Show retained Auto Learn study progress

openpilot/starpilot/flm/live_tracking.py

@@ -1484,10 +1484,17 @@ class LiveTrackingOptimizer:
       and 0 <= now_ns - self.last_ack[-1] <= 100_000_000
     )
     status, reason = self.status, self.reason
+    paused = self.paused_intent
+    if (self.paused_demand_context != (self.binding, self.certificate, self.domain)
+        or paused is None or paused[0]['baseline'] != self.accepted or type(now_ns) is not int or now_ns >= paused[0]['deadline_ns']):
+      paused = None
+    before, future = (self.before, self.future) if paused is None else paused[1:]
     if state is not None and state.get('accepted', False):
       if acknowledged:
         status = state['phase']
         reason = 'accepted_target_currently_acknowledged' if status == 'confirmed' else 'prior_accepted_baseline_restored'
+        if status == 'rollback' and paused is not None and self.paused_baseline_ns > 0:
+          reason = 'original_input_regime_after_baseline_ack_required'
       else:
         state = None
 
@@ -1503,14 +1510,14 @@ class LiveTrackingOptimizer:
       'trialId': None if state is None else state['trial_id'],
       'phase': None if state is None else state['phase'],
       'requiredEpisodes': None if state is None else state['required_episodes'],
-      'beforeEpisodes': len(self.before),
-      'futureEpisodes': len(self.future),
+      'beforeEpisodes': len(before),
+      'futureEpisodes': len(future),
       'modelStatus': self.model_status if current else 'stale',
       'lagMs': self.lag_ns / 1e6,
-      'baselineErrorRms': mean(self.before, 'error_rms'),
-      'futureErrorRms': mean(self.future, 'error_rms'),
-      'baselineReversalHz': mean(self.before, 'reversal_hz'),
-      'futureReversalHz': mean(self.future, 'reversal_hz'),
+      'baselineErrorRms': mean(before, 'error_rms'),
+      'futureErrorRms': mean(future, 'error_rms'),
+      'baselineReversalHz': mean(before, 'reversal_hz'),
+      'futureReversalHz': mean(future, 'reversal_hz'),
       'applicationAcknowledged': acknowledged,
       'applicationSettledMonoTime': self.last_ack[-1] if acknowledged and self.last_ack is not None else self.settled_ns,
       'lastCorrectionMonoTime': self.last_correction or None,

openpilot/starpilot/flm/tests/test_staged_report.py

@@ -0,0 +1,49 @@
+"""Retained numerical progress must not become current target authority."""
+
+from dataclasses import replace
+
+from openpilot.starpilot.flm.tests.test_staged_source_study import pause, baseline_ack
+
+
+def test_held_progress_does_not_claim_target_or_confirmed_application():
+  o, _original, scores, incoming = pause()
+  report = o.report(incoming.end_ns)
+  assert report['phase'] == 'rollback' and not report['applicationAcknowledged']
+  assert (report['beforeEpisodes'], report['futureEpisodes'], report['requiredEpisodes']) == (3, 2, 3)
+  assert report['reason'] == 'trial_input_regime_ended' and not any(o.accepted.values())
+  ack = baseline_ack(o, incoming.end_ns + 10_000_000)
+  report = o.report(ack)
+  assert report['status'] == report['phase'] == 'rollback'
+  assert report['applicationAcknowledged'] and report['reason'] == 'original_input_regime_after_baseline_ack_required'
+  assert (report['beforeEpisodes'], report['futureEpisodes']) == (len(scores[0]), len(scores[1]))
+  assert report['lastCorrectionMonoTime'] is None and not any(o.accepted.values())
+
+
+def test_stale_baseline_does_not_report_current_restoration():
+  o, _original, _scores, incoming = pause()
+  ack = baseline_ack(o, incoming.end_ns + 10_000_000)
+  report = o.report(ack + 100_000_001)
+  assert not report['applicationAcknowledged'] and not report['bindingCurrent']
+  assert report['phase'] is None and report['modelStatus'] == 'stale'
+
+
+def test_expired_changed_or_discarded_study_has_no_retained_progress():
+  for change in ('expiry', 'certificate', 'baseline', 'manual'):
+    o, original, _scores, incoming = pause()
+    now = baseline_ack(o, incoming.end_ns + 10_000_000)
+    if change == 'expiry':
+      now = original['deadline_ns']
+    elif change == 'certificate':
+      o.certificate = replace(o.certificate, model_revision='9' * 64)
+    elif change == 'baseline':
+      o.accepted['ff_gain_left'] = 0.001
+    else:
+      o.interrupt('manual_override')
+    report = o.report(now)
+    assert report['beforeEpisodes'] == report['futureEpisodes'] == 0
+  for invalid_now in (None, True, 'now', float('nan')):
+    o, _original, _scores, incoming = pause()
+    baseline_ack(o, incoming.end_ns + 10_000_000)
+    report = o.report(invalid_now)
+    assert report['beforeEpisodes'] == report['futureEpisodes'] == 0
+    assert not report['applicationAcknowledged'] and not report['bindingCurrent']

openpilot/starpilot/galaxy/tests/test_flm_tracking_results.mjs

@@ -1,7 +1,7 @@
 import assert from 'node:assert/strict'
 import { compile } from '../web/vendor/vue/vue.esm-browser.js'
 import { FlmPage } from '../web/js/flm.js'
-import { SURFACE_NAMES, PID_UNSUPPORTED, validTrainingResults, validTrackingResults, trackingView, surfaceRows, measurementText, validTrackingValues, trackingCoordinates, validTrackingPlot, trackingPlotView, sourceSeconds } from '../web/js/flm-training-results.js'
+import { SURFACE_NAMES, PID_UNSUPPORTED, validTrainingResults, validTrackingResults, trackingView, surfaceRows, measurementText, validTrackingValues, trackingCoordinates, validTrackingPlot, trackingPlotView, sourceSeconds, tunerStatus } from '../web/js/flm-training-results.js'
 
 const coordinates = amount => SURFACE_NAMES.map((name, i) => [name, i === 0 ? amount : 0])
 const tracking = () => ({ version: 1, status: 'trial_active', reason: 'scoring_future_tracking', group: 'feedforward',
@@ -60,6 +60,24 @@ const neutralRollback = report(); neutralRollback.current.tracking.status = 'rol
 neutralRollback.current.application.applied = false
 neutralRollback.current.details['Actual full surface coordinates'] = coordinates(0)
 assert.equal(trackingView(neutralRollback, true).label, 'Correction rollback acknowledged')
+const held = structuredClone(neutralRollback)
+held.current.tracking.reason = 'original_input_regime_after_baseline_ack_required'
+held.current.tracking.futureEpisodes = 2
+assert.equal(trackingView(held, true).label, 'Baseline restored; waiting for matching road')
+assert.equal(tunerStatus(held, true).title, 'Waiting for matching road')
+assert.equal(trackingView(held, true).futureEpisodes, 2)
+assert.equal(trackingView(held, true).requiredEpisodes, 3)
+assert(!/confirmed|saved|correction applied/i.test(tunerStatus(held, true).title))
+for (const mutate of [value => value.current.tracking.applicationAcknowledged = false,
+  value => value.current.tracking.bindingCurrent = false,
+  value => value.current.fresh = false,
+  value => value.current.tracking.phase = 'trial',
+  value => delete value.current.details['Actual full surface coordinates']]) {
+  const noProof = structuredClone(held); mutate(noProof)
+  assert.equal(trackingView(noProof, true).acknowledged, false)
+  assert.notEqual(tunerStatus(noProof, true).title, 'Waiting for matching road')
+}
+assert.notEqual(tunerStatus(held, false).title, 'Waiting for matching road')
 for (const mutate of [value => value.current.tracking.applicationAcknowledged = false,
   value => value.current.tracking.bindingCurrent = false,
   value => value.current.tracking.applicationSettledMonoTime = null,

openpilot/starpilot/galaxy/web/js/flm-training-results.js

@@ -175,6 +175,7 @@ const REASONS = {
   trial_oscillation_batch_worse: "Comparable trial windows showed more oscillation, so the correction was rolled back.",
   predeclared_independent_future_look_required: "Collecting the next planned set of comparison windows.",
   prior_accepted_baseline_restored: "The controller returned to its prior baseline.",
+  original_input_regime_after_baseline_ack_required: "The baseline is restored. Completed comparison samples are kept while waiting for the same driving conditions.",
   scoring_later_matched_whole_episodes: "Comparing new steering windows with the baseline.",
   accepted_target_currently_acknowledged: "The controller acknowledged the accepted correction.",
 }
@@ -204,10 +205,12 @@ export function tunerStatus(report, live = false, choiceLive = false, error = ""
   if ((tracking?.reason || current.reason) === "controller_collection_required")
     return { key: "waiting", tone: "muted", title: "Waiting for active steering", description: learningReason("controller_collection_required") }
   if (tracking && !tracking.bindingCurrent)
-    return { key: "paused", tone: "warning", title: "Learning paused", description: "The steering context changed. A fresh qualified window is required." }
+    return { key: "paused", tone: "warning", title: "Learning paused", description: "Current steering evidence is unavailable. A fresh qualified window is required." }
   if (tracking?.acknowledged && tracking.status === "confirmed") return { key: "confirmed", tone: "success", title: "Correction applied", description: "The comparison passed and the controller acknowledged this correction. Disk saving is a separate result." }
   if (tracking?.acknowledged && tracking.status === "trial_active") return { key: "testing", tone: "active", title: "Testing a small correction", description: "The controller acknowledged the trial. Comparable new driving windows decide whether to keep it." }
   if (tracking?.status === "trial_pending") return { key: "pending", tone: "active", title: "Waiting for the controller", description: "A trial was offered. It has not yet been acknowledged as applied." }
+  if (tracking?.status === "rollback" && tracking.acknowledged && tracking.reason === "original_input_regime_after_baseline_ack_required")
+    return { key: "waiting", tone: "ready", title: "Waiting for matching road", description: learningReason(tracking.reason) }
   if (tracking?.status === "rollback") return { key: "rollback", tone: "warning", title: "Returning to the baseline", description: tracking.acknowledged ? "The controller acknowledged the rollback." : "A rollback was requested. Current application is not yet acknowledged." }
   if (tracking?.status === "paused" || current.status === "unavailable") return { key: "paused", tone: "warning", title: "Learning paused", description: learningReason(tracking?.reason || current.reason) }
   return { key: "collecting", tone: "active", title: tracking?.status === "calibrating" ? "Measuring steering response" : "Collecting driving samples", description: learningReason(tracking?.reason || current.reason) }
@@ -256,8 +259,9 @@ export function trackingView(report, live = false) {
   if (value.status === 'trial_active' && acknowledged) label = 'Evaluating correction'
   if (value.status === 'confirmed' && acknowledged) label = 'Confirmed correction acknowledged'
   if (value.status === 'rollback' && acknowledged) label = 'Correction rollback acknowledged'
+  if (value.status === 'rollback' && acknowledged && value.reason === 'original_input_regime_after_baseline_ack_required') label = 'Baseline restored; waiting for matching road'
   if (value.status !== 'paused' && (value.phase !== null && !matchingPhase || expectedPhase !== undefined && !matchingPhase)) label = 'Learning state reported; trial application unknown'
-  if (live && !value.bindingCurrent) label = 'Learning paused; steering context changed'
+  if (live && !value.bindingCurrent) label = 'Learning paused; current steering evidence unavailable'
   if (!live) label = 'Earlier learning report; current application unknown'
   return { ...value, label, acknowledged, changedParameters: value.names.map(readable).join(', '),
     measurements: [

prebuilt.json

@@ -1 +1 @@
-{"protected_inventory_sha256": "9653f67cf25c7e37488533864943c5949cca94b4c7dadbe1ad1af400eb36d88d", "schema_version": 1}
+{"protected_inventory_sha256": "a65a390042e7d02f1180f79f6ad3e95c80ce96be9f44cc1230ff7afe2fad885c", "schema_version": 1}