Every number in this document was produced in June 2026 by a script committed under
experiments/, on the data committed (or fetched by a committed script) under
dataset_examples/. The methodology, evaluation protocol, and literature context are
in METHODOLOGY.md.
Versions ≤ 0.3.x produced predictions that were independent of the input data (the time-index embedding replaced the series; full account in METHODOLOGY.md §8). The before/after evidence:
Guarded by tests/test_learning.py: the fixed model's predictions respond to their
inputs, training loss falls by >2x, and held-out MAE beats seasonal-naive
(measured ratio 0.59 on the synthetic trend+seasonality regression benchmark).
python experiments/benchmark_multidomain.py — 6 datasets, 12-step horizon,
30 epochs, MAE relative to seasonal-naive (<1.00 beats it). Measured 2026-06-11:
| Dataset | APDTFlow | seasonal-naive | naive-last | Linear | Holt-Winters |
|---|---|---|---|---|---|
| Daily min temperature (real) | 0.73 | 1.00 | 0.92 | 0.74 | 0.80 |
| Regime-switching nonlinear | 0.77 | 1.00 | 1.44 | 0.83 | 0.86 |
| Trend + dual seasonality | 0.85 | 1.00 | 2.57 | 0.50 | 0.38 |
| Retail-like mult. seasonal | 1.01 | 1.00 | 7.77 | 0.68 | 0.81 |
| Electric production (real, 397 pts) | 1.52 | 1.00 | 3.68 | 1.03 | 1.23 |
| Random walk (stochastic) | 1.86 | 1.00 | 1.00 | 1.15 | 1.12 |
Reading: APDTFlow beats seasonal-naive on 3 of 6 domains (parity on a 4th) and beats all baselines on the two with nonlinear structure (temperature, regime-switching). Simple baselines win where the structure is linear or short (trend+dual-seasonality, the 397-point electric series), and nothing beats naive on a random walk — which is exactly why it is in the table.
All audits use the adversarial protocol of experiments/audit_predict_when.py:
held-out units (unseen batteries / engines), thresholds and normalization defined on
training units only, baselines = persistence and linear extrapolation, asymmetric
time-space conformal calibration (METHODOLOGY.md §4).
python experiments/battery_eol_demo.py — leave-one-battery-out, horizon 30
measured cycles, threshold 1.4 Ah, direction below.
Measured 2026-06-12 (cells modeled in state-of-health terms — capacity relative to initial capacity, the standard battery-RUL normalization):
| Held-out cell | Events | APDTFlow | Linear | Persistence | Catch | Coverage (90% target) |
|---|---|---|---|---|---|---|
| B0005 | 30 | 2.76 | 3.47 | 15.23 | 83% | 96% |
| B0006 | 31 | 13.71 | 15.65 | 15.47 | 6% | — |
| Pooled | 61 | 8.33 | 9.66 | 15.36 | 44% | 96% |
Timing errors in measured cycles (multiply by ~4 for chronological cycles — the cells are measured every ~4 cycles). B0007 never reaches EOL inside the horizon; the model correctly censored 42 of its 109 windows (61% false-alarm rate on that just-above-threshold cell — reported, not hidden).
Honest reading: pooled timing beats both baselines, and on the typical cell (B0005) the model is decisively better-calibrated than anything else we ran. But cross-cell transfer to the atypically fast-fading B0006 is weak (6% catch): with only two training cells, the model does not extrapolate to a degradation rate it has never seen. Fleet-scale battery datasets (Stanford/MIT-Toyota 124-cell) are the roadmap fix (CONTRIBUTING.md).
python experiments/turbofan_when_demo.py — sensor s11 indicator, threshold from
training engines only, audit on unseen engines, horizon 40 cycles.
Measured 2026-06-12 — 40 unseen engines, 1,990 held-out windows, 271 real events, 25 epochs:
| Metric | APDTFlow | Linear | Persistence |
|---|---|---|---|
| Timing MAE, full event set (cycles) | 8.33 | 8.65 | 11.46 |
| Timing MAE, matched subset (n=38) | 8.49 | 9.68 | — |
| False alarms (1,719 no-crossing windows) | 0.64% | — | — |
| Catch rate | 26.6% | — | — |
| 90%-window coverage | 40.3% | — | — |
Honest reading: APDTFlow beats both baselines on timing — including on the matched subset where linear also produces an estimate — and its false-alarm rate is near zero, the property that matters most against alarm fatigue. The model is deliberately conservative (27% catch), and the calibrated windows under-cover on this audit (40% vs the 90% target): cross-engine transfer stretches the calibration beyond what the training engines support. Both numbers are printed by the script.
python experiments/turbofan_multivariate_demo.py — same audit, with
fit(feature_cols=[s12, s4, s7, s15]) fusing five sensors into a learned health
indicator.
Measured 2026-06-12, same audit and engines as §3.2:
| Metric | Univariate | Multivariate (5-sensor fusion) |
|---|---|---|
| Timing MAE, full event set (cycles) | 8.33 | 10.11 |
| Timing MAE, caught events (cycles) | 8.34 | 5.85 |
| Event catch rate | 26.6% | 7.0% |
| False alarms | 0.64% | 0.00% |
| 90%-window coverage | 40.3% | 26.3% |
Honest reading: the multivariate fusion did not reproduce the across-the-board
improvement we hoped for. It makes the model sharper where it speaks (caught-event
error drops by 30%, false alarms reach zero) at the cost of much higher
conservatism, and full-event error does not improve. The learned fusion weights
remain a useful interpretability artifact (sensor_importance_). Both sides of the
trade-off are printed by the script and shown in the figure.
python experiments/fd002_robustness_demo.py — 6 operating regimes,
regime_normalize statistics from training engines only, multivariate indicator,
audit on unseen engines.
Measured 2026-06-12 — 110 unseen engines, 3,028 held-out windows, 389 real events, 25 epochs, per-regime normalization statistics from training engines only:
| Metric | APDTFlow (multivariate) | Linear | Persistence |
|---|---|---|---|
| Timing MAE, full event set (cycles) | 9.20 | 8.10 | 11.26 |
| Timing MAE, matched subset (n=46) | 6.81 | 7.81 | — |
| Timing MAE, caught events (cycles) | 6.55 | — | — |
| False alarms (2,638 no-crossing windows) | 0.00% | — | — |
| Catch rate | 13.4% | — | — |
| 90%-window coverage | 53.9% | — | — |
Fleet snapshot (one mid-life window per engine): calibrated windows covered 81% of actual crossings; the act-by edge preceded the actual crossing for 23% of crossing engines.
Honest reading: linear extrapolation wins this audit on the full event set, and we publish that. APDTFlow's measured advantages on FD002 are the matched subset (6.81 vs 7.81 — sharper where both methods commit), zero false alarms, and beating persistence by 2.1 cycles. The under-target coverage (54% vs 90%) repeats the cross-unit-transfer pattern of the other audits. The shipping rule (Section 5) means FD002 is not advertised as a win anywhere in this repository — it is advertised as exactly what it measured.
python experiments/benchmark_missing_data.py — held-out MAE vs the TRUE series,
40 epochs. Measured 2026-06-11:
| Method | 30% missing | 50% missing |
|---|---|---|
| APDTFlow, plain ffill imputation | 1.77 | 2.81 |
| APDTFlow, mask + time-since-observation features | 2.05 | 3.31 |
| linear (ffill) | 1.63 | 2.63 |
| seasonal-naive (ffill) | 1.68 | 2.79 |
Mask/delta-t features made accuracy worse at both missingness levels — confirmed negative result; the feature is not shipped.
python experiments/prototypes/odernn_gate.py — 53% bursty missingness, 40 epochs.
Measured 2026-06-11:
| Method | held-out MAE vs true series |
|---|---|
| seasonal-naive + ffill | 2.03 |
| APDTFlow + ffill | 2.48 |
| linear + ffill | 2.51 |
| ODE-RNN encoder | 2.87 |
Consistent with arXiv:2505.00590. Research track only — APDTFlow makes no irregular-sampling claims.
The sunspots series is kept as a calibration fixture (its predict_when time-window
coverage is unit-tested at ≥85%), not as a skill demo. The full audit
(audit_predict_when.py, threshold 80, horizon 18, 168 held-out windows, measured
2026-06-12) confirms the rejection: seasonal-naive catches 62% of crossings at
4.7-month timing error vs the model's 31% at 5.2 — the shipping-rule verdict is
FAIL, and we publish that verdict instead of the demo.
A predict_when domain demo is publishable only if it beats all of
(a) persistence, (b) linear extrapolation, (c) seasonal-naive/climatology where the
series is seasonal — on both event capture and timing MAE, on held-out data, via
experiments/audit_predict_when.py. See CONTRIBUTING.md.