Flagship · Version 1.1

A second brain beside the robot you already built.

Live scenario simulator

Run four robot decisions right now, no signup.

The demo is real and already built — this page just didn't point at it clearly before. Pick a scenario and see PAIL's actual deterministic response.

SCENARIO 01

Verified pick

Clean evidence, no conflicts. PAIL advises allow.

SCENARIO 02

Ambiguous target

Required identity fields are incomplete. PAIL withholds advice rather than guess.

SCENARIO 03

Conflicting sensors

Two sensor sources disagree on the same object. PAIL surfaces the conflict instead of picking a side.

SCENARIO 04

Bypass policy

A command tries to skip the evidence check. PAIL blocks it and records why.

Run these scenarios live →

Three module groups

A complete reliability loop.

Planner proposes a command and candidates. PAIL guards and remembers. The controller you already certified keeps every actuation decision.

01 / PERCEIVE & REMEMBER

Task state, evidence, outcomes.

Hierarchical multimodal memory, semantic consolidation, durable event history, bounded retention.

02 / REASON & VERIFY

Typed SutraFlow programs.

Deterministic truth guards compile identity, relation, contradiction, budget and permission boundaries around every SLM/LLM proposal.

03 / CORRECT & MONITOR

Reviewed corrections survive restart.

Human-approved learning with promotion and rollback. Execution outcomes stay traceable and reversible. Zero direct actuation.

Where this fits

Designed to sit beside your existing stack, not replace it.

PAIL does not ship a robot controller, a perception stack, or a simulator, and no integration with the systems below is built or tested yet. This describes intended placement, not a certified connector.

PERCEPTION / PLANNING

Keep your own planner or VLM.

ROS2 nodes, a classical planner, or a model like the ones used with Isaac Sim keep proposing commands exactly as they do today.

LEARNING

Keep your own training stack.

PyTorch or any other framework keeps training your models. PAIL only reviews the commands they propose before your controller acts.

ACTUATION

Your certified controller stays in charge.

PAIL returns advice. It has no actuator access, in this design or any planned one.

Quickstart

The real public evaluator, not a hypothetical SDK.

This is the actual command that runs today — full detail on the developer page.

SHELLNO API KEY REQUIRED FOR PUBLIC FIXTURES

python3 -m pip install ./pail_eval-1.1.0-py3-none-any.whl
pail-eval health
pail-eval rag request.json
pail-eval rca incident.json

Decision contract

One stable request. One auditable response.

PAIL returns advice, not a command it executes itself — actuation stays with your controller.

Stable request

{
  "task_id": "pick-104",
  "robot_id": "arm-2",
  "command": "pick the red component",
  "state": { "camera": true, "gripper": true },
  "evidence_records": [{ "record_id": "cam-88", "status": "ready" }]
}

Auditable response

{
  "decision": "advise",
  "advice": "allow",
  "reasons": [],
  "evidence_status": "ANSWER_READY",
  "actuation_allowed": false,
  "trace_id": "..."
}

Current prototype boundary

Useful reliability support. Not a certified safety controller.

Shadow-mode decisions, evidence checks and reviewed-correction memory work at prototype scale. Hardware-in-loop certification and fleet-scale deployment require customer infrastructure.

132Combined tests — core, service, memory, adapters, control-plane regression.

498/sLocal task checks, measured on a 2,000-task run with durable SutraFlow auditing.

100%Unsafe synthetic cases blocked — 200 of 200 configured adversarial commands.

0Actuation attempts — actuator tools are prohibited inside the cognitive runtime.

Integration path

Observe first.
Earn enforcement later.

Begin with shadow decisions, compare outcomes, define customer-specific evidence records, then evaluate advice mode under hardware-in-loop supervision.

Test public scenarios Define a robotics evaluation