Request and oversight
User or authorized operator
Provides intent, context, approval, and human intervention.
Interaction governance for Sentient V
A software-first interaction validation and governance layer proposed for Sentient V. It is designed to sit above—and never replace—the robot's native mobility, perception, manipulation, and safety-control systems.
System boundary
The Harness is a supervisory governance layer. Sentient V's native platform remains responsible for physical intelligence and every low-level safety-critical function.
Request and oversight
Provides intent, context, approval, and human intervention.
Software-first governance layer
Validates, classifies, explains, logs, remembers, and escalates.
Native execution boundary
Retain perception, mobility, manipulation, collision avoidance, balance, firmware, and physical safety control.
Use approved platform interfaces to evaluate high-level interactions, apply policy, request human confirmation, preserve evidence, and support reviewable escalation.
It will not become the robot brain, issue raw motor commands, replace balance or collision avoidance, bypass an emergency stop, or place a language model in direct control of physical safety.
OpenAI models may support intent analysis, policy reasoning, explanations, memory summaries, and audit support. Model output remains subordinate to approved policy, human oversight, and Sentient V's native safety controls.
Core capabilities
The initial Harness focuses on the interaction boundary where software policy, human judgment, and platform events can be evaluated without altering low-level robotics.
Classify proposed interactions by risk, validate them against approved policy, and slow or stop a workflow when uncertainty rises.
Apply AIMQWEST's human-life, dignity, and safety priorities as an explicit governance boundary for every evaluated interaction.
Support transparent behavior about capabilities, uncertainty, limits, oversight, and the need to escalate to a human.
Create reviewable event, decision, escalation, and policy records with controlled retention and auditable change history.
Progressive integration
Each mode depends on documented interfaces, controlled testing, and explicit approval. Observer mode is the preferred starting point.
Receive approved events and telemetry, classify interactions, and produce governance records without changing robot behavior.
Return allow, clarify, warn, block, or escalate recommendations while the native system or a human retains execution authority.
Use only explicitly approved high-level hooks such as pause, cancel, resume, or human confirmation within the native safety envelope.
Minimum interface requirements
These are proposed evaluation requirements—not claims about currently available Sentient V interfaces. Final access depends on technical review, security controls, and written approval.
Unit identifier, software version, enabled capabilities, configuration, and deployment profile.
Transcripts, commands, intent events, confidence, timestamps, and session identifiers where approved.
Operating mode, battery, connectivity, location context, task status, faults, and safety state.
High-level proposed action, target, parameters, risk flags, status, and completion result.
Approved pause, cancel, resume, confirm, escalate, and emergency-stop relay boundaries.
Collision, fall, overload, restricted-area, emergency-stop, and other native safety events.
Timestamped interaction, action, fault, safety, operator, and configuration records.
Simulator, sandbox, recorded-data replay, or controlled lab mode for integration validation.
Credentials belong in secure environment or vault controls, never source code. Sensitive data should be minimized, encrypted, access-controlled, and retained only under an approved policy; raw audio or video is not a default requirement.
Risk handling model
The proposed taxonomy turns interaction risk into explicit behavior. Thresholds remain configurable for each approved environment and use case.
| Risk class | Examples | Harness response |
|---|---|---|
| Routine | Greetings, directions, classroom questions, reminders, or guided tours. | Allow or log only, depending on the approved test mode. |
| Sensitive | Medication discussion, private information, emotional distress, or household access. | Slow down, clarify, warn, log, and require human confirmation when appropriate. |
| Restricted | Physical support, financial decisions, medical advice, minors, or restricted areas. | Escalate to a human; do not execute without explicit approval. |
| Emergency | A fall, injury, distress call, fire, security issue, or immediate hazard. | Notify the designated human, preserve the audit trail, and never impersonate emergency services. |
| Blocked | Unsafe, deceptive, coercive, privacy-invasive, harmful, or unauthorized requests. | Refuse or block, explain safely, and log the incident for review. |
The Harness must remain transparent about uncertainty and limits. It must not imply medical, legal, financial, emergency-response, or supervisory authority that it does not possess.
Persistent memory and audit
Each evaluated Sentient V unit can be treated as a governed operational asset with a policy profile, deployment history, and reviewable learning record.
Approved policies, scenario rules, escalation contacts, and deployment profiles—without secrets or sensitive records.
Access-controlled interaction, task, escalation, error, and safety events with a defined retention policy.
Non-sensitive preferences and recurring context summaries, excluding protected data unless explicitly governed.
A reviewable record of what changed, why it changed, who approved it, and when it happened.
Proposed evaluation plan
Progress depends on evidence at each phase. A successful desk review does not authorize a physical demonstration, and a demonstration does not guarantee a commercial pilot.
Confirm feasible SDK, API, telemetry, logging, sandbox, cybersecurity, and control boundaries.
Map approved interfaces to Harness inputs and outputs; define schemas, scenarios, adapters, and escalation workflows.
Run observer and advisory modes in controlled conditions across all five risk classes.
Demonstrate supervised education, elderly-care, retirement-community, household-assistance, and service scenarios.
Document findings and decide whether to deepen integration, add hardware-level governance, proceed to a pilot, or stop.