Plan My Day v2: measured follow-up
Status: complete. Generated 2026-09-07T11:40:53.941193+00:00.
This report preserves the original pilot. It tests a corrected conversational teacher, matched four-action training probabilities, fixed-student mistake trajectories, and balanced replay.
Teacher signal and development-only selection
Development status: completed.
Common pre-travel SFT baseline, evaluated with the same phase-three memory on all 72 development days: 79.2% complete schedules, 0.167 hard violations/day, and 0.986 correction flags/day.
Warm-start model
| Prompt condition | Accuracy | Correct-action probability | Accuracy on current mistakes | Correct-action probability on current mistakes |
|---|---|---|---|---|
| unconditioned | 59.4% | 49.7% | 0.0% | 18.1% |
| legacy_malformed | 94.8% | 72.1% | 89.7% | 59.8% |
| corrected_conversational | 97.9% | 74.9% | 94.9% | 62.7% |
Observed current mistakes: 39. Feedback probability gain on those mistakes: +44.7%; corrected-versus-malformed placement: +2.9% (absolute probability differences).
Legacy placement only, with the identical richer v2 feedback; not v1 wording replay.
Common pre-travel SFT checkpoint
| Prompt condition | Accuracy | Correct-action probability | Accuracy on current mistakes | Correct-action probability on current mistakes |
|---|---|---|---|---|
| unconditioned | 70.8% | 67.8% | 0.0% | 13.8% |
| legacy_malformed | 94.8% | 84.5% | 82.1% | 68.8% |
| corrected_conversational | 97.9% | 92.7% | 92.9% | 80.5% |
Observed current mistakes: 28. Feedback probability gain on those mistakes: +66.7%; corrected-versus-malformed placement: +11.7% (absolute probability differences).
Legacy placement only, with the identical richer v2 feedback; not v1 wording replay.
Warm-start teacher gate: {'correct_probability_gain_on_mistakes': 0.4465137207164215, 'mistake_count': 39, 'passed': True, 'required_gain': 0.05}.
Pre-travel teacher gate: {'correct_probability_gain_on_mistakes': 0.6671588645216876, 'mistake_count': 28, 'passed': True, 'required_gain': 0.05}.
| Development candidate | Steps | Historical fraction | Distillation weight | Complete schedules | Hard violations/day | Correction flags/day | Promotion gate |
|---|---|---|---|---|---|---|---|
| develop-r25-w100-step64 | 64 | 25.0% | 1.0 | 68.1% | 0.264 | 1.153 | False |
| develop-r25-w100-step192 | 192 | 25.0% | 1.0 | 47.2% | 0.458 | 2.931 | False |
| develop-r25-w50-step64 | 64 | 25.0% | 0.5 | 68.1% | 0.278 | 1.250 | False |
| develop-r25-w50-step192 | 192 | 25.0% | 0.5 | 59.7% | 0.347 | 1.542 | False |
| develop-r50-w100-step64 | 64 | 50.0% | 1.0 | 59.7% | 0.125 | 1.722 | False |
| develop-r50-w100-step192 | 192 | 50.0% | 1.0 | 63.9% | 0.319 | 1.250 | False |
| develop-r50-w50-step64 | 64 | 50.0% | 0.5 | 63.9% | 0.153 | 1.736 | False |
| develop-r50-w50-step192 | 192 | 50.0% | 0.5 | 65.3% | 0.194 | 0.931 | False |
Frozen selection: {'distillation_weight': 1.0, 'lr': 3e-05, 'name': 'r50-w100', 'replay_fraction': 0.5, 'steps': 64}. Continued-SFT budget: 110 updates/phase.
Selection label: diagnostic only; no promotion; eligible development candidates: 0. approximate training wall time, excluding evaluation/upload callbacks; not FLOPs.
Confirmatory held-out results
72 shared test days; optimizer seeds 17, 29, 43. The frozen comparator is evaluated once per phase. All trained values below are means of seed-level means.
| Arm | After phase | Complete schedules | Conflict-free final schedules | Hard violations/day | Correction flags/day | Deferred tasks/day |
|---|---|---|---|---|---|---|
| Frozen SFT + memory | 1: normal | 63.9% | 69.4% | 0.375 | 1.556 | 0.111 |
| Frozen SFT + memory | 2: deadline | 61.1% | 66.7% | 0.472 | 1.556 | 0.083 |
| Frozen SFT + memory | 3: travel | 65.3% | 70.8% | 0.361 | 1.528 | 0.069 |
| Continued SFT + memory | 1: normal | 79.6% | 81.5% | 0.185 | 1.093 | 0.019 |
| Continued SFT + memory | 2: deadline | 77.8% | 80.1% | 0.199 | 0.931 | 0.028 |
| Continued SFT + memory | 3: travel | 84.3% | 85.2% | 0.148 | 0.829 | 0.009 |
| Corrected SDPO + memory | 1: normal | 67.6% | 73.1% | 0.296 | 1.352 | 0.088 |
| Corrected SDPO + memory | 2: deadline | 65.3% | 72.2% | 0.329 | 1.542 | 0.088 |
| Corrected SDPO + memory | 3: travel | 65.7% | 75.5% | 0.301 | 1.602 | 0.204 |
Paired SDPO minus SFT differences
| Stage | Complete schedules, percentage points [95% CI] | Hard violations/day [95% CI] | Correction flags/day [95% CI] |
|---|---|---|---|
| 1 | -12.037 [-20.370, -3.704] | +0.111 [+0.019, +0.204] | +0.259 [+0.125, +0.384] |
| 2 | -12.500 [-21.308, -3.241] | +0.130 [+0.032, +0.231] | +0.611 [+0.449, +0.778] |
| 3 | -18.519 [-25.926, -10.648] | +0.153 [+0.060, +0.245] | +0.773 [+0.648, +0.898] |
Positive completion differences favor SDPO; negative violation/correction differences favor SDPO. Percentile intervals bootstrap the same day IDs jointly across all seed pairs, averaging seeds within each day. They are conditional on these three optimizer seeds and their shared v1 SFT warm start: not 216 independent test days, three independent pretrained models, or uncertainty over every possible training run.
Seed-specific raw checkpoint gates
| Arm | Seed | Stage | Versus phase-matched frozen | Versus preceding stage, all weeks | Earlier-week retention |
|---|---|---|---|---|---|
| sft | 17 | 1 | PASS | n/a | n/a |
| sft | 17 | 2 | FAIL | FAIL | FAIL |
| sft | 17 | 3 | PASS | PASS | PASS |
| sft | 29 | 1 | PASS | n/a | n/a |
| sft | 29 | 2 | PASS | PASS | PASS |
| sft | 29 | 3 | PASS | FAIL | FAIL |
| sft | 43 | 1 | PASS | n/a | n/a |
| sft | 43 | 2 | PASS | FAIL | FAIL |
| sft | 43 | 3 | PASS | FAIL | FAIL |
| sdpo | 17 | 1 | PASS | n/a | n/a |
| sdpo | 17 | 2 | FAIL | FAIL | FAIL |
| sdpo | 17 | 3 | FAIL | PASS | PASS |
| sdpo | 29 | 1 | FAIL | n/a | n/a |
| sdpo | 29 | 2 | PASS | FAIL | FAIL |
| sdpo | 29 | 3 | FAIL | FAIL | FAIL |
| sdpo | 43 | 1 | FAIL | n/a | n/a |
| sdpo | 43 | 2 | FAIL | FAIL | PASS |
| sdpo | 43 | 3 | FAIL | FAIL | FAIL |
Gates require no measured worsening in hard violations, final conflict-free rate, or complete schedules overall and in each compared week. Correction counts cannot override failures. These are raw checkpoint checks, not a test-selected checkpoint or automatic deployment.
Week-specific means
| Arm | Stage | Test week | Complete schedules | Hard violations/day | Correction flags/day |
|---|---|---|---|---|---|
| frozen_sft | 1 | normal | 62.5% | 0.333 | 1.458 |
| frozen_sft | 1 | deadline | 62.5% | 0.417 | 1.458 |
| frozen_sft | 1 | travel | 66.7% | 0.375 | 1.750 |
| frozen_sft | 2 | normal | 62.5% | 0.333 | 1.458 |
| frozen_sft | 2 | deadline | 58.3% | 0.458 | 1.750 |
| frozen_sft | 2 | travel | 62.5% | 0.625 | 1.458 |
| frozen_sft | 3 | normal | 62.5% | 0.333 | 1.458 |
| frozen_sft | 3 | deadline | 58.3% | 0.458 | 1.750 |
| frozen_sft | 3 | travel | 75.0% | 0.292 | 1.375 |
| sft | 1 | normal | 77.8% | 0.208 | 0.861 |
| sft | 1 | deadline | 84.7% | 0.153 | 1.042 |
| sft | 1 | travel | 76.4% | 0.194 | 1.375 |
| sft | 2 | normal | 76.4% | 0.236 | 0.833 |
| sft | 2 | deadline | 83.3% | 0.167 | 0.889 |
| sft | 2 | travel | 73.6% | 0.194 | 1.069 |
| sft | 3 | normal | 83.3% | 0.167 | 0.833 |
| sft | 3 | deadline | 86.1% | 0.139 | 0.833 |
| sft | 3 | travel | 83.3% | 0.139 | 0.819 |
| sdpo | 1 | normal | 62.5% | 0.319 | 1.139 |
| sdpo | 1 | deadline | 73.6% | 0.222 | 1.264 |
| sdpo | 1 | travel | 66.7% | 0.347 | 1.653 |
| sdpo | 2 | normal | 66.7% | 0.292 | 1.347 |
| sdpo | 2 | deadline | 68.1% | 0.361 | 1.514 |
| sdpo | 2 | travel | 61.1% | 0.333 | 1.764 |
| sdpo | 3 | normal | 68.1% | 0.333 | 1.569 |
| sdpo | 3 | deadline | 72.2% | 0.236 | 1.597 |
| sdpo | 3 | travel | 56.9% | 0.333 | 1.639 |
Per-seed means
| Arm | Stage | Seed | Complete schedules | Hard violations/day | Correction flags/day |
|---|---|---|---|---|---|
| frozen_sft | 1 | single frozen evaluation | 63.9% | 0.375 | 1.556 |
| frozen_sft | 2 | single frozen evaluation | 61.1% | 0.472 | 1.556 |
| frozen_sft | 3 | single frozen evaluation | 65.3% | 0.361 | 1.528 |
| sft | 1 | 17 | 77.8% | 0.194 | 1.139 |
| sft | 1 | 29 | 77.8% | 0.208 | 1.014 |
| sft | 1 | 43 | 83.3% | 0.153 | 1.125 |
| sft | 2 | 17 | 66.7% | 0.306 | 1.125 |
| sft | 2 | 29 | 88.9% | 0.083 | 1.000 |
| sft | 2 | 43 | 77.8% | 0.208 | 0.667 |
| sft | 3 | 17 | 79.2% | 0.208 | 0.819 |
| sft | 3 | 29 | 84.7% | 0.139 | 0.847 |
| sft | 3 | 43 | 88.9% | 0.097 | 0.819 |
| sdpo | 1 | 17 | 80.6% | 0.181 | 1.222 |
| sdpo | 1 | 29 | 66.7% | 0.319 | 1.181 |
| sdpo | 1 | 43 | 55.6% | 0.389 | 1.653 |
| sdpo | 2 | 17 | 66.7% | 0.361 | 1.250 |
| sdpo | 2 | 29 | 69.4% | 0.319 | 1.653 |
| sdpo | 2 | 43 | 59.7% | 0.306 | 1.722 |
| sdpo | 3 | 17 | 75.0% | 0.167 | 1.097 |
| sdpo | 3 | 29 | 69.4% | 0.292 | 1.319 |
| sdpo | 3 | 43 | 52.8% | 0.444 | 2.389 |
Runtime, exposure, and compute
Development-calibrated approximate training wall time, not equal optimizer steps, examples, tokens, or FLOPs.
| Phase of experiment | Arm | Completed training segments | Optimizer updates | Training seconds excluding callbacks |
|---|---|---|---|---|
| develop | sdpo | 4 | 768 | 723.306 |
| develop | sft | 2 | 128 | 69.881 |
| final | sdpo | 9 | 576 | 542.453 |
| final | sft | 9 | 990 | 533.069 |
Exposure is not equal just because updates match: SFT nominally presents four examples/update; SDPO draws one prompt with four sampled one-token actions/update. Dataset composition, native observed generation counts, distinct prompt texts, nominal exposures, optimizer continuity, reported token metrics, and individual timings are retained in report.json when present in the source artifacts. Repeated prompt draws are not distinct examples.
Recorded Job running time: 4691.000 seconds across 4/4 timed Jobs. Estimated compute: $1.3031, covering 4/4 recorded Jobs.
Only recorded v2 Job compute, including recorded failed attempts. Estimates are not invoices and exclude hosting, prior pilot runs, and other services.
Boundaries and provenance
- Synthetic public-rubric feedback and four offered actions, not measured human effort or real-user efficacy.
- Correction flags are the original cost-minimization metric, not actual corrections automatically applied to trajectories.
- Complete schedule means final_feasible and zero deferred tasks; original metrics remain available unchanged.
- The frozen SFT comparator retains the shared warm-start adapter and receives phase-limited memory; it is evaluated once, not triplicated as an optimizer replicate.
- Matched teacher-prompt diagnostics use identical v2 feedback in both placements, not an exact v1-wording replay.
- Paired intervals resample shared day IDs jointly across three observed optimizer seeds; 216 seed/day evaluations are not 216 independent days.
- All optimizer seeds share one v1 SFT warm-start checkpoint and one pretrained base; uncertainty is conditional on that shared warm start, not three independent pretrained-model replicates.
- Passing descriptive checkpoint gates is not a statistical guarantee, monotonic learning claim, or live-demo promotion.
- The fixed collector supplies identical historical student-state data; this is not continually recollected on-policy data for each changing arm.
{
"data_revision": "8b8f3763cdc52766e84791c28985539ff3db4f69",
"ledger_hash": "bd33e30f144155374d5af22484cd3d7cdaa575a32f037b51c104547361506ec0",
"manifest_hash": "45c8c60816f6f02f064ec905ec6240c9d6631ef49182a98c04558bbd5953b8c3",
"source_revision": "fa257386d857130bda7188c6df7e1fdeff20efc8",
"development_revision": "3e003eb15fb6a407cf63bdeb968f6b55b96c6ebe"
}