dschlueter
0f1ab04a8d
- sandbox_warning: fehlt bubblewrap oder ist SOLVER_SANDBOX=aus, steht das im Log und in SolverRun.warnings (vorher lief der Code stillschweigend ungeschützt) - Dijkstra-Markierung: der Report erklärt den Begriff selbst (DIJKSTRA_EXPLANATION), check_text_quality zählt das als Erklärung; Stamm 'markier' -> 'markierung' - Pruef-Prompt Regel 7: entscheidungsunabhängige Terme ohne Zahlenwert dürfen weggelassen oder symbolisch mitgeführt werden (wagner_whitin_dijkstra: 1+1 statt 2+1 Korrekturrunden, n=2) - revision_reasons: jede Korrekturrunde nennt ihren Auslöser im Log Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
96 lines
4.3 KiB
Python
96 lines
4.3 KiB
Python
from core.model_schema import AssumptionReview, CheckIssue, CheckSeverity
|
||
from checks.review import select_best
|
||
|
||
|
||
def make_review(matches=True, verdict="justified", missing=(), units_consistent=True):
|
||
return AssumptionReview.model_validate({
|
||
"objective_check": {
|
||
"task_evidence": "Bestimme die Aufteilung mit der geringsten Volatilität",
|
||
"asked_quantity": "Standardabweichung der Portfoliorendite",
|
||
"modeled_quantity": "Standardabweichung" if matches else "Varianz",
|
||
"matches": matches,
|
||
"units_consistent": units_consistent,
|
||
"unit_explanation": "Rendite €, Risiko €, Kosten €" if units_consistent else "Rendite €, Risiko dimensionslos, Kosten €",
|
||
"explanation": "-",
|
||
},
|
||
"checks": [{"unclear_point": "-", "verdict": verdict, "task_evidence": "-", "explanation": "-"}],
|
||
"missing_assumptions": list(missing),
|
||
})
|
||
|
||
|
||
def test_clean_review_needs_no_revision():
|
||
assert not make_review().needs_revision
|
||
|
||
|
||
def test_objective_mismatch_triggers_revision():
|
||
assert make_review(matches=False).needs_revision
|
||
|
||
|
||
def test_faulty_assumption_triggers_revision():
|
||
assert make_review(verdict="stricter_than_task").needs_revision
|
||
|
||
|
||
def test_missing_assumption_triggers_revision():
|
||
missing = [{"unclear_point": "-", "assumption": "-", "alternative": "-", "impact": "-"}]
|
||
assert make_review(missing=missing).needs_revision
|
||
|
||
|
||
# --- Auswahl der besten Korrekturrunde (checks.review.select_best) ---
|
||
|
||
ERROR = CheckIssue(severity=CheckSeverity.ERROR, code="CPSAT_WITH_CONTINUOUS_VARS", message="-")
|
||
WARNING = CheckIssue(severity=CheckSeverity.WARNING, code="UNDECLARED_SYMBOL", message="-")
|
||
|
||
|
||
def test_later_hallucinated_round_is_discarded():
|
||
# beobachtet in portfolio_qp: Runde 1 sauber, Runde 2 beschreibt ein Fremdproblem mit Check-Error
|
||
candidates = [
|
||
(None, [], make_review(matches=False)),
|
||
(None, [WARNING], make_review()),
|
||
(None, [ERROR, WARNING], make_review(matches=False)),
|
||
]
|
||
assert select_best(candidates) == 1
|
||
|
||
|
||
def test_clean_later_round_wins_tie():
|
||
candidates = [(None, [], make_review()), (None, [WARNING], make_review())]
|
||
assert select_best(candidates) == 1
|
||
|
||
|
||
def test_fewer_errors_preferred_when_all_have_errors():
|
||
candidates = [(None, [ERROR], make_review(matches=False)), (None, [ERROR, ERROR], make_review(matches=False))]
|
||
assert select_best(candidates) == 0
|
||
|
||
|
||
def test_mixed_units_trigger_revision():
|
||
# beobachtet: Rendite/Kosten in Euro, Risiko als dimensionslose Summe w_i^2 sigma_i^2 (Faktor 100.000 zu klein)
|
||
assert make_review(units_consistent=False).needs_revision
|
||
|
||
|
||
# --- Grund jeder Korrekturrunde im Log (vorher nur „Probleme gefunden“) ---
|
||
|
||
def test_revision_reasons_name_every_trigger():
|
||
from checks.review import needs_work, revision_reasons
|
||
from core.model_schema import CheckIssue, CheckSeverity
|
||
fehlend = {"unclear_point": "Einkaufspreis nicht beziffert", "assumption": "-", "alternative": "-", "impact": "-"}
|
||
review = make_review(matches=False, verdict="weaker_than_task", missing=[fehlend])
|
||
issues = [CheckIssue(severity=CheckSeverity.ERROR, code="MIXED_OBJECTIVE_UNITS", message="Risiko dimensionslos"),
|
||
CheckIssue(severity=CheckSeverity.WARNING, code="MISSING_DATA", message="kein Wert für p")]
|
||
gruende = revision_reasons(review, issues)
|
||
assert needs_work(review, issues)
|
||
assert any(g.startswith("Zielgröße passt nicht") for g in gruende)
|
||
assert any("weaker_than_task" in g for g in gruende)
|
||
assert any("Einkaufspreis nicht beziffert" in g for g in gruende)
|
||
assert any("MIXED_OBJECTIVE_UNITS" in g for g in gruende)
|
||
assert not any("MISSING_DATA" in g for g in gruende) # löst keine Runde aus, gehört nicht in den Grund
|
||
|
||
|
||
def test_clean_state_has_no_revision_reasons():
|
||
from checks.review import revision_reasons
|
||
assert revision_reasons(make_review(), []) == []
|
||
|
||
|
||
def test_review_prompt_allows_both_forms_of_decision_independent_costs():
|
||
# wagner_whitin_dijkstra: Einkaufspreis ohne Zahl – je 1–2 Korrekturrunden, weil Regel 4 beide Formen beanstanden konnte
|
||
from pathlib import Path
|
||
prompt = (Path(__file__).parent.parent / "checks" / "Pruef-Prompt.txt").read_text(encoding="utf-8")
|
||
assert "entscheidungsunabhängige Terme" in prompt and "Verlange keine Umstellung" in prompt
|