dschlueter
b74445b8d6
Übersetzt natürlichsprachige Optimierungsaufgaben (Deutsch) per LLM in eine geprüfte Modellspezifikation, erzeugt Solver-Code und führt ihn in einer bubblewrap-Sandbox aus. Deterministische Prüfungen ohne LLM, Report als Markdown/HTML, Benchmark mit 19 verifizierten Fällen. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
207 lines
10 KiB
Python
207 lines
10 KiB
Python
"""Verständliche Ergebnistabellen: Ausgabe im Programm, Übernahme ins Ergebnis, Summenprüfung, Report."""
|
||
import pytest
|
||
|
||
from model_schema import PipelineResult, SolverRun, SolverStatus
|
||
from report import build_document, to_html
|
||
from solver_stage import parse_result_tables, run_solver_code
|
||
from test_model_checks import make_spec
|
||
from test_review import make_review
|
||
|
||
# Portfolio-Ergebnis (hard_portfolio_kardinalitaet); Beiträge summieren sich zum Zielwert 0.09756
|
||
PORTFOLIO_TABLE = {
|
||
"title": "Portfolio",
|
||
"columns": ["Aktie", "Anteil", "Rendite", "Beitrag zum Ziel"],
|
||
"rows": [
|
||
["A02", "30,00 %", "14,4 %", 0.03433],
|
||
["A08", "30,00 %", "13,3 %", 0.03093],
|
||
["A03", "28,43 %", "11,5 %", 0.02587],
|
||
["A04", "6,20 %", "9,8 %", 0.00326],
|
||
["A09", "5,37 %", "9,8 %", 0.00317],
|
||
["Summe", "100 %", "", 0.09756],
|
||
],
|
||
"objective_column": "Beitrag zum Ziel",
|
||
"note": "Anteil = Kapitalanteil je gewählter Aktie.",
|
||
}
|
||
|
||
|
||
def test_table_is_printed_and_parsed_without_warnings():
|
||
code = f"report('optimal', 0.09756, {{'w[A02]': 0.3}}, tables=[{PORTFOLIO_TABLE!r}])"
|
||
run = run_solver_code(code)
|
||
assert run.status == SolverStatus.OPTIMAL
|
||
assert run.warnings == []
|
||
assert run.tables[0].title == "Portfolio" and len(run.tables[0].rows) == 6
|
||
|
||
|
||
def test_standalone_output_shows_box_table(tmp_path):
|
||
import subprocess, sys
|
||
from solver_stage import assemble_script
|
||
path = tmp_path / "solve.py"
|
||
path.write_text(assemble_script(f"report('optimal', 0.09756, {{}}, tables=[{PORTFOLIO_TABLE!r}])"), encoding="utf-8")
|
||
out = subprocess.run([sys.executable, str(path)], capture_output=True, text=True, timeout=60).stdout
|
||
assert "┌" in out and "│ A02 │ 30,00 % │" in out and "Anteil = Kapitalanteil" in out
|
||
|
||
|
||
def test_sum_mismatch_is_warned():
|
||
tables, warnings = parse_result_tables([PORTFOLIO_TABLE], objective_value=0.2)
|
||
assert len(tables) == 1
|
||
assert any("weicht vom Zielwert" in w for w in warnings)
|
||
|
||
|
||
def test_inconsistent_sum_row_is_warned():
|
||
table = dict(PORTFOLIO_TABLE, rows=PORTFOLIO_TABLE["rows"][:-1] + [["Summe", "100 %", "", 0.5]])
|
||
_, warnings = parse_result_tables([table], objective_value=0.09756)
|
||
assert any("Summenzeile" in w for w in warnings)
|
||
|
||
|
||
@pytest.mark.parametrize("table, expected", [
|
||
(dict(PORTFOLIO_TABLE, objective_column="Gibt es nicht"), "existiert nicht"),
|
||
(dict(PORTFOLIO_TABLE, rows=[["A02", "30 %", "14,4 %", "0,034"]]), "nicht nur Zahlen"),
|
||
({"title": "kaputt"}, "ungültig"),
|
||
])
|
||
def test_faulty_tables_are_warnings_not_errors(table, expected):
|
||
_, warnings = parse_result_tables([table], objective_value=0.09756)
|
||
assert any(expected in w for w in warnings)
|
||
|
||
|
||
def test_non_additive_objective_without_objective_column_is_fine():
|
||
table = dict(PORTFOLIO_TABLE, objective_column=None)
|
||
_, warnings = parse_result_tables([table], objective_value=0.117) # z. B. Volatilität
|
||
assert warnings == []
|
||
|
||
|
||
def test_report_shows_tables_and_grouped_variables_with_meaning():
|
||
spec = make_spec(variables=[
|
||
{"name": "w_i", "var_type": "continuous", "domain": "[0,1]", "description": "Kapitalanteil in Aktie i"},
|
||
{"name": "y_i", "var_type": "binary", "domain": "{0,1}", "description": "1, falls Aktie i ausgewählt wird"},
|
||
])
|
||
tables, warnings = parse_result_tables([PORTFOLIO_TABLE], 0.09756)
|
||
result = PipelineResult(
|
||
task="Aufgabe", specification=spec, review=make_review(), revisions=0,
|
||
solver_run=SolverRun(status=SolverStatus.OPTIMAL, objective_value=0.09756, tables=tables, warnings=warnings,
|
||
solution={"w[A02]": 0.6, "y[A02]": 1.0, "w[A08]": 0.4, "y[A08]": 1.0}),
|
||
)
|
||
doc = build_document(result)
|
||
headings = [b[1] for b in doc if b[0] in ("h3", "h4")]
|
||
assert headings.index("Portfolio") < headings.index("Alle Variablenwerte (≠ 0)")
|
||
assert "w – Kapitalanteil in Aktie i" in headings and "y – 1, falls Aktie i ausgewählt wird" in headings
|
||
group_tables = [b for b in doc if b[0] == "table" and b[1] == ["Variable", "Wert"]]
|
||
assert [t[2][-1] for t in group_tables] == [["Summe", "1"], ["Summe", "2"]]
|
||
assert "Anteil = Kapitalanteil" in to_html(doc)
|
||
|
||
|
||
def test_report_shows_table_warning_prominently():
|
||
tables, warnings = parse_result_tables([PORTFOLIO_TABLE], 0.2)
|
||
result = PipelineResult(
|
||
task="Aufgabe", specification=make_spec(), review=make_review(), revisions=0,
|
||
solver_run=SolverRun(status=SolverStatus.OPTIMAL, objective_value=0.2, tables=tables, warnings=warnings),
|
||
)
|
||
assert '<div class="warn">⚠ Tabelle' in to_html(build_document(result))
|
||
|
||
|
||
DECOMPOSITION = { # beobachtet: Summenzeile heißt "Zielwert", nicht "Summe"
|
||
"title": "Zielwert-Zerlegung",
|
||
"columns": ["Bestandteil", "Wert"],
|
||
"rows": [["Erwartete Rendite", 0.127133], ["Risiko (lambda * Varianz)", -0.006376],
|
||
["Fixkosten (Transaktion)", -0.0232], ["Zielwert", 0.097557]],
|
||
"objective_column": "Wert",
|
||
}
|
||
|
||
|
||
def test_sum_row_recognised_by_label_zielwert():
|
||
tables, warnings = parse_result_tables([DECOMPOSITION], objective_value=0.097557)
|
||
assert warnings == [] and len(tables) == 1
|
||
|
||
|
||
def test_sum_row_recognised_by_value_when_label_is_unusual():
|
||
table = dict(DECOMPOSITION, rows=DECOMPOSITION["rows"][:-1] + [["Netto nach Abzügen", 0.097557]])
|
||
_, warnings = parse_result_tables([table], objective_value=0.097557)
|
||
assert warnings == []
|
||
|
||
|
||
def test_real_sum_error_is_still_found():
|
||
table = dict(DECOMPOSITION, rows=[["Rendite", 0.2], ["Kosten", -0.05], ["Zielwert", 0.15]])
|
||
_, warnings = parse_result_tables([table], objective_value=0.097557)
|
||
assert any("weicht vom Zielwert" in w for w in warnings)
|
||
|
||
|
||
def test_negligible_term_hints_at_mixed_units():
|
||
# beobachtet: Risiko 0,0066 (dimensionslos) neben Rendite 12.808 € und Gebühren 186 €
|
||
table = {"title": "Zielwert-Zerlegung", "columns": ["Bestandteil", "Wert"],
|
||
"rows": [["Erwartete Rendite", 12808.80], ["Risiko", -0.0066], ["Gebühren", -185.62],
|
||
["Zielwert", 12623.17]],
|
||
"objective_column": "Wert"}
|
||
_, warnings = parse_result_tables([table], objective_value=12623.17)
|
||
assert any("Größenordnungen kleiner" in w and "Risiko" in w for w in warnings)
|
||
|
||
|
||
def test_normal_decomposition_has_no_magnitude_warning():
|
||
_, warnings = parse_result_tables([DECOMPOSITION], objective_value=0.097557)
|
||
assert warnings == []
|
||
|
||
|
||
def test_report_shows_multistart_stability():
|
||
from model_schema import Multistart, SolverRun
|
||
from report import build_document, to_markdown
|
||
from test_pipeline import FakeSession
|
||
from TextToModel import run_pipeline
|
||
from test_review import make_review
|
||
from test_script_export import STRUCTURED_CODE
|
||
from model_schema import SolverCode
|
||
|
||
session = FakeSession([make_spec()], [make_review()], [SolverCode(code=STRUCTURED_CODE)])
|
||
result = run_pipeline("Aufgabe", session=session, max_revisions=0, max_restarts=0)
|
||
result.solver_run.multistart = Multistart(starts=5, converged=4, best=1.5, worst=9.0, mean=3.2, stddev=2.8)
|
||
markdown = to_markdown(build_document(result))
|
||
assert "Stabilität über mehrere Startpunkte" in markdown
|
||
assert "nicht nachweislich das globale Optimum" in markdown
|
||
|
||
|
||
def test_transliterated_umlauts_are_reported():
|
||
# beobachtet: "Der guenstigste Ersatzplan", Tabelle "Gewaehlte Nutzungszeitraeume"
|
||
from model_schema import ResultTable, SolverRun, SolverStatus
|
||
from solver_stage import check_text_quality
|
||
run = SolverRun(status=SolverStatus.OPTIMAL, headline="Der guenstigste Plan kostet 29 €.",
|
||
tables=[ResultTable(title="Gewaehlte Zeitraeume", columns=["Kauf"], rows=[])])
|
||
befunde = check_text_quality(run)
|
||
assert any("guenstigste" in b for b in befunde)
|
||
assert any("Gewaehlte" in b and "Zeitraeume" in b for b in befunde)
|
||
|
||
|
||
def test_correct_german_is_not_flagged():
|
||
# Gegenprobe: echte Wörter mit ue/oe/ae dürfen keinen Fehlalarm auslösen
|
||
from model_schema import ResultTable, SolverRun, SolverStatus
|
||
from solver_stage import check_text_quality
|
||
run = SolverRun(status=SolverStatus.OPTIMAL,
|
||
headline="Die Quelle liefert neue Koeffizienten für tägliche Werte.",
|
||
tables=[ResultTable(title="Gewählte Zeiträume", columns=["Menge", "Beitrag"], rows=[])])
|
||
assert check_text_quality(run) == []
|
||
|
||
|
||
def test_transliterations_are_repaired_after_being_reported():
|
||
# Die Warnung dokumentiert den Prompt-Verstoss, der Report soll trotzdem lesbar sein
|
||
from model_schema import Answer, ResultTable, SolverRun, SolverStatus
|
||
from solver_stage import check_text_quality, fix_transliterations, repair_texts
|
||
run = SolverRun(status=SolverStatus.OPTIMAL, headline="Der guenstigste Plan kostet 29 €.",
|
||
answers=[Answer(question_id="a", text="Die Loesung waehlt zwei Aktien.")],
|
||
tables=[ResultTable(title="Gewaehlte Zeitraeume", columns=["Naechste Periode"],
|
||
rows=[["gewaehlt", 3]], note="Gemaess Aufgabe.")])
|
||
assert check_text_quality(run) # erst melden
|
||
repair_texts(run) # dann beheben
|
||
assert run.headline == "Der günstigste Plan kostet 29 €."
|
||
assert run.answers[0].text == "Die Lösung wählt zwei Aktien."
|
||
assert run.tables[0].title == "Gewählte Zeiträume"
|
||
assert run.tables[0].columns == ["Nächste Periode"]
|
||
assert run.tables[0].rows == [["gewählt", 3]]
|
||
assert run.tables[0].note == "Gemäß Aufgabe."
|
||
|
||
|
||
def test_repair_leaves_names_and_english_alone():
|
||
from solver_stage import fix_transliterations
|
||
assert fix_transliterations("Mueller AG, Quelle, value, Koeffizient") == "Mueller AG, Quelle, value, Koeffizient"
|
||
assert fix_transliterations("") == ""
|
||
|
||
|
||
def test_repair_covers_stems_seen_in_the_lot_sizing_case():
|
||
# beobachtet in wagner_whitin_dijkstra: "110 Baelle", "Das heisst:" im Antworttext
|
||
from solver_stage import fix_transliterations
|
||
assert fix_transliterations("Das heisst: 110 Baelle, der Ball") == "Das heißt: 110 Bälle, der Ball"
|