Fifth Lexis 360 intake (31 briefings, source Stands 2025-12 to
2026-07), extending the corpus to 236 curated entries in 33 clusters.
This batch opens the third briefings chapter 'Arbeitszeit' (AZG
working-time law, the structural backbone for Work-Entry-based
payroll).
- 8 new clusters: arbeitszeitgrenzen (4), arbeitszeitmodelle (7,
incl. Kurzarbeit and Gleitzeit/Schicht), meldepflichten (4),
nachtschwerarbeit (3), ruhezeiten (4), schwangerschaft (1),
uberstunden (4), zeiterfassung (3); plus lb-jug-08 appended to
the jugendliche cluster
- one URL-encoded export filename (sharp-s) normalized; no
duplicates; no fossil ids this time - the specific-first keyword
ordering held
- cross-batch reconciliation: three procurement gaps closed and
back-referenced - 'Jugendliche - Arbeitszeit' (lb-jug-04 and
lb-jug-05 -> lb-jug-08) and 'Aufzeichnungspflicht, Formen der
Zeiterfassung' (lb-jug-05 -> lb-zer-01/02)
- RECHTSQUELLEN cross-checks: 68-EStG overtime free limits 2026
(15 UStd/170 EUR) confirmed via lb-zus-07; the Kurzarbeit/
Lehrlinge corpus conflict is now triangulated - lb-azm-05
(2026-07) sides with lb-leh-07 (apprentices excluded since
2023) against lb-leh-11's '100%'
- flagged batch-5 points: NSchG contribution 2026 (3.8%) cited
without BGBl evidence (lb-nsc-01), 'para 7 Abs 1' vs 'Abs 6 AZG'
citation divergence for the overtime refusal right
(lb-ues-01/04), MAV-START EU conflict (lb-rhz-03), KAB-Richtlinie
computing the 39.05% partial amount with outdated SV rates
(lb-azm-06), 12-Abs-2a-vs-12a designation conflict (lb-zer-01)
- batch-6 candidates updated in the RUNBOOK (Entgeltfortzahlung an
Feiertagen, Motivkuendigung, Gefahrenevaluierung u. a.);
README cluster table extended to batches 1-5 incl. the third
chapter
Validated: --registry 236 entries/33 clusters, --check 0 problems,
structural scan over all 236 curated docs, py_compile.
Move the licensed PDF exports and the Layer-1 full texts from the
project root into the dot directory .lexis360/, keeping the root
listing to versioned content plus the tool checkouts.
The location also encodes provenance, distinct from .firecrawl/:
.lexis360/ holds user-supplied licensed exports that cannot be
re-fetched by an agent - if missing, stop and ask the user
(documented in MEMORY), whereas .firecrawl/ remains the home of
reconstructible web fetches.
- plain move of the unversioned directory (batch2/ subfolder spotted
and left untouched for its own intake run)
- .gitignore, tool constants/messages, source.pdf/source.text
frontmatter paths of all 55 curated entries, README/RUNBOOK/MEMORY
mentions; kb.json/INDEX.md regenerated
- batch 2 (53 PDFs) recorded in MEMORY as staged in .lexis360/batch2/
Validated end to end: --extract 55/55 with stable ids, --registry
55 entries/6 clusters, --check 0 problems (every frontmatter path
resolves at the new location), no stale path references remain.
Curated knowledge base from the first batch of Lexis 360 PDF exports
(LexisNexis "Lexis Briefings Personalrecht", 55 documents, source
Stands 2024-03 to 2026-09). Purpose: development reference alongside
RECHTSQUELLEN-*.md and a retrieval-ready corpus for a future
question-answering copilot.
Licensing split (decision D1): the PDF exports and the extracted full
texts stay local and unversioned (Lexis360/ gitignored, like
.firecrawl/); only the curation in own words with short quotes and
source citations is versioned.
- wissensbasis/dokumente/: one curated entry per briefing (55) with
approved frontmatter (frozen lb-<cluster>-<nn> ids, English
structural keys, German values, stand YYYY-MM, ASCII topic/tags
slugs) and sections Zusammenfassung, Kernwerte & Fristen (Stand
marked), Rechtsgrundlagen (only citations actually in the source,
verified/plausible/open marks), Payroll-Relevanz (Odoo), Verweise
- clusters: altersteilzeit 15, lehrlinge 15, jugendliche 7,
ferialpraktikanten 4, beschaftigungsformen 8, teilzeit 6
- kb.json + INDEX.md: generated from the frontmatter by --registry
(single source of truth, schema-validated)
- tools/build_lexis_kb.py: stdlib-only pipeline - --extract
(pdftotext, breadcrumb/stand parsing incl. dossier-style exports
with wrapped breadcrumbs), --registry (frontmatter validation:
required keys, id pattern, id prefix vs topic, stand format,
dangling cross_refs), --check (layer 1 <-> layer 2 completeness);
ids are frozen across batches, numbers never reused
- RUNBOOK.md: batch intake workflow, validation steps, known
limitations
Corpus contradictions flagged in the entries (not silently resolved):
AMS replacement rate for blocked Altersteilzeit 2026 (lb-atz-07 28.5%
vs lb-atz-09/12 27.5%), DAG limit 2023 (751.36 vs 751.37 EUR),
Kurzarbeit for Lehrlinge (lb-leh-11 vs lb-leh-07); lehrlinge cluster
partly older Stands (2024-03/2025-08).
Validated: extract 55/55, registry 55 entries/6 clusters, check 0
problems, structural scan of all 55 curated entries, py_compile.
fegger