Validate FLM directional evidence

openpilot/starpilot/flm/event_physical.py

@@ -39,10 +39,11 @@ def fit_event_physical(identity, context, rows, *, lag_ns, now_ns, min_excitatio
   # never creates another independent observation or another heldout ID.
   for group in (train, heldout):
     for direction in ('left', 'right'):
-      supporting = {row.block_id for row in group if row.direction == direction}
-      if sum(row.direction == direction for row in group) < 80:
+      selected = tuple(row for row in group if (row.command > 0 if direction == 'left' else row.command < 0))
+      supporting = {row.block_id for row in selected}
+      if len(selected) < 80:
         return None
-      if sum(sum(row.block_id == block and row.direction == direction for row in group) >= 20 for block in supporting) < 2:
+      if sum(sum(row.block_id == block for row in selected) >= 20 for block in supporting) < 2:
         return None
   mean_x = sum(row.command for row in train) / len(train)
   mean_y = sum(row.physical_response for row in train) / len(train)

openpilot/starpilot/flm/physical_response.py

@@ -96,6 +96,7 @@ def fit_physical(identity, domain, context, rows, *, lag_ns, min_excitation, max
   for row in rows:
     if (not row.healthy or row.overridden or row.saturated or row.context != context or
         row.direction not in ('left', 'right') or
+        row.command > 0 and row.direction != 'left' or row.command < 0 and row.direction != 'right' or
         type(row.command_ns) is not int or type(row.response_ns) is not int or
         not 0 < row.command_ns <= row.response_ns <= now_ns or row.command_ns < now_ns - history_ns or not math.isfinite(row.command) or
         not math.isfinite(row.physical_response) or row.command_ns <= previous or
@@ -105,7 +106,8 @@ def fit_physical(identity, domain, context, rows, *, lag_ns, min_excitation, max
   train = [row for row in rows if row.block_id in train_ids]
   heldout = [row for row in rows if row.block_id in heldout_ids]
   for group in (train, heldout):
-    if any(len({row.block_id for row in group if row.direction == direction}) < 2 for direction in ('left', 'right')):
+    if any(len({row.block_id for row in group if (row.command > 0 if direction == 'left' else row.command < 0)}) < 2
+           for direction in ('left', 'right')):
       return None
   mean_x = sum(row.command for row in train) / len(train)
   mean_y = sum(row.physical_response for row in train) / len(train)

openpilot/starpilot/flm/tests/test_bounded_stage_regression.py

@@ -28,7 +28,7 @@ def known_fit(block_count=8):
       (-1 if block % 2 else 1) * (0.1 + j * 0.01),
       2 * (-1 if block % 2 else 1) * (0.1 + j * 0.01) + 0.01,
       'calibration-v1',
-      'left' if block % 2 else 'right',
+      'right' if block % 2 else 'left',
       True,
       False,
       False,

openpilot/starpilot/flm/tests/test_event_physical.py

@@ -56,3 +56,30 @@ def test_local_model_does_not_extrapolate_or_accept_quiet_unidentified_input():
   unsupported = tuple(replace(row, command=row.command * 2, physical_response=row.physical_response * 2)
                       if row.block_id in ('4', '5', '6', '7') else row for row in rows)
   assert fit_event_physical('a' * 64, 'physical-context', unsupported, lag_ns=50_000_000, now_ns=unsupported[-1].response_ns) is None
+
+
+def test_zero_and_negative_zero_do_not_supply_negative_direction_support():
+  for zero in (0.0, -0.0):
+    for change_train in (True, False):
+      rows = tuple(replace(row, command=zero, physical_response=.002, direction='right')
+                   if row.command < 0 and (change_train or row.block_id in ('4', '5', '6', '7')) else row
+                   for row in episodes())
+      assert fit_event_physical('a' * 64, 'physical-context', rows,
+                                lag_ns=50_000_000, now_ns=rows[-1].response_ns) is None
+
+
+def test_tiny_positive_has_no_negative_direction_support_and_no_epsilon_is_used():
+  for tiny, expected in ((1e-15, False), (-1e-15, True)):
+    rows = tuple(replace(row, command=tiny, physical_response=1.98 * tiny + .002,
+                         direction='left' if tiny > 0 else 'right') if row.command < 0 else row
+                 for row in episodes())
+    model = fit_event_physical('a' * 64, 'physical-context', rows,
+                               lag_ns=50_000_000, now_ns=rows[-1].response_ns)
+    assert (model is not None) is expected
+
+
+def test_true_bilateral_original_control_still_qualifies():
+  rows = episodes()
+  assert any(row.command > 0 for row in rows) and any(row.command < 0 for row in rows)
+  assert fit_event_physical('a' * 64, 'physical-context', rows,
+                            lag_ns=50_000_000, now_ns=rows[-1].response_ns) is not None

openpilot/starpilot/flm/tests/test_physical_direction_support.py

@@ -0,0 +1,77 @@
+"""Neutral commands and direction labels cannot manufacture bilateral evidence."""
+from dataclasses import replace
+import math
+
+import pytest
+
+from openpilot.starpilot.flm.event_physical import fit_event_physical
+from openpilot.starpilot.flm.physical_response import PhysicalRow, fit_physical
+
+
+IDENTITY = 'a' * 64
+CONTEXT = 'exact-independent-sign-support'
+LAG = 50_000_000
+
+
+def cohort(signs):
+  rows = []
+  for block, sign in enumerate(signs):
+    for sample in range(80):
+      stamp = 1_000_000_000 + (block * 80 + sample) * 50_000_000
+      command = sign * (.10 + sample * .004)
+      rows.append(PhysicalRow(str(block), stamp, stamp + LAG, command,
+                              1.98 * command + .002, CONTEXT,
+                              'left' if command > 0 else 'right', True, False, False))
+  return tuple(rows)
+
+
+def event_fit(rows):
+  return fit_event_physical(IDENTITY, CONTEXT, rows, lag_ns=LAG, now_ns=rows[-1].response_ns)
+
+
+@pytest.mark.parametrize('neutral', [0.0, -0.0])
+def test_two_positive_two_neutral_episodes_per_chronological_half_cannot_fit(neutral):
+  rows = cohort((1., 1., neutral, neutral, 1., 1., neutral, neutral))
+  assert len(rows) == 640 and len({row.block_id for row in rows}) == 8
+  for half in (rows[:320], rows[320:]):
+    assert sum(row.command > 0 for row in half) == 160
+    assert sum(row.command == 0 for row in half) == 160
+    assert not any(row.command < 0 for row in half)
+    assert sum(row.direction == 'right' for row in half) == 160
+  # Independent measured-response rule has ample positive-vs-zero variance
+  # and exact heldout OLS; the sole missing qualification is negative support.
+  train = rows[:320]
+  mean = sum(row.command for row in train) / len(train)
+  assert sum((row.command - mean)**2 for row in train) / len(train) > .05**2
+  assert all(math.isclose(row.physical_response, 1.98 * row.command + .002, abs_tol=1e-15) for row in rows)
+  assert event_fit(rows) is None
+
+
+@pytest.mark.parametrize('neutral', [0.0, -0.0])
+def test_neutral_observations_remain_valid_when_real_bilateral_support_exists(neutral):
+  data = cohort((1., -1., 1., -1., 1., -1., 1., -1.))
+  # Keep60 substantive sign rows perblock, hence120/sign/half and2independent
+  # blocks, while preserving all neutral points inOLS/heldout/error evaluation.
+  rows = tuple(replace(row, command=neutral, physical_response=.002, direction='right')
+               if index % 80 < 20 else row for index, row in enumerate(data))
+  model = event_fit(rows)
+  assert model is not None
+  assert math.isclose(model.slope, 1.98, abs_tol=1e-12)
+  assert model.train_blocks == ('0', '1', '2', '3')
+  assert model.heldout_blocks == ('4', '5', '6', '7')
+  assert model.heldout_rmse < 1e-12
+
+
+def test_real_negative_rows_only_in_one_episode_do_not_become_two_with_zero_episode():
+  rows = cohort((1., -1., 1., 0., 1., -1., 1., 0.))
+  assert event_fit(rows) is None
+
+
+def test_legacy_wrong_sign_labels_cannot_create_bilateral_support():
+  # Unlike event rows, legacyblocks must each have one direction. Preserve
+  # chronology, nonzero variance and exactphysicalOLS, but forge half labels.
+  data = cohort((1., 1., 1., 1., 1., 1., 1., 1.))
+  rows = tuple(replace(row, direction='left' if int(row.block_id) % 2 == 0 else 'right') for row in data)
+  assert all(row.command > 0 for row in rows)
+  assert fit_physical(IDENTITY, 'normalized_torque', CONTEXT, rows, lag_ns=LAG,
+                      min_excitation=.05, max_rmse=.05, now_ns=rows[-1].response_ns) is None

openpilot/starpilot/flm/tests/test_physical_response.py

@@ -10,7 +10,7 @@ def rows():
                            1_110_000_000 + (block * 40 + index) * 50_000_000,
                            (-1 if block % 2 else 1) * (.1 + index * .005),
                            2 * (-1 if block % 2 else 1) * (.1 + index * .005) + .01,
-                           'stable_vm_roll_lag', 'left' if block % 2 else 'right', True, False, False)
+                           'stable_vm_roll_lag', 'right' if block % 2 else 'left', True, False, False)
                for block in range(8) for index in range(40))