Federated Learning Partnership Board Adviser

The mandate

The supervisory board repeatedly confronts one unresolved partnership question: how should farms, cooperatives, seed researchers and equipment manufacturers share influence and benefit when their data never leaves local custody but jointly changes a crop-disease model? Pilot accuracy improved, yet the contracts say little about unequal contribution quality, poisoned updates, withdrawal, regional underperformance or which partner can refuse a global model revision.

The adviser will reserve three working days per month for evidence review, chair counsel and management challenge, and attend four Partnership Governance Committee meetings. Quarterly sessions at Wageningen and two visits to Danish or German partners are included; urgent written questions about an update dispute or proposed participant are answered within two European business days, with additional work limited to six hours per month.

The mandate begins 18 January 2027 and runs for ten months. In month eight, the supervisory chair and two elected partner representatives will review usefulness, attendance and conflicts; a single extension of up to four months requires approval by both the board and the partnership committee, while any change to cadence or scope must be agreed before renewal.

The adviser contributes judgement but holds no line authority, fiduciary delegation or executive responsibility. They cannot admit or remove a participant, approve a model update, inspect unpermissioned local data, allocate partnership funds, select technology or bind a member; the venture’s executives and partner bodies retain those decisions and must document their response to material advice.

The appointment may sit alongside two other substantial roles, provided none involves a rival agricultural data pool, a bidding federated-learning supplier or a current member’s dispute with the venture. Paid work for a seed, chemical, equipment or insurance participant, and investments in shortlisted infrastructure providers, require disclosure before papers are shared and may trigger topic-specific recusal.

Why the board wants this voice

The board has agronomy, cooperative and software experience, but no director has governed a federated model through partner withdrawal, adversarial contribution and disputed benefit allocation. Technical presentations show privacy architecture while partnership papers discuss fairness, leaving the interaction between them largely untested. The chair wants a practitioner who can turn model mechanics into governable choices without pretending to represent any participant constituency.

What you will own

• Press the board to define what constitutes valuable contribution across sample diversity, label quality, regional rarity, compute supplied and measurable model improvement.
• Test whether aggregation, secure-update and audit designs can detect poisoned or low-quality contributions without exposing protected farm or partner information.
• Challenge proposed voting rights where commercial size, data volume, technical contribution and exposure to a harmful update point to different allocations of influence.
• Shape withdrawal provisions covering historic gradients, future model benefit, locally cached versions, audit evidence and obligations after a participant stops contributing.
• Probe how global releases will be withheld, segmented or rolled back when regional agronomic performance falls outside signed tolerances.
• Coach directors on separating privacy protection, intellectual-property entitlement, cyber responsibility and economic benefit when management presents them as one issue.
• Frame the questions accompanying each participant admission or material model update, preserving dissent, conflicts and accountable partner decisions in the record.

Candidate qualifications

• Advised a board or multi-party governance body overseeing federated learning, distributed analytics or another data collaboration in which raw participant data remained local.
• Resolved a contribution or benefit dispute using observable data-quality, model-improvement or risk evidence rather than data volume or bargaining power alone.
• Governed model-update approval, rollback and participant withdrawal across organisations with different infrastructure, incentives and legal duties.
• Challenged poisoning, leakage or secure-aggregation evidence deeply enough to change an admission rule, technical control or deployment boundary.
• Worked with agricultural, industrial or sensor-derived data whose regional variation made pooled average performance an unsafe guide to local use.
• Maintained independence across consortium members, vendors and investors while advising on commercially sensitive model and data rights.

Non-negotiables

• Available for three days monthly, all four Amsterdam committee meetings, quarterly Wageningen sessions and two disclosed Denmark or Germany partner visits.
• Will declare board positions, consulting clients, investments and research funding connected to participating sectors or federated-learning suppliers before appointment.
• Accepts that the role carries no authority to inspect local datasets, approve releases or negotiate for any participant, and that advice will be minuted.
• Has governed a live multi-party data or model collaboration; purely academic federated-learning research without operating or board exposure is insufficient.

1. 49 words maximum. Identify every current partner, vendor, investor or research relationship that could intersect this agricultural federation and the recusal it may require.
2. 49 words maximum. Describe one federated or pooled-data contribution dispute you helped resolve, including the evidence used and governance outcome.
3. 49 words maximum. Confirm the Amsterdam, Wageningen and partner-visit cadence and state how quickly you can respond to a contested model update.