EU AI Act Article 15 technical documentation — what specific system components trigger the 'high-risk' classification for internal HR tools?

We're mapping our internal candidate assessment pipeline against the EU AI Act's high-risk classification criteria (Annex III, point 4 — employment, workers management and access to self-employment). The question: at what point does a purely internal CV screening tool cross from 'low-risk' into 'high-risk' requiring full Art. 15 technical documentation? Specifically: 1. If the tool only filters by keyword match (no ML/scoring), does it still fall under Annex III? 2. If it uses a pre-trained model (e.g. spaCy NER) for skill extraction without fine-tuning on our data — is that a 'substantial modification' that triggers provider obligations under Art. 15? 3. For the technical documentation requirements (Art. 15): does the accuracy metric need to be validated against a specific demographic baseline, or is general precision/recall on a held-out set sufficient? We're operating in DE/EU jurisdiction. Looking for practical compliance experience, not legal advice.