Hack Apertus 2026 · Track 2A · Challenge by OST

GeaNaVid

Every claim checked against the official voting booklet. Each Yes or No comes with the booklet line behind it.

Gea · Na · Vid — Romansh for Yes · No · Blank

▶ Demo video Built on Apertus v1.5 Team Apertool · solo builder

The problem

Voters hear the claim.
The answer is on page 4 of 40.

The claim

„… die Bevölkerung seit Einführung der Personenfreizügigkeit im Jahr 2002 um rund 500'000 Personen gewachsen ist.“

“… the population has grown by about 500,000 since free movement began in 2002.”

Official booklet · page 4

„Seit Einführung der Personenfreizügigkeit im Jahr 2002 ist die Bevölkerung um rund 1,7 Millionen Personen gewachsen.“

Na contradiction

OST demo set, label 2; Federal Council explanations for 14 June 2026, German edition, PDF page 4.

The task

Booklet + claim in, a ballot mark out.

Voting booklet
PDF · German, French or Italian
Claim
one sentence · DE, FR or IT
→
GeaNaVid
CLI · Apertus v1.5
→
Gea entailment 0
Na contradiction 2
Vid neutral 1
+ the passages it rests on · tokens used · seconds taken
Judged on Macro-F1 on a hidden benchmark evidence quality input tokens inference time

All 9 claim/booklet language combinations. Labels 0/1/2 and judging criteria as set by OST in the Hack Apertus challenge brief.

The name

Three answers, in Romansh.

Gea Yes entailment
Na No contradiction
Vid Blank neutral

Every Gea and every Na comes with the booklet line that decides it. A Vid quotes nothing.

Gea, Na and vid(as) as printed in Romansh voting booklets of 1994 and 1995 (VotingBooklets-v1, Stüssi & Vamvas, SwissText 2026, CC BY 4.0).

The context

Narrow first, then let Apertus read.

1
Top-5 passages of the right proposal, with page neighbours the vote title narrows the booklet, BM25 picks the passages, the section the claim names first
≈ 1,100 tokens of passages
Claim in another language? Apertus translates it into the booklet language first
2
Translation one short call, then the same search
+ ≈ 300 tokens
full
The whole booklet the baseline OST recommends
20,242 tokens

Per claim: 1,461 input tokens with 70B, 2,159 with 8B, instead of 20,242.

Mean input tokens per claim as reported by the API, OST v1.1 test split (493 pairs), Apertus v1.5, 14 October 2026.

The trade-off

Under a ninth of the tokens. Better verdicts.

1.00.750.50.250 Macro-F1 always one label: 0.17 70B 0.9658B 0.959without translation 0.82 / 0.7870B 0.889whole booklet8B 0.720whole booklet 5001k2k5k10k20k50k100k Input tokens per claim (log scale)
Apertus 70B submitted 0.965 · 1,461 tokens 1.3 s mean · 2.2 s p95
Apertus 8B 0.959 · 2,159 tokens 0.37 s mean · 0.55 s p95
Beyond the test split Held-out booklets 0.932 / 0.874. Older booklets 1998–2018: 0.890 / 0.836 (70B / 8B).

OST v1.1 test, 493 pairs; 95 % CI 0.944–0.983 / 0.933–0.980; time: cold, 50 cases; caveats: report §4.

Across languages

Claim in Italian, booklet in German.
The evidence stays German.

DE
FR
IT
DE
0.980.98
0.940.98
0.910.91
FR
0.960.98
1.000.98
0.860.91
IT
1.000.94
0.960.97
1.000.98

Macro-F1 per pair: top Apertus 8B, below 70B. Rows: claim language · columns: booklet language. Framed: same language.

0.70 → 0.94 8B
0.74 → 0.94 70B

macro-F1 on the 330 cross-lingual test claims, before and after Apertus translates the claim (ablation step; full defaults 0.943 / 0.958).

The search runs on the translation, the evidence is quoted from the booklet as it is. No language field is needed: the claim language is detected.

OST v1.1 test, 38–74 claims per pair (cells ± 0.05); same / cross language 0.993 / 0.943 (8B), 0.980 / 0.958 (70B).

Evidence

Every Yes or No points to a line in the booklet.

Na contradiction · label 2
Claim (Italian): the reduced VAT rate falls from 2.5 to 2.4 per cent
„Der dafür geltende reduzierte Mehrwertsteuersatz steigt von 2,5 auf 2,6 Prozent; …“
25.09.2022 · DE · page 28 · AHV 21: additional financing by VAT
input tokens 2,794 output tokens 52 Apertus 8B
Quotes, not paraphrase The evidence is booklet sentences, verbatim, each with its PDF page, from the section the claim names, chosen after the verdict without another model call.
Evidence first, label second Apertus names the passages before it gives the verdict, in strict JSON.
Not voting advice GeaNaVid says what the booklet says. It never says how to vote.

OST case v1-1484 (dev split). By OST's scoring rule a quote lies in the gold passage in 96.3 % (70B) / 96.9 % (8B) of 324 cases.

Our ask

Score it on the hidden booklets as submitted,
and read the lines it quotes.

Niklas Wichter Team Apertool · solo builder · one of four tools at apertool.ch

Scan: watch the 2-minute demo

geanavid.apertool.ch

Hack Apertus 2026, Track 2A, challenge by OST. GeaNaVid is an independent entry, not a product of OST or the Federal Chancellery.