| --- |
| license: cc-by-4.0 |
| task_categories: |
| - text-classification |
| language: |
| - en |
| tags: |
| - evaluation |
| - rubrics |
| - inter-annotator-agreement |
| - llm-as-a-judge |
| - benchmark-annotation |
| pretty_name: AI Summit Barcelona 2026 — Rubric Agreement Workshop |
| size_categories: |
| - n<1K |
| --- |
| |
| # AI Summit Barcelona 2026 — Rubric Agreement Workshop |
|
|
| Every evaluation filed during a live 80-minute workshop at AI Summit Barcelona on |
| 22 September 2026, in which the room scored one expert-authored benchmark item exactly the |
| way an annotator does it on [EuroExec](https://sovrano.ai/euroexec/) (arXiv:2608.04549), |
| Sovrano's expert-graded benchmark of frontier models on European executive decisions. |
|
|
| Attendees read one executive task and its graded checklist, then blind-scored three |
| frontier-model answers: five dimensions from 1 to 5, then Hit / Partial / Miss on every |
| checklist row, then a ranking and a written rationale. Which model wrote which answer was |
| hidden until the end, and the A/B/C labels were the same on every screen, so one attendee's |
| numbers sit next to another's. |
|
|
| Sovrano's own evaluation of the same three answers is published alongside, which is what |
| makes this comparable: `scorings.jsonl` carries each attendee's Cohen's κ against that |
| expert key, over the same Hit / Partial / Miss calls. On EuroExec, two blind scorers below |
| κ 0.7 halt production of an item. |
|
|
| ## What's in it |
|
|
| | File | Rows | One row is | |
| | --- | --- | --- | |
| | `scorings.jsonl` | 16 | One attendee's finished evaluation, with their agreement against the expert key | |
| | `dimension_scores.jsonl` | 240 | One attendee scoring one response on one dimension, 1–5 | |
| | `coverage.jsonl` | 240 | One attendee's Hit / Partial / Miss call on one checklist row of one response | |
| | `responses.jsonl` | 3 | One of the three answers under test, with the model that wrote it | |
| | `checklist.jsonl` | 5 | One row of the graded checklist, and whether it is critical | |
| | `dimensions.jsonl` | 5 | One of the five scoring dimensions, as it was shown | |
| | `golden.jsonl` | 3 | Sovrano's expert evaluation of one response, with a note per checklist row | |
|
|
| The task all three answers respond to: |
|
|
| > Act as a senior product strategist for a B2B subscription SaaS Platform providing healthcare referral tracking system using AI in a startup that just received seed funding of €1M. You have piloted the product among 50 hospitals in Germany for a one-time trial fee of €2500. From July onwards you would like to start billing all 50 customers €250 monthly as a subscription fee. However, your customer support team has reported that user reviews have been increasingly overwhelming and unsatisfactory and that there are at least 30% 1 star (very bad) ratings coming in. The board expects evidence-based recommendations before subscription billing begins. Your task is to create a product strategy report based on customer feedback and subscription launch readiness. |
| > Diagnose the business situation identifying at least 5 risks to customer journey, subscription sign ups and product adoption. Design a review analysis process capable of reviewing 50 reviews per day. The process must categorize per theme, its frequency, calculate severity and distinguish feature requests from defects. Define a measurement framework containing the KPIs and metrics you come up with. Do not invent perfect outcomes, explicitly identify uncertainties and additional data that should be collected before scaling. |
|
|
| ## Caveats that matter |
|
|
| - **37 unfinished evaluations are not here.** A scoring is exported once |
| the attendee submitted it, which the app only accepts when every dimension is scored, |
| every checklist row is marked and a rationale is written. Partial work was left out rather |
| than published as if it were a considered call. |
| - **Scoring was not blind.** The workshop app holds nothing back: the model behind each |
| answer and the expert key are on the results screen from the start, and an attendee could |
| read them before scoring. Most did not — the screens lead you through scoring first — but |
| this data cannot be used as evidence about blind annotation, and the figures here should |
| be read as what a room of non-specialists produced with the key available. |
| - **The expert key is a small sample of one.** It is Sovrano's grading of three answers |
| against five rows. κ against it says how closely a reader applied the same checklist, not |
| who was right. |
| - **Attendees are not expert annotators.** They are conference attendees working at speed, |
| which is the point of the comparison with EuroExec, not a defect to correct for. |
| - **Contributors are pseudonymous.** `contributor_id` is a random per-browser identifier. |
| No personal data appears in this dataset and nothing in it can be traced back to a person. |
| Attendees gave their details to the organisers on the way in; those were stored |
| separately, were never linked to a contributor id, and are not part of this release. |
|
|
| ## Licence |
|
|
| CC BY 4.0. Please cite this dataset and EuroExec if you build on it. |
|
|
