Files
Jyotisha/tests/test_reported_offset_research.py
T
jesse-uxandClaude Code 932f2fffba
Independent Staging Quality Gate / validate (push) Successful in 12m7s
Independent Staging Quality Gate / publish (push) Successful in 3m46s
research(rectification): add reported-offset evaluation and frozen rerun integrity
Preserve closed confirmation gates and previously-exposed dataset boundaries. Add auditable 900-trial sensitivity results, current scorer freshness checks, and the v5 collection protocol.

Co-Authored-By: Claude Code <noreply@anthropic.com>
2026-09-20 12:03:52 +08:00

118 lines
6.1 KiB
Python

from __future__ import annotations
import json
from datetime import datetime, timedelta
import pytest
from scripts.research import reported_offset_sweep as sweep
from scripts.research.sealed_holdout_rerun import DATASET, opaque_order
def public_case():
return json.loads(DATASET.read_text(encoding="utf-8"))["cases"][0]
def test_zero_offset_centres_on_truth_and_has_complete_grid():
case = public_case()
truth = datetime.fromisoformat(f"{case['birth']['date']}T{case['birth']['time']}:00")
request, candidates = sweep.shifted_window(case, 0, 15)
assert candidates[15] == truth
assert len(candidates) == 31
assert request["start_time"] == candidates[0].strftime("%H:%M")
assert request["end_time"] == candidates[-1].strftime("%H:%M")
assert truth in candidates
assert not any("truth" in key or "offset" in key for key in request)
@pytest.mark.parametrize("offset", [-30, -20, 20, 30])
def test_offset_beyond_radius_has_no_truth_candidate(offset):
case = public_case()
truth = datetime.fromisoformat(f"{case['birth']['date']}T{case['birth']['time']}:00")
_, candidates = sweep.shifted_window(case, offset, 15)
assert truth not in candidates
rows = [{"time": moment.strftime("%H:%M"), "score": 1} for moment in candidates]
metrics = sweep.reveal_metrics(rows, candidates, candidates, truth, "test", "test")
assert metrics["truth_in_window"] is False
assert metrics["true_rank"] is None
assert metrics["delivery_covers_truth"] is False
def test_opaque_tie_break_is_independent_of_truth_and_input_order():
# Explicitly synthetic unit inputs, not a fabricated engine-contract fixture.
moments = [datetime(2000, 1, 1, 12) + timedelta(minutes=i) for i in range(3)]
rows = [{"time": moment.strftime("%H:%M"), "score": 1} for moment in moments]
ordered = opaque_order("test", "synthetic", rows)
assert ordered == opaque_order("test", "synthetic", list(reversed(rows)))
winner = next(moment for moment in moments if moment.strftime("%H:%M") == ordered[0]["time"])
for truth in moments:
metrics = sweep.reveal_metrics(rows, moments, moments, truth, "test", "synthetic")
assert metrics["top_1_hit"] is (truth == winner)
assert metrics["true_rank"] == 1 + next(i for i, row in enumerate(ordered) if row["time"] == truth.strftime("%H:%M"))
@pytest.mark.parametrize("clock,offset", [("00:02", -5), ("23:58", 5)])
def test_shifted_window_preserves_dates_across_midnight(clock, offset):
case = public_case()
# Synthetic boundary mutation of the public case; no real user's birth data.
case["birth"] = {**case["birth"], "time": clock}
request, candidates = sweep.shifted_window(case, offset, 15)
truth = datetime.fromisoformat(f"{case['birth']['date']}T{clock}:00")
assert candidates[15] == truth + timedelta(minutes=offset)
assert request["birth_date"] == candidates[0].date().isoformat()
rows = [{"time": moment.strftime("%H:%M"), "score": int(moment == truth)} for moment in candidates]
metrics = sweep.reveal_metrics(rows, candidates, candidates, truth, "test", "synthetic")
assert metrics["top_1_minute_error"] == 0
assert metrics["truth_in_window"] is True
assert metrics["delivery_covers_truth"] is True
assert metrics["delivery_width_minutes"] == 31
def test_cross_midnight_real_engine_scores_match_candidate_date_replay():
case = json.loads(DATASET.read_text(encoding="utf-8"))["cases"][5]
request, candidates = sweep.shifted_window(case, 0, 60)
assert len({candidate.date() for candidate in candidates}) == 2
contexts = sweep.compute_candidate_static_contexts(request, candidates=candidates)
grouped = sweep.score_window(request, contexts)
expected = []
for context in contexts:
dated = {**request, "birth_date": context["candidate_at"].date().isoformat()}
built = sweep.build_event_contribution_matrix(dated, static_contexts=[context])
expected.extend(sweep.score_from_matrix(dated, built))
assert grouped == expected
old_matrix = sweep.build_event_contribution_matrix(request, static_contexts=contexts)
old_rows = sweep.score_from_matrix(request, old_matrix)
assert old_rows != expected
def test_recorded_specification_and_all_prespecified_cells():
report = json.loads((sweep.ROOT / "docs/research/reported_offset_2026_09_20.json").read_text(encoding="utf-8"))
spec = report["specification"]
assert spec["ayanamsa"] == "raman"
assert spec["node_mode"] == "mean"
assert spec["radii_minutes"] == list(sweep.RADII)
assert spec["offsets_minutes"] == list(sweep.OFFSETS)
assert spec["minute_step"] == 1
assert spec["implementation_sha256_prefix"] == spec["implementation_sha256"][:16]
assert spec["dataset_sha256"] == sweep.file_sha256(DATASET)
assert spec["evaluator_sha256"] == sweep.file_sha256(sweep.ROOT / "scripts/research/reported_offset_sweep.py")
assert spec["production_scoring_sha256"] == sweep.implementation_sha256(spec["production_scoring_files"])
assert spec["research_implementation_sha256"] == sweep.implementation_sha256(spec["research_files"])
assert spec["replay_revision"] == "candidate_date_grouped_v2"
assert spec["truth_hidden_from_ranker"] is True
assert spec["is_blind_evaluation"] is False
assert report["trial_count"] == 20 * len(sweep.RADII) * len(sweep.OFFSETS)
assert len(report["summary"]) == len(sweep.RADII) * len(sweep.OFFSETS)
assert report["summary"] == sweep.summarize(report["trials"], sweep.RADII, sweep.OFFSETS)
assert {(row["case_ordinal"], row["radius_minutes"], row["offset_minutes"]) for row in report["trials"]} == {
(case, radius, offset) for case in range(1, 21) for radius in sweep.RADII for offset in sweep.OFFSETS
}
assert report["widening_geometry"]["radius_120_scored"] is False
for trial in report["trials"]:
assert not ({"predicted_time", "birth", "latitude", "longitude"} & trial.keys())
for row in report["summary"]:
expected = float(abs(row["offset_minutes"]) <= row["radius_minutes"])
assert row["truth_in_window_rate"] == expected
assert row["delivery_coverage_rate"] <= expected
assert row["top_1_rate"] <= expected