Reference framework: EASA Concept Paper — Guidance for safety-related artificial intelligence applications, Proposed Issue 03 (June 2026), supporting EASA AI Roadmap 2.0 and RMT.0742 – https://www.easa.europa.eu/en/downloads/143701/en

This is an Abbreviated Version of this Positioning Paper.

For the Full Version of the Paper in PDF format, please send an email request to info@amrisaviation.com or to rickadams@aviationvoices.com

1. Executive summary

Amris Competency Management System (originally Amelia) is an AI-based pilot-competency and predictive-safety platform used to capture, structure and analyse evidence-based training data under the EASA Evidence-Based Training (EBT) framework. It comprises an assessment engine (ORCA — Observe, Record, Classify, Assess) that scores the nine EASA competencies from multi-modal session evidence, and a predictive intelligence layer that estimates a calibrated, explainable probability of attritional safety events at the level of pilot, competency and Observable Behaviour.

This paper positions Amris with the EASA Concept Paper (Issue 03) — EASA’s anticipated framework for approving safety-related AI. The central message is straightforward: Amris CMS is a ground-based, human-augmentation application (EASA Level 1) in which a qualified instructor or training manager remains fully in command of every decision. That classification keeps Amris in the most tractable region of the framework, and Amris’s architecture — structured data provenance, calibrated outputs, competency-level explainability, immutable audit trail, and Just-Culture data handling — already reflects the intent of the four EASA trustworthiness building blocks.

2. Characterisation of the AI application

Under the EASA framework, trustworthiness analysis begins with a characterisation of the AI application, its function, and its Operational Design Domain (ODD). Amris CMS is characterised as follows:

• Nature — a cloud-native, ground-based training and safety-support platform. It is not an airborne system and performs no airworthiness or real-time control function.
• Purpose — to augment human instructors and training managers by structuring training evidence, scoring EASA competencies, and surfacing competency drift before it becomes an incident. Outputs are advisory.
• Operational Design Domain — commercial air-transport operators running EBT / mixed-EBT programmes on defined aircraft types and fleets, within an approved training organisation’s management system. Each model’s applicability is bounded to the operators, fleets and phases on which it was validated.
• AI functions — two AI functions.
  • (1) ORCA — a data-driven assessment engine mapping audio, instructor notes and flight data to competency grades and Observable Behaviours.
  • (2) A predictive risk model — a Bayesian latent-skill estimator feeding a calibrated, explainable classifier that outputs attritional-incident probability over a defined horizon.
• Decision authority — the qualified instructor, evaluator or Training Data Coordinator. Amris CMS never revalidates a licence, closes remediation, or exercises training privileges; it informs a human who does.

3. Classification under the EASA AI levels

The EASA Concept Paper classifies applications by the degree of automation and the division of authority between human and AI. Amris CMS sits deliberately at Level 1 — assistance to human — and does not enter Level 2 or Level 3.

EASA level	Definition	Relevance to AMRIS
Level 0 — Low automation	Information acquisition/analysis with no link to decision-making.	Below Amris: Amris CMS explicitly supports human decisions.
Level 1A — Human augmentation	AI augments the human’s perception/analysis; human decides and acts.	Primary fit: ORCA evidence structuring, competency trend surfacing, cohort analytics.
Level 1B — Human support in decision & action selection	AI proposes options/rankings; human selects and remains accountable.	Fit: predictive attritional-risk flags and targeted-training recommendations.
Level 2 / Level 3 — Cooperation, collaboration, advanced automation	AI shares tasks with, or acts ahead of, the human.	Not applicable: Amris CMS never acts autonomously on safety decisions.

Positioning consequence. Because Amris CMS is a Level 1 human-augmentation tool, the depth of guidance EASA applies — governed by the criticality- and classification-based proportionality in Chapter D of the Concept Paper — is modest relative to airborne or autonomous applications. Amris CMS carries no Development Assurance Level obligation of a flight-critical system; the demonstration centres on data quality, output trustworthiness, explainability and human oversight, all of which are core to the product design.

4. Mapping to the EASA trustworthiness building blocks

The EASA AI Roadmap identifies four building blocks essential to trustworthy AI. Trustworthiness analysis is always required in full; the other three are scaled by classification and criticality. The table below maps each block to Amris CMS’s current design.

Building block	EASA intent	AMRIS alignment
Trustworthiness analysis	Characterisation, safety/risk assessment, information security, ethics-based assessment, and continuous (in-service) risk assessment. The gate to all other building blocks.	Amris CMS characterises each AI function, its Operational Design Domain (fleet, aircraft type, EBT programme) and ConOps as a ground-based, advisory decision-support tool. Because outputs inform — never command — a qualified instructor or Training Data Coordinator, the residual safety contribution is bounded. A non-punitive, Just-Culture data policy and role-based de-identification address the ethics dimension directly.
AI assurance	Learning assurance (the W-shaped process), data management, and development/post-ops explainability — replacing classical development assurance for the data-driven parts of the system.	Amris CMS is trained and validated on structured ORCA evidence (200M+ observation evidences, 100k+ competency grades). The predictive risk model is specified with an explicit data-management and validation strategy: time-based, pilot-disjoint and operator-disjoint splits, calibration of output probabilities, and counterfactual sanity checks — mirroring the intent of the learning-assurance objectives.
Human factors for AI	Human-centred design, AI operational explainability, human-AI cooperation, and error management so the human retains authority and situational awareness.	Every Amris CMS output is explainable at the level the instructor already works in: Shapley contributions aggregate to the 9 EASA competencies and to Observable Behaviours, with traceability back to source evidence (audio, notes, flight data). The instructor observes, records, classifies and assesses; Amris augments that workflow rather than replacing judgement.
AI safety risk mitigation	Additional mitigation means required only for advanced automation (Level 2B / Level 3) where human oversight is reduced or delegated.	Largely out of scope at Amris’s classification: a qualified human remains fully in command of every training, revalidation and remediation decision. Amris therefore does not rely on the Level 3 mitigation means, which simplifies the compliance demonstration and lowers approval risk.

5. Explainability and human oversight — the core of the story

EASA treats operational explainability and human oversight as decisive for Level 1 systems: the human must understand the AI output well enough to accept, question or override it. Amris CMS is engineered for exactly this.

• Explained in the regulator’s own language — every predictive output decomposes, via Shapley values, into contributions aggregated to the nine EASA competencies and to individual Observable Behaviours — the same vocabulary instructors are trained and standardised in.
• Traceable to evidence — each grade and flag traces back to source evidence (in-browser audio, structured instructor notes, flight-data linkage), with an immutable, timestamped audit trail and assessor identity, satisfying the training-records and data-integrity expectations of ORO.FC.231 and the EASA guidance.
• Calibrated and validated — probabilities are calibrated so a stated risk means what it says, and models are validated on pilot- and operator-disjoint splits to demonstrate generalisation rather than memorisation.
• Human-in-command by design — Amris CMS supports the ORCA workflow (Observe-Record-Classify-Assess); the human performs the assessment and owns the outcome. This preserves the division of authority EASA requires at Level 1.

6. Net safety benefit

The Concept Paper extends the net-safety-benefit concept as a basis for accepting safety-related AI. Amris’s value proposition is expressed directly in those terms: by structuring the richest pilot-performance signal in the industry and surfacing competency degradation before it manifests as an attritional event, Amris CMS enables targeted, evidence-based intervention in place of blanket retraining. The intended effect is a measurable reduction in attritional incidents — the low-severity, high-frequency events that precede most serious occurrences — with the human decision-maker firmly in the loop. This is a safety-positive contribution delivered without ceding authority to the machine.

7. Positioning statement and forward path

Amris CMS is positioned as a trustworthy, EASA-aligned Level 1 AI application: a ground-based decision-support platform whose outputs are advisory, explainable in EASA competency terms, traceable to evidence, calibrated, and governed by Just-Culture data handling under an operator’s approved management system. It sits in the most approvable region of the EASA framework and already embodies the intent of the trustworthiness building blocks. To convert alignment into a defensible compliance demonstration as RMT.0742 matures, Amris will maintain: a living trustworthiness analysis and ODD definition per deployed model; a documented learning-assurance and data-management dossier (data provenance, validation splits, calibration and monitoring); an in-service monitoring regime for continued risk assessment and model drift; and an ethics-based assessment covering non-punitive use, de-identification and role-based access. Andy O’Shea’s chairmanship of the EASA Aircrew Training Policy Group and Amris Aviation’s EBT operating heritage give Amris CMS direct line of sight to the evolving guidance.