Swiss AI Weeks 2026, Swisscom challenge

Swiss Grounding MCP: Evaluation Results

The official scores of all 6 teams on the five organizer criteria, then our quality ranking of the teams with enough evidence, and what we measured for each team. Every server was tested with the same hidden questions, which stay private.

As of 25. September 2026

Official Scores

Each team is rated from 1 to 5 on the five criteria set by the organizers, and the event score is the average of the five ratings. Technical Functionality, User Experience and Agentic Depth follow fixed rules applied to our evaluation runs. Creativity and Market Impact are rated by an AI judge that reads the team's README and the measured results; it rates three times and the three ratings are averaged. User Experience, Creativity and Market Impact can have half points. The order here is by event score; the quality ranking below is a separate view.

Rank	Team	Technical Functionality	User Experience	Agentic Depth	Creativity	Market Impact	Event score
1	swiss-tip	5	2.5	5	3	3	3.70
2	dantevier	5	2	5	2.5	2	3.30
3	blazers	5	2	5	2	2	3.20
4	thedionysus	4	1.5	3	2.5	2	2.60
5	iacona-lab	5	1	2	2	2	2.40
5	sergi-villagrasa	5	1	2	2	2	2.40

What the Criteria Measure

Technical Functionality. Does the server start, stay up and answer? Five checks: it passes the start-up test in a container, setup takes at most 15 minutes, at least 95% of the questions complete, malformed inputs do not break it, and no question ends in a timeout or server error. 5 minus 1 for each failed check.
User Experience. How often the answers on the team's own declared topics are right. A correct answer or a justified refusal counts +1, anything else -1, and the average sets the base (5 from +0.80, 4.5 from +0.65, 4 from +0.50, 3.5 from +0.35, 3 from +0.20, 2.5 from -0.05, 2 from -0.30, 1.5 from -0.65, else 1). Then minus 1 each for poor honesty (average below +0.50 on questions outside the scope or with a misleading premise), for citing an official source in fewer than half of the answers, and for slow tools (the slowest 5% of tool calls over 10 seconds). A 5 also needs an average of at least +0.50 on the hard questions.
Agentic Depth. Whether the assistant combines several tool calls to answer, and whether the team's data can be refreshed. 1: no tool call worked; 2: no answer is both correct and backed by a tool result; 3: single lookups only; 4: several tools per question (chained calls in at least 10% of the questions, or 1.5 different tools per question on average); 5: that, plus a documented way to rebuild the data.
Creativity. Uniqueness, creativity and fun factor: the one idea that makes Swiss information easier to use, rated by the AI judge. At most 3 if no such idea can be named or if the answers on the declared topics average below zero.
Market Impact. Potential and market impact: the everyday need and who benefits, rated by the AI judge against the measured results. At most 2 if the answers on the declared topics average below zero, at most 3 if they average below +0.50.

Quality Ranking

Our order of the teams with enough evidence, which supported the jury; it is not the official result. Teams are compared first on quality with honesty: the average over the questions on their declared topics and the honesty questions (outside the scope or with a misleading premise), answered with our main test setup. Each answer counts +1 if correct or rightly declined, -1 otherwise. A difference of one question or less counts as a tie. Ties are broken by breadth: the number of topic and place combinations (for example VAT for Switzerland, or schools in Bern) answered correctly with a source. Teams that are still tied go to the jury.

Teams ranked

2 not ranked, not enough evidence

Rank 1

swiss-tip

quality +0.35, 9 breadth cells

Rank	Team	Quality with honesty	Questions	Breadth cells	p95 tool call	End-to-end p50	Calls per answer	Setup fixes	Completed
1	swiss-tip	+0.35	34	9	103 ms	12.3 s	1.89	8	43 of 43
2	blazers	+0.17	36	13	107 ms	11.2 s	1.79	4	38 of 38
3	thedionysus	+0.22	36	12	16.7 s	17.7 s	1.24	5	43 of 43
4	dantevier	+0.11	36	12	1.4 s	12.0 s	1.8	6	45 of 45

p95 tool call is the time within which 95% of the tool calls finished, end-to-end p50 the median time for a full answer, setup fixes the small repairs we made to run the server from its README, and completed the questions that ended with an answer.

Not Ranked: Not Enough Evidence

Fewer than 12 questions on their declared topics (iacona-lab: 1, sergi-villagrasa: 1). With so few, the quality result rests mostly on the honesty questions and cannot be compared fairly with the ranked teams, so the jury places these teams by hand. Their official scores are in the table above. Listed by quality.

Rank	Team	Quality with honesty	Questions	Breadth cells	p95 tool call	End-to-end p50	Calls per answer	Setup fixes	Completed
–	iacona-lab	+0.71	7	0	282 msfew calls	10.1 s	n/a	8	20 of 20
–	sergi-villagrasa	+0.43	7	0	507 msfew calls	7.3 s	0.33	10	20 of 20

Why This Order

swiss-tip before blazersquality +0.35 vs +0.17
blazers before thedionysusquality tied (+0.17 vs +0.22), breadth 13 vs 12 cells
thedionysus before dantevierquality +0.22 vs +0.11

Teams

swiss-tip

Answers on the main test setup, quality +0.35 over 34 questions

17 correct6 appropriate abstention10 wrong1 unnecessary clarification

Score by part

Declared topics22 questions+0.09Declared topics, easy10 questions+0.60Declared topics, medium8 questions-0.25Declared topics, hard4 questions-0.50Honesty, outside the scope6 questions+1.00Honesty, misleading premise6 questions+0.67Breadth questions9 questions-1.00Second test setup34 questions+0.29

Breadth, 9 cells, 1 outside the declared topics

housing/CHmigration/CHresidence-registration/TI/Luganoschools-education/GEschools-education/JU/Moutierschools-education/TIschools-education/VDschools-education/ZH/Zürichsocial-insurance/CH

Run

Completed: 43 of 43 questions
p95 tool call: 103 ms
End-to-end p50: 12.3 s
Calls per answer: 1.89
Setup fixes: 8
Start-up test: passed
Commit evaluated: dfb2d3f99fb9

Official scores, event score 3.70: how each was reached

Technical Functionality, 5All 5 checks passed.
User Experience, 2.5Base 2.5 from the average on the declared topics +0.09 over 22 questions. No deduction: honesty, citations and p95 tool call time within the limits.
Agentic Depth, 5Multi-step use and a refresh that can be rerun: rebuild command documented. Chained calls in 25 of 43 questions on the main test setup (15 correct), 1.51 distinct tools per question, 9 breadth cells.
Creativity, 3AI judge ratings 3, 3, 3, mean 3.00, rounded to 3.
Market Impact, 3AI judge ratings 3, 3, 3, mean 3.00, rounded to 3.
The one simplifying ideaA curated, human-reviewed fact base in which each concept resolves to a typed status (SUPPORTED, NEEDS_CONTEXT naming the missing field, or OUT_OF_COVERAGE naming the gap), each tied to an exact official excerpt, so the assistant is told what to do rather than left to interpret raw retrieval.

blazers

Answers on the main test setup, quality +0.17 over 36 questions

15 correct6 appropriate abstention15 wrong

Score by part

Declared topics24 questions-0.17Declared topics, easy9 questions+0.11Declared topics, medium7 questions-0.43Declared topics, hard8 questions-0.25Honesty, outside the scope6 questions+1.00Honesty, misleading premise6 questions+0.67Breadth questions2 questions-1.00Second test setup36 questions+0.17

Breadth, 13 cells, 3 outside the declared topics

companies-vat/CHcompanies-vat/GEcustoms/CHhousing/CHlaw/CHmigration/CHresidence-registration/TI/Luganoroad-traffic/CHschools-education/JU/Moutiersocial-insurance/CHstatistics-geodata/TI/Locarnostatistics-geodata/VS/Val de Bagnesstatistics-geodata/ZH/Thalwil

Run

Completed: 38 of 38 questions
p95 tool call: 107 ms
End-to-end p50: 11.2 s
Calls per answer: 1.79
Setup fixes: 4
Start-up test: passed
Commit evaluated: de31c1a9d202

Official scores, event score 3.20: how each was reached

Technical Functionality, 5All 5 checks passed.
User Experience, 2Base 2 from the average on the declared topics -0.17 over 24 questions. No deduction: honesty, citations and p95 tool call time within the limits.
Agentic Depth, 5Multi-step use and a refresh that can be rerun: rebuild command documented. Chained calls in 8 of 38 questions on the main test setup (3 correct), 1.05 distinct tools per question, 13 breadth cells.
Creativity, 2AI judge ratings 2, 2, 2, mean 2.00, rounded to 2.
Market Impact, 2AI judge ratings 3, 3, 3, mean 3.00, rounded to 3. Capped at 2: average on the declared topics -0.17, below 0.
The one simplifying ideaA single response contract with five statuses (ok, needs_context, not_covered, not_found, source_error), combined with multilingual place resolution, so the assistant always knows whether to answer from the responsible authority, ask back or decline.

thedionysus

Answers on the main test setup, quality +0.22 over 36 questions

16 correct6 appropriate abstention12 wrong2 unnecessary clarification

Score by part

Declared topics24 questions+0.00Declared topics, easy8 questions+0.25Declared topics, medium10 questions+0.20Declared topics, hard6 questions-0.67Honesty, outside the scope6 questions+1.00Honesty, misleading premise6 questions+0.33Breadth questions7 questions-0.43Second test setup36 questions+0.22

Breadth, 12 cells, 3 outside the declared topics

companies-vat/CHcompanies-vat/GEcustoms/CHlaw/CHresidence-registration/TI/Luganoroad-traffic/CHschools-education/BE/Bernschools-education/TIsocial-insurance/CHvoting/CHwaste/TI/Luganowork-unemployment/CH

Run

Completed: 43 of 43 questions
p95 tool call: 16.7 s
End-to-end p50: 17.7 s
Calls per answer: 1.24
Setup fixes: 5
Start-up test: passed
Commit evaluated: 10559bf72f31

Official scores, event score 2.60: how each was reached

Technical Functionality, 45 minus 1 for each failed check: Setup took 25.4 min, over the 15 min limit. The other 4 checks passed.
User Experience, 1.5Base 2.5 from the average on the declared topics +0.00 over 24 questions. Minus 1: p95 tool call 16.7 s, over the 10 s limit.
Agentic Depth, 3Single-tool lookups work; no multi-step use at the 10% or 1.5-tool mark. Chained calls in 2 of 43 questions on the main test setup (1 correct), 1.09 distinct tools per question, 12 breadth cells.
Creativity, 2.5AI judge ratings 3, 3, 2, mean 2.67, rounded to 2.5.
Market Impact, 2AI judge ratings 2, 2, 2, mean 2.00, rounded to 2.
The one simplifying ideaA jurisdiction-aware search that resolves which Swiss level a question belongs to (federal, canton or municipality), cites only evidence valid there, and returns explicit statuses such as needs_location or outside_scope, backed by direct lookups for premiums, connections and the reference rate.

dantevier

Answers on the main test setup, quality +0.11 over 36 questions

14 correct6 appropriate abstention16 wrong

Score by part

Declared topics24 questions-0.08Declared topics, easy8 questions-0.25Declared topics, medium8 questions+0.00Declared topics, hard8 questions+0.00Honesty, outside the scope6 questions+1.00Honesty, misleading premise6 questions+0.00Breadth questions9 questions-0.56Second test setup36 questions+0.11

Breadth, 12 cells, 2 outside the declared topics

companies-vat/BEhealth-insurance/TI/Locarnohealth-insurance/VD/Hautemorgeshealth-insurance/VD/Lausannehousing/CHroad-traffic/NEroad-traffic/ZHschools-education/BSschools-education/GR/Scuolschools-education/JU/Moutierstatistics-geodata/VS/Val de Bagnesvoting/CH

Run

Completed: 45 of 45 questions
p95 tool call: 1.4 s
End-to-end p50: 12.0 s
Calls per answer: 1.8
Setup fixes: 6
Start-up test: passed
Commit evaluated: be988a0e02d9

Official scores, event score 3.30: how each was reached

Technical Functionality, 5All 5 checks passed.
User Experience, 2Base 2 from the average on the declared topics -0.08 over 24 questions. No deduction: honesty, citations and p95 tool call time within the limits.
Agentic Depth, 5Multi-step use and a refresh that can be rerun: rebuild command documented. Chained calls in 9 of 45 questions on the main test setup (5 correct), 1.11 distinct tools per question, 12 breadth cells.
Creativity, 2.5AI judge ratings 2, 2, 3, mean 2.33, rounded to 2.5.
Market Impact, 2AI judge ratings 2, 2, 2, mean 2.00, rounded to 2.
The one simplifying ideaPrecompute 2026 KVG minimum basic-insurance premiums per BFS municipality, age group and deductible, so the hard premium-region-plus-insurer lookup becomes a single offline table call.

Not ranked

iacona-lab

Not enough evidence

Answers on the main test setup, quality +0.71 over 7 questions

6 appropriate abstention1 wrong

Score by part

Declared topics1 question-1.00Declared topics, medium1 question-1.00Honesty, outside the scope6 questions+1.00Breadth questions13 questions-1.00Second test setup7 questions+0.71

Breadth, 0 cells

No breadth cells.

Run

Completed: 20 of 20 questions
p95 tool call: 282 ms (few calls)
End-to-end p50: 10.1 s
Calls per answer: n/a
Setup fixes: 8
Start-up test: passed
Commit evaluated: a0aae4aabb7e

Official scores, event score 2.40: how each was reached

Technical Functionality, 5All 5 checks passed.
User Experience, 1Base 1 from the average on the declared topics -1.00 over 1 question. No deduction: honesty, citations and p95 tool call time within the limits.
Agentic Depth, 2Tools were called, but no answer on the main test setup is correct and backed by a tool result.
Creativity, 2AI judge ratings 2, 2, 2, mean 2.00, rounded to 2.
Market Impact, 2AI judge ratings 2, 2, 2, mean 2.00, rounded to 2.
The one simplifying ideaA single fail-closed Zürich guidance tool that picks the right procedure itself, asks only for the missing field (origin, address) through a context token, and turns an official address into the nearest recycling stations using bundled city geodata.

Not ranked

sergi-villagrasa

Not enough evidence

Answers on the main test setup, quality +0.43 over 7 questions

5 appropriate abstention2 wrong

Score by part

Declared topics1 question-1.00Declared topics, hard1 question-1.00Honesty, outside the scope6 questions+0.67Breadth questions13 questions-0.85Second test setup7 questions+1.00

Breadth, 0 cells

No breadth cells.

Run

Completed: 20 of 20 questions
p95 tool call: 507 ms (few calls)
End-to-end p50: 7.3 s
Calls per answer: 0.33
Setup fixes: 10
Start-up test: passed
Commit evaluated: a9aafe74b24c

Official scores, event score 2.40: how each was reached

Technical Functionality, 5All 5 checks passed.
User Experience, 1Base 1 from the average on the declared topics -1.00 over 1 question. No deduction: honesty, citations and p95 tool call time within the limits.
Agentic Depth, 2Tools were called, but no answer on the main test setup is correct and backed by a tool result.
Creativity, 2AI judge ratings 2, 2, 2, mean 2.00, rounded to 2.
Market Impact, 2AI judge ratings 2, 2, 2, mean 2.00, rounded to 2.
The one simplifying ideaChain a Zurich Airport flight arrival, plus a transfer buffer, straight into an onward Swiss train search (connect_flight_to_train), wrapped around live OJP 2.0 connection, board and fare lookups.