For European funders and public agencies
Evaluation panels
you can defend
Every application is read separately by each member of an AI panel, every score comes with a written reason, and the whole round exports as one record you can hand to an auditor — so your assessors spend their time where it actually matters.
- Residency policy enforced on every call
- Every step timestamped and on the record
- On-premises or air-gapped deployment
The decision record
A decision record is the whole assessment of one application: what each panel member scored, where they disagreed, who approved it and when, and the reasoning in writing. This is the artefact itself, not a dashboard about it.
Illustrative example — a synthetic application, shown in the real record's structure
Strengthening local democracy through youth councils
Reference GR-26-047 · Application round 2026 · 1–6 scale · Five AI panel members
Scores by criterion
Relevance to the call
Feasibility of the plan
Organisational capacity
Expected impact
Where the panel disagreed
Four panel members scored organisational capacity 4; the fifth scored 2, citing an unaudited financial statement for the previous year. A spread of 2 meets this round's threshold, so the application was flagged for a person and entered the review queue before the round closed.
Human approval
Approved by the round's assessor, who saw the divergence, the dissenting member's written reason, and the application itself alongside every score.
Written justification
The plan is concrete, names its partner schools and sets a realistic timeline, and its relevance to the call is direct. Organisational capacity is the weak point: last year's financial statement is unaudited and the projected staffing assumes a funding decision by June. Recommended for funding on condition that an audited statement is filed before any payment is made.
Decisions that get questioned
An evaluation platform for organisations whose decisions are read by auditors, boards and appeals bodies.
Auditable deliberation
Every panel run records each step's outcome — passed, degraded or failed — and pauses for human approval where you require it. When a residency policy sends a run to a different model, the audit row keeps the model that was asked for, the model that ran, and the reason they differed.
Residency enforced per call
Each organisation picks one rule for where its data may go — anywhere, the EU preferred, the EU only, or never off your own machines. Every model call, search and outgoing message is checked against that rule first, and anything it forbids is stopped rather than quietly sent elsewhere. Run in our EU cloud, on your own premises, or fully air-gapped.
A process per round
Rubrics, scales and processes change year to year. Each round owns its own process and records which round it was copied from, and it surfaces the applications that most need human attention.
The last word is human
Every application leaves a panel run by one of three routes, and none of them is “the model decided”.
Agreed — and still sampled
Where the panel agrees and the application sits clear of the funding line, its draft stands. A fixed share of these is routed to a human anyway. That share defaults above zero, and the round records the rate it used.
Disagreement goes to a person
Four things send an application to a person: a wide spread between panel members, weak agreement, a criterion only one member could score, or a rank close to the funding line. Any one of them puts the application in the review queue — and where the round is set to pause, the run stops there until someone looks.
Below the line — and still sampled
Applications the panel places below the funding line carry the same oversight sample as the confident ones. Being ranked out is not a reason for nobody to look.
This follows Article 14 of the EU AI Act on human oversight, and the Swedish authority IMY's position that a confident model is not a reason for nobody to look. It is enforced in the software rather than promised in a policy document: the share of confident cases sent to a person starts above zero, and an automated check blocks any release that would quietly remove it.
How it works
From rubric to defensible decision in three steps
Define the round
Set the rubric, scoring scale and criteria — your own, not a template's.
Assemble the panel
Each panel member takes one of the rubric's roles, on cloud or local models.
Run and review
Applications are scored, debated and justified. The panel pauses for human approval where you require it, and every step's outcome is on the record you export.
The same rubric, three ways
Write your assessment scheme once — the questions, the scale, the weights. Three different jobs then run from it.
Score a round
Upload a round's applications as one archive and scoring starts. Each criterion keeps its score, its weight, the reason written for it, and the model that produced it.
Draft against it
Write a short brief and get a first draft of every section, built from your own documents. It is also how you test a process before a real round: draft a strong and a weak application, and see how your panel scores them.
Advise against it
Point the same panel at a draft instead of a decision. Every question that falls short comes back with why it did, and what to change to lift it — ranked by the points still on the table.
Drafting and advice run inside your organisation, for your own staff. There is no applicant-facing portal yet — an applicant cannot be sent a link of their own.
Book a walkthrough
Bring your own rubric and a handful of real applications. We will run them in front of you, show you the record that comes out the other end, and answer whatever your procurement will want to ask.