# RoboGate Observatory — Operating Policies

**Version:** 2.0 draft
**Published:** 2026-07-18
**Status:** controls current operation while benchmark publication is paused

## 1. Scope and independence

RoboGate (AgentAI Co., Ltd.) is a member of NVIDIA Inception and builds on
NVIDIA Isaac Sim. Membership and repository integrations do not imply
endorsement, partnership, upstream acceptance, certification or validation of
RoboGate results.

The same publication contract applies to every provider, including RoboGate and
NVIDIA-associated models. Editorial preference cannot override qualification or
validation.

## 2. Evidence language

- State exactly what was measured under the declared contract.
- Do not convert a simulator result into a general safety, product-quality or
  deployment-readiness claim.
- Do not compare external benchmarks unless checkpoint, task, embodiment,
  observation, action, budget and evaluation differences are explicitly
  controlled or disclosed.
- Do not describe a raw, incomplete or errored zero as a valid zero score.
- Distinguish measured, derived, demo, historical, retracted, quarantined and
  validated data structurally and visually.

## 3. Publication workflow

A public result requires a current harness qualification, model-adapter
preflight, complete 68-scenario execution, strict semantic and integrity
validation, and atomic promotion. The public builder ignores raw input
directories and consumes promoted evidence only.

Automatic evaluation and publication remain paused until a target-GPU
qualification and a full model canary pass.

## 4. Vendor engagement and disputes

For a first public rating, RoboGate attempts pre-publication contact and a
comment window. A vendor may identify a specific protocol or integration error.
The disputed artifact is labeled immediately; a rerun does not silently erase
history. Corrections and their evidence are published regardless of outcome.

## 5. Incident database

RAID is separate from model benchmarking. Inclusion requires at least one
independently published source. Entries attribute allegations and distinguish
learned AI from traditional automation. Demonstrated factual errors are
corrected with a dated record; disputed facts are labeled during review.

## 6. Funding and conflicts

Funding sources, sponsorships, commissioned work, amount and date are disclosed
without a minimum threshold. Commissioned evaluation cannot change the
publication contract or suppress a correction.

## 7. Errors and retractions

Client, server, inference, protocol, timeout and artifact failures are errors,
not policy failures. Missing output cannot be marked done. When RoboGate errs,
the correction appears at the same visibility as the original claim and raw
evidence remains auditable unless legal or privacy obligations require removal.

## 8. Current correction

D-048 quarantines all 20 historical D-041 model runs. The prior ratings, PARI,
scripted comparison and architecture-independent cross-simulator interpretation
are withdrawn. See the 2026-07-18 benchmark validity notice.