General-Purpose AI Technical Documentation Expert — EU Evidence Dossier

The mandate

The provider has released fourteen model and fine-tune variants, but architecture records, training provenance, compute and energy calculations, evaluation results and downstream limitations cannot be reliably matched to immutable release identifiers. The defined problem is to make those evidence chains complete and maintainable, not to redesign the models or supply a legal opinion.

The named deliverable is a version-controlled GPAI Evidence Dossier mapped to Article 53 and the applicable information in Annexes XI and XII of the EU AI Act. It includes authority-facing technical documentation, a downstream-provider information pack, an evidence trace matrix, controlled redaction rules and a lifecycle procedure for keeping both documentation sets current.

Work starts on 7 December 2026. Milestone one on 23 December is the approved scope map and evidence-gap register; milestone two on 29 January 2027 is the populated dossier and downstream pack for two representative releases; milestone three on 26 February is the remediated portfolio framework, witnessed retrieval exercise and final operating procedure.

Acceptance occurs when the Chief Product Officer can reproduce every sampled model claim from release-tagged source evidence, General Counsel approves the documented trade-secret treatment, and independent regulatory counsel finds no unexplained high-priority gap in its sample. The full dossier must be retrievable inside twenty-four hours, with every residual omission assigned to an accepting executive.

The client will provide repository access, release manifests, experiment and evaluation records, data-provenance artefacts, compute and energy evidence, incident material, product documentation and a secure authoring environment. Engineering and legal sponsors will nominate evidence owners and decide unresolved legal interpretations; delayed or withheld source material will enter the formal exception log.

Why this is external work

The engineers understand individual artefacts, but no internal owner spans the complete chain from training run to released model and downstream disclosure. The compliance team can interpret obligations yet lacks the technical depth to test whether a documented claim is reproducible. A specialist external build also gives the sponsors a politically neutral account of missing evidence.

What you will own

• Define the dossier schema linking model identity, architecture, training process, data characteristics, computational resources, evaluation evidence, limitations and release controls.
• Reconcile model cards, repository tags, experiment trackers and deployment manifests to decide which artefacts describe each released or materially modified version.
• Compile authority-facing technical files with source citations, evidence owners, freshness dates, confidentiality classifications and explicit treatment of unavailable information.
• Produce a downstream-provider pack explaining intended tasks, integration requirements, input and output characteristics, capabilities, limitations and applicable acceptable-use boundaries.
• Trace every material evaluation statement to executable method, dataset version, run configuration, result file, reviewer decision and released model hash.
• Rehearse a time-bound information request and record retrieval failures, inconsistent claims, redaction disputes and remediation ownership before final acceptance.
• Deliver the update procedure, responsibility matrix, change triggers, approval workflow, retention logic and controlled templates needed for subsequent model releases.

Candidate qualifications

• Prepared EU general-purpose AI technical documentation, downstream-provider information or an equivalent regulator-facing model dossier for a foundation-model provider.
• Interpreted model architecture, training, fine-tuning, evaluation and deployment records deeply enough to challenge engineers on missing or non-reproducible evidence.
• Linked compute-resource and energy calculations to identifiable training runs rather than relying on aggregate infrastructure estimates.
• Protected trade secrets while producing documentation usable by authorities and downstream integrators with different information entitlements.
• Led a mock regulatory information request or evidence-room exercise that exposed and closed material model-lifecycle documentation gaps.
• Built a controlled documentation system that remained current across model changes, releases, fine-tunes and retirement decisions after the initial project ended.

Non-negotiables

• The proposed lead must personally author and defend the evidence architecture; a legal-only team or delegated documentation factory will not be accepted.
• Paris presence is required three days each week during discovery and for both final acceptance weeks, with secure remote work between milestones.
• The engagement excludes legal opinions, copyright remediation and model redesign, and the consultant must preserve those scope boundaries.
• Any concurrent instruction from a direct foundation-model competitor or retained AI-systemic-risk assessor must be disclosed before contracting.

1. 49 words maximum. Describe how you matched technical evidence to a specific model release when repositories, model cards and deployment manifests used different identifiers.
2. 49 words maximum. Which Annex XI evidence field have you found hardest to substantiate, and what source artefact ultimately supported it?
3. 49 words maximum. Explain your first-three-week method for separating a genuine evidence gap from documentation that is merely stored under another owner.