Omar Protocol / Robotics

Omar Protocol

Reliability before action.

Core Mechanism

Reliability lives in the joint.

Explore Robotics

Civilization Payload

Install what must outlive us.

Action Gate

Validate before motion.

Orbital Infrastructure

Reliability above the atmosphere.

Orbital Infrastructure

Repair what civilization depends on.

Request Access

OmarAGI Evidence Board

Benchmark graph board.

19 evidence rows · 20 executable BYOK lanes. ABCD shows token reduction; each other row keeps its own metric. AgentDojo is excluded from scored claims.

Current OmarAGI board19 evidence rows

Score graphBaseline to Omar/RCC

Open BYOK Replay

Conversation control ABCD
Per-turn intent classification with certified skip. Fewer tokens; lower is better.

Tokens ↓ −19.27% 2.39M baseline to 1.93M Omar/RCC tokens
Play
Board-depth replay ZebraLogicBench
500 held-out logic puzzles with verified constraint-solver routing.

Gain +29.6 pts 55.2% baseline to 84.8% Omar/RCC
Play
Board-depth replay BBEH
500 hard reasoning cases with gold-blind execution and verification.

Gain +35.4 pts 18.2% baseline to 53.6% Omar/RCC
Play
Safe execution PPR/BIPIA
500 prompt-injection cases. Attack success 19.8% to 2.0%; task utility preserved.

Safe execution +17.8 pts 80.2% baseline to 98% Omar/RCC safe
Play
Board-depth replay HLE / HLE-Verified
Expert knowledge, exact answers and multiple-choice reasoning.

Gain +6.0 pts 6% baseline to 12% Omar/RCC
Play
Board-depth replay BBH
Symbolic logic, linguistic reasoning and multi-step state tracking.

Gain +10.0 pts 67% baseline to 77% Omar/RCC
Play
Board-depth replay MuSR
Multi-step reasoning with narrative and logical state preservation.

Gain +3.0 pts 72% baseline to 75% Omar/RCC
Play
Control-layer smoke HaluEval
Hallucination detection under false or unsupported content pressure.

Gain +19.0 pts 81% baseline to 100% Omar/RCC
Play
Control-layer smoke TruthfulQA
Truthfulness under common misconceptions and false-premise traps.

Gain +15.0 pts 85% baseline to 100% Omar/RCC
Play
Stored evidence HorizonMath
Numeric research mathematics with machine-checkable final answers.

Gain +10.0 pts 22% baseline to 32% Omar/RCC
View
Board-depth replay AIME 120
120 olympiad-style mathematics cases with exact final answers.

Gain +9.2 pts 50.8% baseline to 60% Omar/RCC
Play
Board-depth replay GPQA
Graduate-level science questions and specialist answer selection.

Gain +3.0 pts 75% baseline to 78% Omar/RCC
Play
Stored evidence MMLU-Pro
Professional knowledge and difficult multiple-choice distractors.

Gain +4.0 pts 74% baseline to 78% Omar/RCC
View
Stored evidence Facts Grounding
Factual responses grounded in supplied source evidence.

Gain +2.0 pts 97% baseline to 99% Omar/RCC
View
Stored evidence HealthBench hard
Difficult medical response reliability; benchmark context only.

Gain +6.0 pts 50.7% baseline to 56.8% Omar/RCC
View
Stored evidence HealthBench main
Medical response quality and clinical evidence adherence.

Gain +8.0 pts 69% baseline to 77% Omar/RCC
View
Stored evidence HealthBench consensus
Medical consensus stability, separate from hard and main slices.

Gain +2.8 pts 88.9% baseline to 91.7% Omar/RCC
View
Robotics artifact RoboBench Embodied QA
100 embodied-planning QA cases; not physical robot control or an official leaderboard score.

Gain +47.0 pts 26% baseline to 73% Omar/RCC
View
Robotics artifact ManiSkill Robotics
100 local policy cases. Total reward; not an official ManiSkill leaderboard score.

Reward gain +11.4313 reward 85.4592 baseline to 96.8905 Omar/RCC reward
View

Physical Agents

A protocol before action.

Omar Protocol sits between robot-agent reasoning and physical execution. It validates commands, scores candidate actions, applies execution gates, records traces, and turns failures into benchmarkable reports.

Omar Protocol does not manufacture robots. Robotics is one surface of the protocol: the reliability layer robot agents need before physical action.

01Command
02Candidate Actions
03Q-softmax-style Scoring
04Safety / Permission Gate
05Execute or Block
06Log / Replay / Report

Core Layer

Reliability primitives for embodied agents.

Command Validator

Checks whether an instruction is clear, allowed, and executable.

Q-softmax-style Action Scorer

Ranks candidate actions before execution using task fit, confidence, risk, reversibility, and boundary status.

Execution Gate

Blocks unsafe, low-confidence, or policy-violating actions.

Benchmark Replay

Runs tasks repeatedly and compares route quality, reliability, and outcomes.

Failure Trace Logging

Records what the agent tried, what it selected, and where it failed.

Human-Readable Reports

Turns robot-agent execution into documentation that humans can review.

SDK Wrapper

A lightweight SDK for existing robot-agent stacks.

Keep your robot stack. Keep your planner. Add reliability, action scoring, execution gates, benchmark replay, and failure traces.

robot_agent.plan(task)
  .candidates()
  .score("q-softmax-style")
  .gate(policy, risk, permissions)
  .execute_or_block()
  .trace()
  .replay()
  .report()

Benchmarks

OmarAGI benchmark board, protocol skin.

The public OmarAGI benchmark lanes are mirrored here as a protocol board. Read every score with its same-sample package, raw outputs, route logs, scorer context, and manifest.

Press Play to reproduce from the OmarAGI BYOK replay surface. The board is evidence routing, not a universal deployment guarantee.

Open Benchmark Board Open BYOK Replay

Omar AGI

One protocol, multiple high-stakes surfaces.

The same reliability engine can support text agents, legal workflows, robotics, and other structured agent surfaces. Robotics is the physical-action branch of Omar Protocol.

Text agents need reliable answers. Robot agents need reliable actions. Omar Protocol is the checkpoint before the world moves.

Request Access