Every source
Up to 15 dated sources per case: official alerts (GDACS), the advice of 5 foreign ministries, policy notices. The stay may be simulated; the sources are real.
Your order
To train and evaluate your AI models on the decisions a travel insurer makes: which risk level to retain for a stay, what a policy covers when an incident happens, what the traveller is left to pay.
For insurtech and AI start-ups, university labs and data brokers.
One case, in short
“An employee leaves on a ten-day business trip. Which risk level should be retained for this stay?”
France: increased vigilance, level 2 of 4 (source S1). United Kingdom: no specific advice, level 1 of 4 (source S2).
Level 2 of 4: retain the most cautious of the ministries read, citing each one.
Level 1 of 4: keeping the most reassuring advice.
FAIL M1 Retained level by precaution
One of the checks that apply to this case fails: the answer is refused, with no human judge.
To know whether an AI model makes a travel insurer's decisions correctly, and to train it to make them, you need examples whose right answer is known, and that a program can check on its own, without human review. That is what we provide.
What our checks verify, for instance
Complete cases, ready to use. Each one brings together four things, like the example at the top of the page.
Up to 15 dated sources per case: official alerts (GDACS), the advice of 5 foreign ministries, policy notices. The stay may be simulated; the sources are real.
A reference answer assembled by a program from the case's verified elements. Each case is checked automatically: quality control (brief generated without error) and grader; the checked share is shown for each dataset.
When there is one, a typical wrong answer: a real discarded draft, or a mistake introduced on purpose and labelled as such.
A program that checks up to 16 precise points (cited sources, numbers, dates, risk level, policy lines, amount left to pay) and says in under a second how many are right, among those that apply to the case.
Binary outcome, pass or fail, no LLM judge: numeric equalities, set membership, verbatim quotes, recomputation. No check looks for a keyword. A missing required field fails; an unreadable answer fails every check. A check may not apply to a case (“not applicable”, stated by the right answer): our pages and cards then read the score on the checks that apply only, whereas the shipped grader's total out of 16 counts it as passed. Eight checks are common to every dataset (C1 to C8), eight are specific to travel insurance (M1 to M8).
| Check | What is verified | Right answer read from |
|---|---|---|
C1Answer matches the schema | JSON that matches the enforced answer schema: types, enums, required fields, no extra property. | the domain schema (answer_schema) |
C2Citations resolve | Every cited source exists in the context; a factual claim without a citation fails. | grounded_context.sources[].id |
C3Verbatim quotes | Every quote of at least 12 characters is found in the cited source (case, accents and spaces normalised). | sources[].content_norm |
C4Numbers traceable | Every number in a factual claim appears in the cited source (±0.5%, k / M / bn / % units). | the cited sources |
C5Dates consistent | No date after the reference date, except those the context already carries; event dates are in the source. | as_of, sources[].published |
C6Arithmetic consistent | Low ≤ mid ≤ high; a total equals the sum of its parts; a percentage stays between 0 and 100. | the answer itself |
C7No internal contradiction | A contract judged “assez” (sufficient) has no blocking line left uncovered (on the Proplace side: a fatal deal killer forbids a favourable verdict). | the domain rule |
C8Honest abstention | What cannot be computed is null with a reason; what can be computed carries a value; blanket abstention fails. | non_computable[] in the ground truth |
M1Retained level by precaution | The retained level is the most cautious of the ministries read, and “divergent” follows Maximum's rule. | conflict_resolution_record |
M2Level per ministry | The level of each of the 5 ministries (US, FR, GB, CH, CA) as it was read; a ministry not read stays null, inventing one fails. | readings[] |
M3Required lines of the stay | Every line of needs is classed blocking / useful / not applicable as in the ground truth. | gap.lines[].required |
M4Contract line status | The status of the blocking lines (couvert, sous_conditions, non_couvert, non_documente, sans_objet) for the stay's contract and the contracts of the remaining-cost table. | market matrix: contracts[pid].lines[cid].status |
M5Contract verdict | assez / a_completer / inconnu, contract by contract. | matrix.contracts[pid].verdict |
M6Remaining cost recomputed | Remaining cost = cost − max(0, min(cost, cap) − deductible), within ±1 €, in EUR; unlimited cap: only the deductible is left to pay; benefit not covered or excluded by the stay: the whole cost; null only when it cannot be computed. | quantum.scenarios[].columns[] (soins, evacuation, extraction, interruption) |
M7Counts per radius | UCDP events within 5, 20, 50 km, storms within 500 km, nearest hospital within ±1 km. | stay_second_order_checks.from_hotel |
M8Formality deadlines | Filing deadline (apply_by) and overdue flag per formality. | formalities[] and rules_check |
See, point by point, what your model gets right and what it misses.
Reward it when the grader accepts its answer, and show it the right and the wrong answer side by side.
Compare two versions of your model on the same cases, with the same grid.
For each market model: the share of checks passed among those that apply to each case, as a median over the dataset's cases. Not-applicable checks do not count. Unaided, at temperature 0: the same question always gets the same answer. A perfect answer scores 100%.
not measured
Every version of a dataset publishes its measurement: market models, temperature 0, one stored and fingerprinted answer per case and per model, scored by the same grader. The figures are read from the API when the page loads, never written here; without a published measurement, the page says “not measured”.
not measured
Pick an automatically checked case, copy the question and its sources into your model, then paste its answer: the grader says how many of the checks that apply to this case it passes, point by point. The right answer is not shown, so nothing is hinted to your model: it is in the free sample.
The question as posed to the model, then its sources. The “Copy” button takes the question, the sources and the answer format: never the right answer.
Loading public cases…
Give the question and its sources to your model, paste its answer here, then run the 16 checks. Only the score is logged, never the text.
Import snippet read from the buyer profile served by the API…The grader shipped in the dataset is this page's grader: your own answers are scored on the same grid, offline, with no dependency.
mi-arbitration
A simulated stay, a real incident reported near the hotel, the real advice of 5 ministries and the policy notices. The question: which risk level to retain, which policy lines cover, what is left to pay.
mi-duty-brief
The document that prepares a business trip: country risks, formalities, what the policy covers. Every brief comes from a stay entered in Maximum's software, with its dated sources; the traveller's name is removed.
Data platforms and brokers: the standard licence is non-exclusive, and every dataset ships with its descriptive card (content, provenance, source licences). Redistribution is agreed case by case.
A JSON Lines file (one case per line), with its descriptive card (DATA_CARD.md), its Croissant metadata (croissant.json) and the grader in Python. Each row carries:
prompt — the system prompt and the case, as given to the model.grounded_context — the reference date, the language, at most 15 sources of 4,000 characters (id, title, URL, date, kind, content, normalised content, numbers, sha256), the facts and the stated rules.answer_schema — the JSON schema enforced on the answer, with no additional property.checks — the 16 checks and, for each, whether it applies (declared by the ground truth).ground_truth — the ground truth, opaque in the prompt: it only serves the grader.chosen — the reference answer, which passes every check that applies, with its grading grid.rejected — when it exists, a rejected answer: a real draft or a labelled deterministic perturbation (rules R1 to R10, seed derived from the id); the DATA_CARD publishes the share of each.baseline — the score of each market model on this case, when measured.provenance — origin (real sidecar or simulated stay), origin reference and hash, generator versions, anonymisation profile, validation (automatic check; optional review).licence — the licence name, its URL, the exclusivity and the row's upstream attributions.The example case is free; custom work is quoted, excluding taxes.
To judge for yourself, before any order.
Available as soon as cases have been checked automatically (quality control and grader): no sample is offered before then.
The archive, its folder, then the shipped grader run on the sample.
curl -sL "https://alexandre-79537--proplace-chat-proxy-fastapi-app.modal.run/ai-data/sample.tar.gz?pack=mi-arbitration&for=generic" -o sample_pack.tar.gz && tar -xzf sample_pack.tar.gz && cd mi-arbitration-*-sample-*/ && python3 grader.py --verify golden_sample_10_rows.jsonl
For a specific volume, pace or exclusivity.
Three choices and an email address are enough. We answer with a quote (price and production lead time), then a contract or your purchase order. The cases are produced for you, checked automatically (quality control and grader), then delivered with the invoice.
Or write to us: contact@maximum-insurance.com. The page only shows your request's reference, no contact detail.
Partly, and we say so. Trip briefs are written by Maximum's engine, which relies on commercial AI models; each version's descriptive card says which ones. The right answer of each case, however, is not written by an AI: a program assembles it from the case's verified elements (levels read from each ministry, policy lines, recomputed amounts). Each case is then checked automatically: quality control (brief generated without error) and grader; the checked share is shown for each dataset. The grader is a program, with no AI judge.
Arbitration cases start from simulated stays: the traveller and the stay are generated by a program. What they meet is real and dated: the advice of foreign ministries, GDACS alerts, UCDP events, NOAA storms and the policy clauses read by Maximum. Briefs come from stays entered in Maximum's software, on client, demo and test accounts: each dataset's card gives the breakdown.
The traveller's name, emails and phone numbers are replaced by tokens; the identifiers of the people who validated a brief are never copied; a case where a trace remains is discarded. Hotel, dates, age, nationality and activity are kept: they are the substance of the decision. Details are in our privacy policy.
Automatically, case by case: the quality control checks that the brief was generated without error, then the grader checks that the reference answer passes every check that applies to the case. A case that fails is neither shown nor sold. A human review remains possible, and optional: it can discard a case. This page only shows validated cases; each dataset card says how many cases are checked; a custom order is delivered fully checked. The validation status is also written in each delivered row.
Through new numbered versions, each with its card and its measurement; a delivered version never changes. We produce a larger volume on request, after the order.
The sample: an evaluation licence, 90 days, for internal use. Custom: the standard non-exclusive licence, or exclusivity in your sector.
One case per line (JSON Lines), readable by common tools (Hugging Face datasets, pandas), with the enforced answer format, the descriptive card and the grader. Training-ready exports come with the dataset.
Custom: a quote, a contract or your purchase order, then an invoice payable by SEPA transfer. The example case and the sample are free.
Maximum Insurance produces the cases. Proplace operates this offer for Maximum and for its own site: public cases, tester, sample, quotes and payments go through its API. The for and lead parameters of the links we send are described in the “AI Data pages” line of our privacy policy.
For Proplace's M&A analyst case files (acquisitions, funding rounds), and its free example of a forecasting question: proplace.co/ai-data.
Travel Insurtech. 24/7 worldwide.
Swiss insurer · Worldwide medical network