Datasets:
license: other
license_name: fedlex-open-data
license_link: https://www.fedlex.admin.ch/
language:
- de
- fr
- it
task_categories:
- question-answering
pretty_name: Swiss Point-in-Time Law (CH-PiT)
size_categories:
- 10K<n<100K
tags:
- legal
- swiss-law
- point-in-time
- benchmark
- fedlex
configs:
- config_name: de
default: true
data_files:
- split: core
path: data/de/core-*.parquet
- split: full
path: data/de/full-*.parquet
- config_name: fr
data_files:
- split: core
path: data/fr/core-*.parquet
- split: full
path: data/fr/full-*.parquet
- config_name: it
data_files:
- split: core
path: data/it/core-*.parquet
- split: full
path: data/it/full-*.parquet
Swiss Point-in-Time Law (CH-PiT)
Version v2026.09. Items per language: de: 5000 (core 500), fr: 5000 (core 500), it: 5000 (core 500). Oracle: 1.000 on every language. Load with
load_dataset("overthelex/ch-pit", "de", split="core"); scorer and runners: https://github.com/overthelex/ch-pit.
Results on core
| system | n | correct % | wrong version % | ungrounded % | date-sensitive: correct % | de / fr / it | recite-easy: correct % | errors |
|---|---|---|---|---|---|---|---|---|
| recite/recite | 1500 | 48.7 | 41.1 | 10.1 | 0.0 | 0.0 / 0.0 / 0.0 | 100.0 | 0 |
| claude-haiku-4.5/closed | 1500 | 0.0 | 0.0 | 100.0 | 0.0 | 0.0 / 0.0 / 0.0 | 0.0 | 0 |
| claude-sonnet-5/closed | 1500 | 0.3 | 0.4 | 99.3 | 0.1 | 0.4 / 0.0 / 0.0 | 0.5 | 0 |
| gpt-5.6-terra/closed | 1500 | 0.1 | 0.1 | 99.9 | 0.0 | 0.0 / 0.0 / 0.0 | 0.1 | 0 |
| deepseek-v4-pro/closed | 1500 | 0.1 | 0.0 | 99.9 | 0.0 | 0.0 / 0.0 / 0.0 | 0.1 | 0 |
| claude-haiku-4.5/current | 1500 | 38.7 | 33.1 | 28.2 | 0.1 | 0.0 / 0.4 / 0.0 | 79.2 | 0 |
| claude-sonnet-5/current | 1500 | 46.3 | 39.1 | 14.6 | 0.7 | 0.0 / 0.8 / 1.1 | 94.3 | 0 |
| gpt-5.6-terra/current | 1500 | 47.5 | 40.6 | 11.9 | 0.1 | 0.0 / 0.0 / 0.4 | 97.4 | 0 |
| deepseek-v4-pro/current | 1500 | 47.7 | 40.0 | 12.3 | 0.1 | 0.0 / 0.0 / 0.4 | 97.8 | 0 |
| claude-haiku-4.5/pit | 1500 | 77.4 | 0.6 | 22.0 | 76.2 | 75.4 / 76.5 / 76.7 | 78.7 | 0 |
| claude-sonnet-5/pit | 1500 | 92.3 | 0.8 | 6.9 | 90.1 | 84.9 / 92.8 / 92.5 | 94.7 | 0 |
| gpt-5.6-terra/pit | 1500 | 95.0 | 0.6 | 4.4 | 92.7 | 90.5 / 93.6 / 94.0 | 97.4 | 0 |
| deepseek-v4-pro/pit | 1500 | 95.3 | 0.4 | 4.3 | 92.8 | 90.1 / 94.0 / 94.4 | 97.8 | 0 |
| claude-haiku-4.5/agentic | 300 | 93.0 | 1.3 | 5.7 | 90.2 | 90.9 / 88.9 / 90.7 | 96.4 | 0 |
| claude-sonnet-5/agentic | 300 | 94.7 | 1.7 | 3.7 | 91.4 | 92.7 / 90.7 / 90.7 | 98.5 | 0 |
| gpt-5.6-terra/agentic | 300 | 95.7 | 1.7 | 2.7 | 94.5 | 92.7 / 94.4 / 96.3 | 97.1 | 0 |
| deepseek-v4-pro/agentic | 300 | 92.3 | 1.7 | 6.0 | 89.6 | 89.1 / 88.9 / 90.7 | 95.6 | 0 |
Summary
CH-PiT asks one question about a system's answer: was it grounded in the version of the Swiss federal article that was actually in force on the date asked?
Swiss federal law is amended constantly, and Fedlex keeps every consolidated edition of every act. A system with no notion of time will happily quote the current wording of an article even when asked about a date years before that wording existed, or years after it was replaced. CH-PiT turns real amendment events into paired questions — "what did Art. X of act Y say as of date D?" — where the correct answer is one specific edition's text and the wrong answer is the adjacent edition's text, and it ships a deterministic scorer that tells the two apart from a free-text answer, not from a multiple-choice label.
Built by Lawrider, Zurich. First build: 2026-08-25 (v2). Current build:
v2026.09 (2026-09-04, v3): federal acts only, Fedlex XML editions
only, plus the fixed core subset every published baseline runs on.
Source and licence
Every item is derived from Fedlex (the Swiss federal law portal,
fedlex.admin.ch), specifically from consolidated Akoma Ntoso XML editions of
federal acts (ch_act_version with stage = 'parsed' in this pipeline's own
database — see services/ch-pipeline/README.md, "What this corpus does not
contain," item 1). Fedlex publishes machine-readable XML consolidations only
for a subset of acts, and only from the point each act's editor started
producing XML at all — that start date differs per act (for SR 220, the
Code of Obligations, the earliest German XML edition is dated 2021-01-01;
other acts differ). CH-PiT inherits that boundary: it can only ask about a
change if both the edition before and the edition after it exist as parsed
XML in this pipeline's database.
Fedlex data may be reused free of charge, provided the source is
acknowledged ("Fedlex data may be reused free of charge with source
attribution" — the exact licence note stamped on every item's licence
field). This card and every item's source/licence fields carry that
attribution forward. See https://www.fedlex.admin.ch/ for Fedlex's own terms.
Construction
Source table: ch_act_change, one row per (act, article, language) pair of
consecutive parsed editions that differ. A change becomes a benchmark item
pair only if all of the following hold:
A federal act.
ch_act.jurisdiction = 'CH'. The same tables now also hold the 26 cantons' legislation; CH-PiT asks about the federal SR collection only.Both editions are Fedlex XML.
ch_act_version.source = 'fedlex'on both sides of the change (the--sourcesflag;build-report.jsonrecords what was used). The pipeline also holds Fedlex's pdf-a consolidations for roughly 1995-2020 (source = 'fedlex_pdf'), whose article text still carries footnote apparatus -- see Known limits. Every item records the source of each edition ingold.source/distractor.source.Modified, not added or removed. Only
change_type = 'modified'rows are used — an article that appears for the first time or disappears has no "before" or "after" counterpart to ask about.Both texts are substantial. After normalisation (see Scorer, below), each edition's article text must be at least 200 characters. This rules out stub or largely-repealed articles with near-empty bodies.
The texts actually differ. The two normalised texts must not be the same string. This rules out a "change" that is only a re-typesetting of whitespace, soft hyphens or quote/dash characters, and nothing else. Deliberately not a similarity threshold: an earlier version of this rule required a
difflib.SequenceMatcherratio below 0.9, which threw away precisely the amendment CH-PiT exists to ask about — a change that swaps one figure leaves a multi-paragraph article ~0.98 similar to its predecessor, so the gate kept only wholesale rewrites and silently dropped every one-token amendment. Whether the pair can be scored apart is a separate question, answered separately and at the unit level by the discriminating-unit rule below, where a one-digit difference is visible.The act is in force.
ch_act.enforcement_status = 0. An act that has been repealed entirely is excluded — the benchmark asks about the live body of law, not archived history.An abbreviation exists for the language. German reads
ch_act.abbreviationdirectly. French and Italian look upch_act_alias(lang, sr_number, abbr, source), preferring acuratedrow over atitle_paren-sourced one when both exist. A change whose act has no abbreviation for that language is skipped (no_abbreviation) — the question templates always name the act by its abbreviation and SR number, never by its full title.The article number is unambiguous. The question can only name an article by its number ("Wie lautet Art. 7 OR …?"), and a system resolving that number back to a text has to pick one row. Swiss acts routinely carry the same article number twice inside one edition — a top-level
art_7and anart_7nested in a transitional-provisions block (disp_u17/art_7). For such a number the question is genuinely ambiguous: two different texts answer it. Changes on a nestede_id(one containing/), and changes on any number that some othere_idin either edition also carries, are excluded and counted asambiguous_article.The article number parsed at all. A change whose
article_numberis NULL cannot be named by any question template, and the ambiguity test above cannot see it either (SQL NULL compares equal to nothing, so both duplicate-number subqueries come up empty and the change looks unambiguous). Excluded first, counted asno_article_number.Both editions are
parsed. The resolver a system is measured against only ever returns aparsededition, so a change built on an edition that never reached that stage would be scored as a system failure on an item the corpus cannot express. Both version joins require it, so builder and oracle share one edition population.The two editions do not overlap. The old edition's
date_end_applicabilitymust be NULL, or strictly earlier than the new edition'sdate_applicability. Fedlex sometimes re-issues a consolidation without retracting the previous edition's end date, leaving two editions that both claim the same day (or days) as in force. Sincedate_end_applicabilityis inclusive, an end date equal to the new edition's start date is already an overlap. For such a change there is no date thebeforequestion can ask about: a covering lookup on the old edition's own last day returns the NEWER edition, and the item's gold answer scoresgrounded_wrong_versionagainst its own gold (13 items on the prod build before this rule). Both halves are dropped, not justbefore— the same overlap makes the change date itself ambiguous, with only the resolver's ordering deciding which edition wins. Counted asoverlapping_editions.The edition pair has at least one discriminating unit. Splitting both texts into paragraph- or sentence-level units (see Scorer), the edition valid on the query date must contain at least one unit that is not contained, word-for-word, anywhere in the other edition's text. If it doesn't — the whole visible change is something this scorer's unit-level matching cannot represent as a difference — the item is dropped (
no_discriminating_unit) rather than shipped unscoreable.Pure deletions are not benchmarkable this way. When an amendment only removes wording — a sentence struck from a paragraph, with nothing added — the shorter edition's text is entirely contained in the longer one's, so the half whose gold is the shorter edition has no discriminating unit and is dropped. This is not a scorer limitation to be tuned away: a correct answer is textually a fragment of the wrong answer, so nothing an answer contains can prove which of the two it meant. (The half whose gold is the longer edition is unaffected and is kept, so a deletion still contributes one item, not two.)
A change that survives all of the above produces two items, before and
after:
before:as_of= the gold edition's last day in force — itsdate_end_applicability, which is inclusive. This is usually the change date minus one day, but not always: consecutive parsed editions can leave a gap, since Fedlex did not publish XML for every consolidation, and the day before the change can fall in that gap where no edition exists to answer the question at all. Only when the old edition carries nodate_end_applicabilitydoes the builder fall back to the change date minus one day. Gold = the edition valid before the change; distractor = the edition valid after it.after:as_of= the change'sdate_applicabilityitself; gold = the edition valid after the change; distractor = the edition valid before it.
Either half can still be dropped independently by the discriminating-unit rule above, so a change can contribute 0, 1 or 2 items.
Caps and sampling. At most 50 changes are kept per act (so one heavily
amended act like the Code of Obligations cannot dominate the sample), and at
most 5,000 items per language in total. The language cap is on ITEMS, and
the builder consumes eligible changes until it is reached — a change yields
one item or two, never a fixed number, so budgeting changes instead would
leave the cap unfilled. Selection above the caps is random,
seeded per language as random.Random(f"{seed}:{lang}") with
seed = 20260825 — each language's sample depends only on its own seed and
its own eligible-change set, never on which other languages were built in
the same run or in what order. This makes bench-fr.jsonl byte-identical
regardless of whether French was built alongside German and Italian or on
its own.
The core subset. The full build is 5,000 items per language; the
scorer is free, so the only cost a baseline has is its model calls, and
without one fixed sample every reproduction picks its own subset and the
numbers stop being comparable. core is the sample every published
baseline runs on: 500 items per language, chosen by
chpipe/bench/core_split.py with random.Random(f"{seed}:{lang}:core").
It takes an equal share of each as_of year present in that language's
items (remainder to the earliest years), and within a year fills the four
(kind, gold_is_current) cells round-robin, so before/after and
current/superseded gold are each as even as the year's pool allows. A year
too thin for its share (2011 has a few dozen items, only in French and
Italian) keeps what it has and the shortfall is filled from the other
years, again round-robin. Note that a before item's gold is the edition
the change replaced, which is current only in the rare case where the old
edition never received an end date, so the current/superseded balance is
really an after property. core-{lang}.jsonl holds the subset and every
item in bench-{lang}.jsonl carries core: true/false. Per-year and
per-cell counts are in build-report.json under {lang}.core.
Every skip reason above (no_abbreviation, identical_or_short,
ambiguous_article, overlapping_editions, no_discriminating_unit, plus
capped for anything left unused by the two caps) is counted in build-report.json, per
language, alongside changes_considered, selected and items.
changes_considered is every modified change on an in-force act in that
language, including the ones excluded in SQL, so the skip counts account
for the whole difference rather than the exclusions shrinking the total.
Fields
Each line of bench-{lang}.jsonl is one JSON object:
| field | type | meaning |
|---|---|---|
build |
string | the build this item was first published in, e.g. "v2026.09" -- items are frozen at publication and later builds only add |
core |
boolean | true for the 500-per-language core subset every published baseline runs on (see Construction) |
id |
string | stable id, first 16 hex chars of sha1(f"{lang}|{act_id}|{sr_number}|{e_id}|{as_of}") — act_id is in the payload because more than one act can share an SR number, and two such acts amended in the same article on the same date would otherwise collide on one id |
lang |
string | de, fr, or it |
act_id |
integer | the exact ch_act row this item's editions come from — resolve editions by this, not by sr_number alone, since more than one act can share a SR number (a predecessor act and its successor filed under the same number) |
sr_number |
string | the act's Systematische Rechtssammlung number (e.g. "220") |
abbreviation |
string | the act's abbreviation in this language (e.g. "OR", "CO") |
article_number |
string | the article number asked about (e.g. "336") |
e_id |
string | the article's element id in the Akoma Ntoso structure, used to join editions |
as_of |
string (ISO date) | the date the question asks about |
kind |
string | before or after — which side of the change as_of falls on |
change_date |
string (ISO date) | the ch_act_change row's date_applicability this item pair is derived from |
question |
string | the rendered natural-language question, in lang |
gold_is_current |
boolean | true when gold.date_end_applicability is null, i.e. the gold edition is still the wording in force today. An item where this is true can be answered correctly by a system that simply recites the current text and ignores the date; only the false items measure point-in-time grounding. The report splits the correct-answer share on this flag |
gold |
object | the edition valid on as_of — see below |
distractor |
object | the adjacent edition (the other side of the same change) — see below |
source |
string | always "Fedlex (fedlex.admin.ch)" |
recite_label |
string or null | how the scorer labels today's edition of the article (fetched from the point-in-time tool without a date, on recite_as_of) against this item: grounded_correct means the date does not matter for this item -- reciting current law answers it; grounded_wrong_version / ungrounded mark the date-sensitive items. Null where the tool could not resolve the article. Stamped at publication from the recite baseline (see Baselines) |
recite_as_of |
string (ISO date) or null | the day recite_label was measured; Fedlex keeps publishing, so the label is a dated fact |
licence |
string | always "Fedlex data may be reused free of charge with source attribution" |
gold and distractor share the same shape:
| field | type | meaning |
|---|---|---|
version_id |
integer | the ch_act_version row id |
date_applicability |
string (ISO date) | the date this edition took effect |
date_end_applicability |
string (ISO date) or null | the LAST DAY this edition was in force (inclusive), or null if still current -- the next edition's date_applicability is this date + 1 day |
eli |
string | the edition's eli_consolidation_uri, Fedlex's own permanent identifier for it |
source |
string | where the edition's text came from: fedlex (Akoma Ntoso XML) -- the only value in this build; fedlex_pdf (pdf-a consolidation) is reserved for a later build |
text |
string | the article's text in this edition, verbatim from ch_act_article.text — no normalisation applied at build time |
Scorer
chpipe/bench/score.py decides, given a free-text answer and the item's
gold.text / distractor.text, whether the answer is grounded in the
correct edition, the wrong one, or neither. It is pure and deterministic:
no database, no network, same three strings in, same verdict out.
Normalisation (normalise()): NFKC fold, lower-case, unify quote
characters (guillemets, curly quotes) to "/' and dash characters (en
dash, em dash, minus sign, non-breaking hyphen) to -, drop soft hyphens
(a Fedlex line-break artefact), collapse whitespace runs to a single space.
This folds cosmetic re-typesetting differences out without touching actual
wording.
Units (units()): each candidate text is split into paragraph-level
units on Fedlex's numbered-paragraph markers (1 , 1bis , 2 , …), or,
if the text has no such markers, into sentence-level units on ./;/:.
A unit shorter than 25 normalised characters is discarded — too short to
tell one edition from another, and short strings inflate similarity scores
by chance. This 25-char floor is the default used everywhere in the scorer
except one case: distractor_all_coverage (see below) is computed with an
8-char floor instead, so a short but real distractor paragraph is not
dropped out of that one computation.
Partition. The gold and distractor texts' units are compared against each other and split three ways: gold-only (a unit of gold whose normalised form does not occur anywhere in the distractor's normalised text — evidence the answer quotes gold), distractor-only (the mirror image), and shared (everything else: a unit contained in the other edition's text, so finding it is evidence of neither). Coverage is the fraction of each partition's units found in the answer; shared-coverage is reported for diagnostics only and never decides the label.
The test is containment, not equality, and that matters for the
commonest amendment shape there is. When an amendment adds words to a
paragraph, the old paragraph is a substring of the new one: the two units
are not equal, but every answer quoting the new (correct) wording verbatim
necessarily contains the old wording too. Treating the old paragraph as
distractor-only therefore scored word-for-word correct answers as
ungrounded — 1,097 items on the first full build did exactly that.
Under containment the old paragraph is shared, and only the added wording
discriminates.
distractor_all_coverage. Containment has a consequence: for an
amendment that only adds wording, no unit is distractor-only, so
distractor_coverage is 0.0 by construction and an answer reciting the old
edition could never be caught as wrong-version. So the scorer also reports
distractor_all_coverage, the share of all the distractor's units —
distractor-only and shared alike — found in the answer, and uses it as the
wrong-version signal in exactly that case (see Labels). It is a diagnostic
everywhere else: on such an item a correct answer scores 1.0 on it too,
which is why it is only ever consulted after the gold-side test has already
failed.
This one computation uses a lower unit-length floor: 8 normalised
characters, not the usual 25. Reusing the normal 25-char floor here
reintroduced a version of the same bug this fallback exists to close: when
the amendment's old (distractor) paragraph is itself short — e.g. a
two-paragraph article where paragraph 2 changes from "Er ist zu
begründen." (20 chars) to "Er ist zu begründen und zu unterzeichnen." — the
25-char floor drops that paragraph out of the set entirely, leaving only
the unrelated, unchanged paragraph 1. An answer that quotes only that
unchanged paragraph then found the one unit that survived the filter,
scored distractor_all_coverage 1.0, and was labelled
grounded_wrong_version despite never touching the amendment. With the
8-char floor the short paragraph is back in the set, and that same answer
now scores 0.5 (one of two units found) — which is also why the fallback's
own threshold was raised to 0.8 (see Labels): on a two-unit set, 0.5 still
clears the normal 0.6 floor.
Discriminating pairs. A Fedlex amendment often changes exactly one number or short word and leaves the rest of the paragraph untouched (e.g. "180 days" becomes "30 days"). Two such near-identical paragraphs can score above 0.92 on plain string similarity even though they mean different things. The rule, in one sentence: if fuzzy matching would find this unit in the other edition, fuzzy matching is not allowed for it. Such a unit is discriminating: it may only be found by an exact substring match in the normalised answer, never by the fuzzy window match described next. This is what lets the scorer catch the one-number-changed case, which is the hard case the benchmark exists to test.
A unit is flagged discriminating when either test fires:
- Window test. The unit is run through the same fuzzy window match
(same window lengths, same 0.92 threshold) against the other edition's
whole normalised text. Asking the flag question with the matching
function is what makes the scorer self-consistent rather than merely
stricter: against an answer that quotes the other edition verbatim, the
flag test and the match test are the same computation on the same string,
so the unit is either flagged (exact-only, correctly not found) or
unflagged (and the window match correctly fails too). Either way the
cross-edition false positive cannot happen. Without this test, 222+ items
on the prod oracle run scored
ungroundedon a word-for-word correct answer — e.g. SR 142.203 Art. 3, where an amendment both rewords a phrase and appends a clause, leaving a distractor-only sentence that scores only 0.84 pairwise against the gold sentence but window-matches inside the gold text above 0.92. - Unit-pair test. The unit scores 0.92 or higher against any of the other edition's units, shared ones included. This is kept alongside the window test, not replaced by it: the window match only probes windows of roughly 0.8x, 1.0x and 1.2x the unit's length, so when the other edition's whole text is shorter than 0.8x the unit — a one-paragraph article with a clause appended — no window exists and the test cannot fire, on a pair that is 0.96 similar unit to unit. The comparison runs against all of the other edition's units, shared ones included, because containment files a unit as shared exactly when it sits inside the other edition's text, which is itself the near-duplicate relationship this guard exists to catch.
Fuzzy window match (0.92). For a unit that is not a discriminating
unit, if it does not occur verbatim in the normalised answer, the scorer
also checks whether some window of the answer roughly the unit's own
length matches it with SequenceMatcher.ratio() >= 0.92. This tolerates a
model re-typing a word slightly wrong, a stray OCR-style slip, or a typo,
while still rejecting a paragraph that is merely topically similar.
Cost of the exact-match rule. Switching fuzzy matching off for discriminating units is what makes the one-number case scoreable at all, and it is not free. Four consequences to read a CH-PiT number with:
- One character wrong in a discriminating unit means "not found." An answer that gets the amended figure exactly right but mistypes a word elsewhere in the same paragraph scores that unit as missed. There is no partial credit inside a discriminating unit.
- A long single-paragraph article is effectively a verbatim-only item.
Units are paragraphs; an article that is one long paragraph is one long
unit, and if that unit is discriminating, the whole item can only be
scored
grounded_correctby an answer that reproduces the entire paragraph character-for-character after normalisation. Such items measure verbatim recall as much as they measure date grounding. - Answers over 20,000 characters get substring matching only. Window
matching is skipped above that length (a bounded-calculation trade — see
the performance cap in
score.py), so a runaway answer loses near-verbatim credit even on non-discriminating units. - A correct answer in the wrong language scores
ungrounded. Units are compared against the item's owngold.text, which is in the item'slang. A model that answers the German question with a correct French quotation of the same article shares no units with either candidate and is scored as grounding in neither — not as correct.
Labels. grounded_correct requires gold_coverage >= 0.6 and
distractor_coverage <= 0.2, and is tested first. grounded_wrong_version
requires gold_coverage <= 0.2 and either distractor_coverage >= 0.6 or —
when there are no distractor-only units at all, the pure-addition case above
— distractor_all_coverage >= 0.8 and the answer containing at least
one unit the distractor edition has and the gold text does not (the
amendment's old wording, at the same 8-char floor). That fallback threshold
is deliberately higher than the 0.6 used everywhere else, in exchange for
the lower 8-char unit floor described above; the extra unit condition is
what stops a long article's shared paragraphs from carrying the 0.8 share
on their own (five shared paragraphs out of six units is 0.833). "Has and
gold does not" is unit-set membership, not substring containment: a
distractor paragraph nested inside a gold paragraph ("Er ist zu
begründen." inside "Er ist zu begründen und zu unterzeichnen.") is exactly
the wrong-version answer this fallback exists to catch. When no such unit
exists at all — every distractor paragraph is also a gold paragraph, the
amendment only appended text — the fallback cannot fire. Everything else —
including an answer that clears neither floor, or one that clears 0.6 on
one side while also leaking past 0.2 on the other — is ungrounded.
Why 0.6 / 0.2, not 0.5 / 0.5. Many changes have as few as one gold-only
and one distractor-only unit, so coverage on that partition can only be 0.0
or 1.0 — 0.6/0.2 is simply "found it" versus "did not find it" in that
common case. The gap between 0.2 and 0.6 is reserved for partitions with
several units, where an answer that got some but not all of the
discriminating wording right is scored ungrounded rather than guessed
at: a partial match to one edition's distinguishing wording is not
evidence the model resolved the date correctly, and treating it as such
would let a benchmark that exists to measure date-grounding reward
recitation-by-memory instead. The split is also deliberately asymmetric in
effect: it biases toward ungrounded over a confident label whenever
coverage is ambiguous, because a false grounded_correct or
grounded_wrong_version corrupts an accuracy number a paper would cite,
while an ungrounded false negative only costs one data point.
Baselines
Every published baseline runs on the core split (500 items per
language) through the chpit runners
(https://github.com/overthelex/ch-pit) with models served by OpenRouter
(pinned slugs; every result line records the model that actually served
the answer, the provider, tokens and cost). Settings: temperature 0,
reasoning.effort = minimal, max_tokens 8192 (16384 on a truncated
reply), workers in parallel, resumable runs; each mode's
run-report-{mode}.json in results/{version}/ carries them. Six
systems:
- oracle -- the database lookup by
act_idand date; must be 1.000. - recite -- no model: the article as
ch_get_act_articlereturns it WITHOUT a date, i.e. today's edition. Correct exactly on the items where the date does not matter (recite_label, above); wrong-version or ungrounded on the date-sensitive ones. This is the floor every model with retrieval has to beat, and the split every table reports. - closed -- the model sees only the
question(act, article, date), no retrieval. System prompt, verbatim: "You are a Swiss legal database. Answer with the verbatim text of the requested article as in force on the given date, in the language of the question, nothing else." - current -- today's edition handed to the model as context; the model is asked to reproduce it. Measures copy fidelity of a wrong-edition context.
- pit -- the edition valid on
as_ofhanded to the model as context. Its distance to the oracle's 1.000 is pure transcription loss. - agentic -- the model gets the three point-in-time tools
(
ch_get_act_article,ch_get_act_history,ch_get_act_text, English descriptions vendored in the repo and checked against the live server for schema drift) and has to decide itself whether to pass the date. The system prompt says the tools default to today's edition and the question names a date; it does not say to passas_of. Every tool call's arguments are logged, andreport --toolsgives the share of items where the model asked for the right date, the right act and article, and how often it answered without calling a tool at all. Up to four tool rounds, then a forced answer.
Read the date-sensitive column, not the headline. About half of
core is answerable by reciting current law (recite is correct on it).
report --hard-from results-recite-recite.jsonl adds, for every system,
the correct share on the items where recite is wrong; that column is the
point-in-time number. gold_is_current is kept for continuity but is an
edition-level flag: a later edition usually changes other articles and
leaves this one's text equal to gold, and an edition published with a
future in-force date is "current" in the database while not being the
text in force today, so it splits items worse than recite_label does.
Caveats every reader should carry: temperature 0 is not determinism for
hosted models; OpenRouter may route a slug to different backends
(recorded per line); the retrieval and agentic baselines call Lawrider's
own MCP server, so they measure the model plus that tool, disclosed as
such; recite, current and everything that fetches without a date are
dated measurements; the tool resolves the act by SR number with an
in-force tiebreak, so a few items whose SR number is shared with a
predecessor act resolve differently than the oracle's act_id (reported
as errors). Baseline outputs, the MCP cache (the exact context every
model saw) and run reports are published under results/{version}/.
Oracle (run_oracle.py): answers every item straight from the
database, resolving act → edition → article the same way the product tool
(ch_get_act_article) does, with no LLM involved. Its only purpose is to
prove the builder and the scorer agree with each other — the oracle must
score 100% grounded_correct. Anything less means a bug in build.py or
score.py, not that "the database got the date wrong."
Bedrock models (v2 history) (run_llm.py, since replaced by the chpit runners): asked a model the exact question field
from the item, with no retrieval — the model has no access to the
gold or distractor text, only the act, article number and date, exactly as
a chat user would type it. System prompt, verbatim:
You are a Swiss legal database. Answer with the verbatim text of the requested article as in force on the given date, in the language of the question, nothing else.
Sampling is 300 items per language, stratified by kind (before/after)
so a systematic bias toward one side of a cutoff date cannot hide inside an
unbalanced sample, seeded the same way item selection is
(random.Random(f"{seed}:{lang}:llm")). Temperature 0, max 2048 output
tokens. Every Bedrock call costs money; run_llm.py refuses to run without
CHPIPE_BENCH_CONFIRM=1, printing a cost estimate first — see the README's
bench section for the exact commands.
Reading the report. report.py reduces a run to one row per (language,
system, kind) plus an all row per (language, system), and reports on
each row the label shares, the mean coverages, an errors count (result
lines the system failed on — a Bedrock exception, or an oracle resolution
step that came up empty), and the correct-answer share split on
gold_is_current. Read the gold_is_current = false share, not the
headline score, as the point-in-time grounding number.
Results
Current build: v2026.09, 2026-09-04 10:22 UTC -- federal acts, Fedlex
XML editions, 5,000 items per language plus the 500-per-language core
subset. Oracle on all 15,000 items: 1.000 in German, French and Italian,
0 errors. Baselines on core (run 2026-09-04, see RESULTS.md):
closed-book 0.0–0.3% for Claude Haiku 4.5, Claude Sonnet 5, GPT-5.6-terra
and DeepSeek V4 Pro; with the edition valid on the date as context
(pit) 76–93% on the date-sensitive items; with the tools and the model
deciding (agentic) 90–95%, every model passing the date on every item. The paragraph below
describes the v2 run of 2026-08-25, kept for history.
Previous build and run: 2026-08-25 23:24 UTC, commit 28618f7d (v2 --
item ids include the source act's act_id, parsed editions only, adds the
no_article_number skip). A first run on the previous build (v1, item ids
without act_id, 2026-08-25 19:56 UTC) gave identical headline numbers.
Full numbers, per-language and per-kind tables (now including mean gold
and mean distractor coverage), the year distribution of as_of dates, the
runner settings, actual spend for both runs, and a qualitative read of the
model answers are all in RESULTS.md. Headline numbers:
- Oracle: 1.000 (
grounded_correct) in German, French and Italian, the builder and the scorer agree with each other. - Haiku 4.5 (
eu.anthropic.claude-haiku-4-5-20251001-v1:0, no retrieval): 0.000 in all three languages. - Sonnet 4.6 (
eu.anthropic.claude-sonnet-4-6, no retrieval): 0.003 in German (0.007 onafteritems, 0.000 onbefore), 0.000 in French and Italian.
Without point-in-time retrieval, a general-purpose model's verbatim recall of the specific dated edition is effectively zero. CH-PiT is measuring whether a system has a retrieval layer that finds the right edition, not how good the underlying language model is.
Dates
The as_of year distribution in the current build includes 191 items in
2027, 13 in 2028 and 11 in 2029, all after the 2026-08-25 23:24 UTC build
date. These are not placeholder or malformed dates: they are editions
Fedlex has already published with a future in-force date (an amendment
enacted now that takes effect on a later date), and CH-PiT asks about the
edition valid on the date it takes effect, which can be after the date the
benchmark was built. See RESULTS.md for the full year-by-year table.
Known limits
- Fedlex XML editions only, mostly from 2021 onward. CH-PiT asks
only about changes where both adjacent editions are parsed Akoma Ntoso
XML (
source = 'fedlex'). Fedlex's XML consolidations do not go back further than roughly 2020 for most acts (SR 220's earliest German XML edition is dated 2021-01-01). The pipeline does hold Fedlex's pdf-a consolidations for roughly 1995-2020, split into articles, andch_act_changehas ~166K federal German "modified" changes for 2000-2019 built from them -- but on a 100-pair hand-read sample of those changes that pass this builder's filters, about 85% were footnote apparatus surviving in the article text (three-digit footnote references glued to words, footnote bodies where no running header marked the page), not amendments. Until the PDF splitter strips that apparatus (tracked as LEXAI-2046), the pre-2021 history stays out; the--sourcesflag and the per-editionsourcefield are how it comes back in a later build without changing the item shape. - Three languages, no Romansh. German, French and Italian only — Fedlex does not publish Romansh consolidations for these acts.
- In-force acts only. An act with
enforcement_status != 0(repealed entirely) contributes no items, even if its amendment history is otherwise well covered. - The hard case is a single number or short word changing. Most real Fedlex amendments change one figure or phrase in an otherwise unchanged paragraph (see Scorer, "Discriminating pairs"). This is the case the benchmark is designed to catch and the case a system with no real point-in-time grounding is most likely to get wrong by reciting the current wording from memory.
- The
2021-01-01"placeholder date" used elsewhere in this repo does not apply here.services/ch-pipeline's decisions pipeline treatsdecision_date = '2021-01-01'as a source placeholder meaning "no date known" for a subset of court decisions (see the README's "The2021-01-01placeholder" section). CH-PiT'sas_ofandchange_datefields are real dates read fromch_act_versionandch_act_change— 2021-01-01 appearing in an item is a real change date, never a stand-in for "unknown." - Long unsegmented articles are near-unscoreable in practice. The
scorer works on paragraph units. An article Fedlex publishes as a single
unnumbered block becomes one unit, so
grounded_correctdemands a character-exact reproduction of the whole thing (see Scorer, "Cost of the exact-match rule"). Such items are kept — they are real questions — but a system's score on them reflects verbatim recall more than date resolution, and a per-item length breakdown is worth looking at before reading a headline number as a grounding measurement. - Some items have a still-current gold edition (11% of v2026.09:
1,657 of 15,000; a third of
core'safteritems by construction). Theafterhalf of a pair whose change is the most recent one asks for wording that is also today's wording, and reciting the current text answers it correctly with no date reasoning at all. This is why every item carriesgold_is_currentand the report splits the correct-answer share on it; thegold_is_current = falseshare is the one that measures what the benchmark is named after. - Amendments that only delete wording contribute one item, not two. See Construction: the half whose gold is the shorter edition has no wording of its own to find and is dropped. A benchmark built this way therefore under-represents repeals relative to their share of real amendment traffic.
- The
distractor_all_coveragefallback is a coarser signal thandistractor_coverage. Its share is measured over all of distractor's units (8-char floor, 0.8 threshold — see Scorer), not just the unit the amendment actually touched, so on a long article the shared paragraphs dominate the arithmetic. The label no longer rests on that share alone — the fallback additionally requires the answer to reproduce a unit gold does not have, which is what makes it a statement about the amendment rather than about the article — but the share itself is still a blunt instrument next todistractor_coverage's proper distractor-only partition, which is scoped to exactly the discriminating wording. The price of the extra condition is the pure-addition-of-whole-paragraphs shape: when every distractor paragraph is also a gold paragraph, no answer can be labelledgrounded_wrong_versionat all, and an answer reciting the old edition readsungrounded. beforeitems are dated by the edition, not by the change. An item'sas_ofis the gold edition's last day in force, which is the change date minus one day only when the two editions are contiguous. Where Fedlex's XML consolidations skip an edition,as_ofsits further back — still a date on which the gold text was genuinely in force, but not always adjacent tochange_date.- No cantonal law. CH-PiT covers only federal acts (the corpus this pipeline builds from Fedlex's federal SR collection); cantonal statutes are out of scope entirely.
How to reproduce
From services/ch-pipeline:
# 1. Build the items (v2026.09: federal acts, XML editions, 500-per-language core).
python -m chpipe.bench.build --langs de,fr,it --out /data/ch-corpus/bench-v3 \
--build v2026.09 --core-per-lang 500 # --sources fedlex is the default
# 2. Oracle. Must come back 100% grounded_correct before step 3 spends anything.
python -m chpipe.bench.run_oracle --items /data/ch-corpus/bench --out /data/ch-corpus/bench
# 3a. Bedrock baselines, DRY RUN: prints a JSON cost estimate, calls nothing, exits 2.
python -m chpipe.bench.run_llm --items /data/ch-corpus/bench --out /data/ch-corpus/bench --sample-per-lang 300
# 3b. Only once that estimate looks reasonable, run for real.
CHPIPE_BENCH_CONFIRM=1 python -m chpipe.bench.run_llm --items /data/ch-corpus/bench --out /data/ch-corpus/bench --sample-per-lang 300
# 4. Report.
python -m chpipe.bench.report --results /data/ch-corpus/bench/results-oracle.jsonl /data/ch-corpus/bench/results-llm-haiku-4-5.jsonl /data/ch-corpus/bench/results-llm-sonnet-4-6.jsonl --items /data/ch-corpus/bench --out /data/ch-corpus/bench/report.json
See services/ch-pipeline/README.md, "Point-in-time benchmark
(chpipe.bench)," for the full command reference, the cost gate, and what
each step writes.
Citation
@misc{chpit2026,
title = {Swiss Point-in-Time Law (CH-PiT): A Benchmark for Date-Grounded Legal Question Answering},
author = {Lawrider},
year = {2026},
note = {Built from Fedlex (fedlex.admin.ch) consolidated legislation},
url = {https://www.fedlex.admin.ch/}
}