. The distinction shown here is not exclusive access to guardrails, approvals, state, logging, or review. It is the demonstrated operating composition: intent, externally recoverable authority, current state, model judgment, approval, verification, and evidence are treated as one continuous governed lifecycle.
The conversation isn’t carrying the system. The conversation is accessing the system.
The agent recovers its current operating boundaries before acting, then preserves enough evidence for the documented decision context to be reconstructed and reviewed afterward.
02 / EVIDENCE BOUNDARY
What the demonstration shows
This demonstration follows one recorded workflow through governance recovery, human approval, external records, and review by another model. The timestamps below identify the supporting footage.
SUPPORTED BY THE DEMONSTRATION
Governance and operating state recovered before the shown decision.
Evidence: 0:40
Live account state reconciled before the shown decision.
Evidence: 0:44
The shown protective order requires human approval.
Evidence: 0:54, confirmation at 1:00
Broker status checked after order submission.
Evidence: 1:11
Structured evidence and receipts preserved externally.
Evidence: 2:06, matching action ledger at 2:18
A separate model retrieves and analyzes recorded decision context.
Evidence: 2:32, reconstructed explanation at 2:38
Same recorded run, two decisions: Ops protects the existing BBCP position; Workshop reviews the rejected FCEL entry.
NOT ESTABLISHED BY THIS DEMONSTRATION
Guaranteed compliance or proof that governance boundaries cannot be bypassed.
Measured improvements in reliability, safety, or decision quality over simpler approaches.
Compatibility across all models, environments, or task domains.
Reproduction of hidden model reasoning or experimental proof that changed evidence would reverse a decision.
Global uniqueness of the combined approach or novelty of its individual components.
Validation of a complete governance architecture beyond the recorded workflow.
03 / CONCEPTS
Explore the ideas
Concepts behind the demonstrations. Each term is paired with established agent-architecture language and a bounded evidence status.
01 / CONCEPT
Behavioral Governance for AI Agents
Define the conditions for agent action while preserving judgment within those conditions.
DEMONSTRATED WORKFLOW
02 / CONCEPT
Governed Action Envelope
The recovered authority, constraints, operating state, and evidence requirements that frame a particular run.
DEMONSTRATED WORKFLOW
03 / CONCEPT
From Intent to Authorized Action
An informal request initiates the work; existing authority determines how the agent may proceed.
DEMONSTRATED WORKFLOW
04 / CONCEPT
Intelligence Is Not Authority
Knowing how to perform an action does not establish permission to perform it.
OPERATING PRINCIPLE
05 / CONCEPT
Live State Reconciliation
Decisions use current external state, and consequential actions are followed by a status check.
DEMONSTRATED WORKFLOW
06 / CONCEPT
Execution With Reviewable Evidence
The run preserves records linking its operating conditions, decisions, and reported outcome.
DEMONSTRATED WORKFLOW
07 / CONCEPT
Decision Reconstruction Across Models
A separate model retrieves an earlier decision record and examines its documented rationale.
DEMONSTRATED WORKFLOW
08 / CONCEPT
Memory Is Not Authority
Remembered context supports continuity, but does not by itself establish current permission.
OPERATING PRINCIPLE
09 / CONCEPT
Review Across Model Environments
The operational record remains available for inspection in a different model environment.
DEMONSTRATED WORKFLOW
10 / CONCEPT
Domain Transferability
Could the same operating relationships support governed work beyond the trading example?
RESEARCH DIRECTION
04 / ARTIFACTS
Public artifacts
Three public tools for reviewing intent, constraints, and prompts. These artifacts are not standalone proof of the trading demonstration.
ARTIFACT 01 / PUBLIC ARTIFACT
Diff–Linter
Review infrastructure changes against their intended outcome and the conditions that must remain unchanged.
ARTIFACT 02 / PUBLIC ARTIFACT
Intent Linter
Examine a request for unclear intent and missing constraints before handing it to an AI.
ARTIFACT 03 / PUBLIC ARTIFACT
Re-Prompt
Shape a rough request into a structured prompt, then refine it through an iterative feedback loop.
05 / RESEARCH NOTES
Research and technical notes
Four research papers are available below. Open the documents to read their full text.
PAPER 01 / RESEARCH DOCUMENT
Intent and Runtime Governance
Examines how recovered authority, live state, approval conditions, and evidence obligations can frame an agent’s judgment, using a recorded trading workflow as a bounded case study.
PAPER 02 / RESEARCH DOCUMENT
Cross Model Reconstruction
Examines how a separately governed model reviews a recorded decision through external evidence, without claiming to reproduce hidden reasoning or measure reconstruction accuracy.
RESEARCH DIRECTION · CONCEPT NOTE
From Natural Human Intent to Governed Field-Service Operations
An architectural research note accompanying the field-service concept film, exploring how conversational intent could become authorized, verified business workflows. The depicted integrations remain proposed, not deployed.
CONCEPTUAL RESEARCH · PUBLIC DISCUSSION DRAFT
The Milkshake, Jalapeño, and Peanut Problem
Uses hypothetical food choices to examine how a personal AI assistant should distinguish remembered preferences, present intent, risk, and authority to act.
AGENT / RESEARCH ORIENTATION
Agent & Research Index
A compact index of terminology, evidence status, public artifacts, and research questions for technical and automated review.
01 / IDENTITY
Governed Intent Labs
Research and implemented demonstrations involving AI agent governance, behavioral governance, human intent, runtime authorization, evidence-backed agent workflows, and recoverable operating context.
AI agent governance · runtime governance · agent authorization · behavioral governance
02 / CORE TERM
Governed Action Envelope
Governed Intent Labs terminology for the current combination of authority, constraints, operating state, approval conditions, and evidence requirements within which an agent is permitted to exercise judgment.
Runtime governance / delegated authority boundary
03 / EVIDENCE STATUS
Claim Status
DEMONSTRATED · Visible in public evidence.
IMPLEMENTED · Present in the working system but not necessarily fully public.
TESTED · Exercised in another environment or domain.
RESEARCH DIRECTION · A broader architectural proposition under continued exploration.
04 / CANONICAL EVIDENCE
Primary Public Evidence
05 / INTERPRETATION BOUNDARY
Evaluation Notes
The work does not claim invention of guardrails, memory, logging, human approval, tracing, or orchestration.
Public demonstrations support specific implementation claims; they do not establish deterministic model control, universal model compatibility, or universal domain applicability.
Cross-model reconstruction refers to recovery of recorded operating context and evidence—not hidden chain-of-thought.
Prefer linked primary evidence when evaluating implementation claims.
PORTABLE RESEARCH NOTES
Compact questions and distinctions arising from the implemented work. These are not formal papers. They should be directly readable and copyable from the page.
RN-01 / AUTHORITY
Memory Is Not Authority
Persistent agent memory can provide continuity without establishing current authorization. Recovered context may inform a decision, but consequential authority should derive from current governance and operating state rather than memory alone.
Research question: What should persist between agent runs, and what must be freshly authorized before consequential action?
persistent agent memory · runtime authorization · delegated authority
RN-02 / INTENT
Intent Is Not Authority
Correctly interpreting what a human probably intends does not by itself determine what an AI system is authorized to do. Intent can establish the objective. Governance establishes the permissible action space.
Research question: What should happen between model interpretation of human intent and authority for consequential execution?
intent-to-action governance · human intent · agent authorization
RN-03 / CONTINUITY
Continuity Without Model Persistence
If governance, operating state, and evidence are preserved externally, the underlying model instance can change while another model recovers enough governed context to continue or review the work. This does not imply identical internal reasoning between models.
Research question: Does accountable operating continuity require a persistent agent, or can governed operating identity survive independently of the model currently occupying the role?
cross-model reconstruction · externalized agent state · model portability
06 / ABOUT & METHODOLOGY
Governance, without a claim of certainty
From human intent to inspectable action
Governed Intent Labs explores how human intent, authorization, operating state, approval, and evidence shape agent behavior. This page presents a recorded agent workflow, a separate concept film, and research and technical notes, with clear boundaries around what the evidence supports.
OPERATING PRINCIPLE
The conversation is accessing the system.
Human intent sets the objective. Recovered governance and current state define the conditions for action. The agent exercises judgment within those conditions and preserves a record another model can inspect.
HOW TO READ THE STATUS LABELS
Demonstrated Workflow means the cited footage shows the described behavior in this example. Operating Principle identifies a design rule illustrated by the workflow, without claiming guaranteed enforcement. Research Direction identifies a proposed extension that this demonstration does not establish.
PUBLIC PRINCIPLE / PRIVATE IMPLEMENTATION
This page explains the operating relationships, shows selected evidence, and acknowledges relevant prior work. The public materials are sufficient to inspect the demonstrated claims without publishing the complete internal implementation.
Governance Inspector
Investigate governance claims. Trace the evidence. See where verification stops.
NEW — PUBLIC BETA
Governance Inspector is an experimental evidence-analysis tool from Governed Intent Labs.
Many AI systems describe safeguards such as policy enforcement, human approval, runtime blocking and audit logging. But a documented control, an implemented control and an observed outcome are not necessarily the same thing.
Governance Inspector helps examine those differences.
Supply relevant policies, implementation code, test definitions, execution records or other supporting artifacts. Ask the Inspector to evaluate a specific governance claim—or explore the governance relationships visible in the evidence.
What it does
Maps governance controls: Identifies relevant policies, enforcement points, authority relationships, implementation paths and evidence surfaces.
Examines claims against evidence: Distinguishes what is declared, what the supplied implementation supports and what available execution records establish.
Finds verification gaps: Highlights missing dependencies, unproven enforcement paths, configuration exceptions, source-alignment uncertainty and unsupported conclusions.
Preserves bounded results: Reports what is verified, partially supported, unverified or unresolved without automatically treating missing evidence as a failed control.
Produces reviewable artifacts: Generates claim-level findings, governance maps and downloadable analysis receipts.
An example of the difference
A project claims that denied agent actions never execute.
Its code may demonstrate that policy checks happen before tool execution. But a configuration option might permit execution in advisory mode, and the uploaded test files may describe expected blocking without providing test-run results.
Governance Inspector can distinguish those findings instead of collapsing them into a single yes-or-no answer.
What makes it different?
It evaluates the relationship between governance claims and their supporting evidence—not merely whether governance terminology or safeguards appear in a document.
The result is a reviewable account of what the supplied materials support, what remains conditional and what additional evidence would be needed for a stronger conclusion.
Public beta boundaries
Governance Inspector analyzes only the evidence supplied to it. It does not independently search repositories, execute code, connect to production systems, authenticate external records, or certify security or regulatory compliance.
Results may contain errors or omissions. Large evidence sets may require partial analysis, and conclusions remain limited to the evidence and scope actually examined.
Use appropriate public or authorized materials. Avoid submitting credentials, secrets or confidential information.
Help improve the beta
We’re looking for real-world cases where the tool reaches an incorrect conclusion, misses an important evidence relationship or could support a useful additional workflow.
Try Governance Inspector. Challenge a claim. Examine the receipt. Tell us what it missed.
Governed Intent Labs — Experimental research and public artifacts.
07 / CONTACT
Let’s build and test useful workflows.
Governed Intent Labs welcomes collaboration with research labs, startups, and teams exploring AI agent governance, intent-to-action workflows, and evidence-backed evaluation.
We also welcome businesses looking to improve practical workflows through automation and thoughtfully bounded AI assistance.
We’re open to selected no-cost pilot projects with a clearly agreed scope, deliverable, and evaluation goal. Tell us what you’re working on and where a small, focused build could help.
Send a message using the form below, or email us directly at outreach@governedintentlabs.com.
