QWALORA
AI transparency

Where AI helps, and where it is never allowed.

Ranger is deliberate about AI. Qualification is never decided by AI. Every qualifying outcome is governed by deterministic, independent verification and sealed with a signature. This page documents how AI is and is not used, which providers are under evaluation, and the human authority that governs every decision.

ai-governance // statusGOVERNED
Candidate-facing inferenceNone
Qualification decisionDeterministic, signed
Training on candidate dataDisabled
human authorityfinal on every outcome
Approved uses and governance

What AI may do, and what it may never do.

AI is a staff tool, governed by policy and kept away from every qualifying decision. Candidates receive no AI assistance in assessment.

  • AI may assist staff with drafting challenge specifications, analyzing scenario diversity and difficulty, and advisory review after a submission is made.
  • AI cannot determine qualification, alter a signed score, override verification, or coach a candidate during an active challenge.
  • Candidates receive no AI assistance in assessment. There are no walkthroughs, no hints, and no generated solutions.
  • Candidate submissions are not used to train models. Protected challenge truth, answer keys, and evaluator logic are never exposed.
ai-governancestatus
candidate-facing inferencenone
open-weight eval (approved)under evaluation
qualification decisiondeterministic · signed
model authenticity scoresadvisory, non-authoritative
training on candidate datadisabled

No specific model is claimed operational until it is deployed and independently verified.

Model & provider disclosures

Provider-independent, and honest about status.

Ranger is evaluating open-weight model inference for internal engineering use only. It is provider-independent: Hugging Face is one prospective inference integration, not the sole technology underpinning any pathway. No model is in production for assessment. Each integration is tracked as proposed, validated, or operational.

Hosted inference Hugging Face and alternative hosted-inference providers, evaluated for internal engineering use. One prospective integration among several, not the definition of any pathway. Proposed
Candidate models Open-weight models such as Qwen and gpt-oss, selected by benchmark against the task. No default model and no single provider. Proposed
Adversarial AI frameworks Microsoft PyRIT, NVIDIA garak, Promptfoo and Inspect AI are candidate evaluation frameworks. They are not deployed assessment components. Proposed
Signed-evidence plane The Ed25519 signed-evidence system that seals and re-verifies outcomes. Operational on the platform today. Operational
Proposed Validated Operational Status reflects integration maturity, not candidate progress or any vendor certification. No framework is listed as deployed until that is verified.
Model evaluation methodology

Benchmark first, adopt only what a challenge needs.

Models and frameworks are selected by benchmark against the specific task, not by brand. No component is adopted to claim breadth. Anything used in assessment must produce reproducible evidence the signing plane can seal.

STEP 01

Task-specific benchmark

Each challenge defines capability, failure surface, cost, latency and reproducibility criteria. Candidates are ranked against it.

STEP 02

Reproducibility gate

A component advances only when it produces evidence that recomputes from recorded inputs and maps to the evidence schema.

STEP 03

Independent re-measurement

Outcomes are re-measured and signed, never self-reported, and verification remains deterministic and human-governed.

Data handling & candidate privacy

Candidate data stays the candidate's.

Submissions and candidate data are handled for assessment and verification only, and are never used to train models.

  • Candidate submissions are not used to train or fine-tune any model
  • Protected challenge truth, answer keys and evaluator logic are never exposed to candidates or to inference providers
  • Evidence is retained for verification and audit; access is authorized and logged
  • Interest and application data is used only to contact and review, with consent, and is kept separate from assessment
Human authority

A person is accountable for every outcome.

AI never qualifies a candidate. Authority rests with people, and the signed record makes every decision auditable.

  • Qualification decisions are deterministic, human-governed and signed
  • Model outputs, where used, are advisory and non-authoritative
  • Admission is selective and decided by administrators, not automation
  • Every outcome is independently re-measured and can be re-verified

Questions about how we use AI?

Transparency is a differentiator, not a disclaimer. If you need detail for a security or procurement review, reach the team directly.

See the verification methodology Contact the team