Part of QSMPC-QKD-QHE-AI-Hybrid, a quantum-safe orchestration demo. This is the plaintext model for the satellite_comms use case; the encrypted path runs a distilled student, not this model.

Measured performance

metric	value
`anomaly_sample_rate`	0.0075528
`event_wise_f0_5`	0.454545
`event_wise_fn`	2
`event_wise_fp`	1
`event_wise_tp`	1
`n_anomaly_intervals`	320
`n_events_scored`	3
`n_parameters`	80486
`point_adjusted_f1`	0.993691
`published_full_channel_f0_5`	0.061
`published_lightweight_f0_5`	0.786
`wall_clock_s`	43.2
`windowed_iforest_event_wise_f0_5`	0.000248

Published baselines this is measured against

Target metric: event-wise F0.5
Baseline to beat: 0.968 - ESA-ADB Supplementary Table 9, Mission1 lightweight (ch. 41-46), Telemanom-ESA-Pruned: corrected event-wise F0.5 0.968 (precision 0.999, recall 0.862) for detection of ANOMALIES ALONE - which is what this project scores, so this is the like-for-like row. Earlier rounds cited Table 2's 0.786 instead. That number is real and correctly attributed, but Table 2 scores ALL EVENTS excluding communication gaps, i.e. anomalies PLUS annotated rare nominal events, so it was the wrong comparator and the lenient one. Verified against the paper (arXiv:2406.17826 v2) in Round 24. THREE protocol differences remain and none of them are closed by changing the number: (1) ESA-ADB's metric multiplies event precision by the TIME-LEVEL true-negative rate, which is why most of its baselines read <0.001 - this project records that corrected value as corrected_event_wise_f0_5 beside the uncorrected one; (2) ESA-ADB splits each mission in half, this project tests on the last 30 percent; (3) ESA-ADB's Mission1 lightweight test partition holds 65 annotated events of which 29 are anomalies, against 3 in the window scored here. So this remains an indicative target, and the sample-size gap is roughly tenfold.
Companion metric shown alongside: event_wise_f0_5_full_channels - reported together because the aggregate figure can look healthy while the class that matters is not.

Gap to the published baseline

Read n_events_scored before comparing this to the baseline. The scored test window holds THREE labelled anomaly events, not the 320 that n_anomaly_intervals counts over the whole timeline - the annotated anomalies sit earlier in the mission, so a temporal split leaves almost nothing to score. Both recorded values decompose uniquely on three events: 0.714286 is tp=1 fp=0 fn=2 and 0.454545 is tp=1 fp=1 fn=2, so the 0.26 swing between two identical-config runs is ONE extra false-alarm run. The detector flags 2 points out of 441,600. Three protocol gaps to the citation remain open and are stated rather than closed: ESA-ADB multiplies event precision by the time-level TNR (recorded here as corrected_event_wise_f0_5), it splits each mission in half where this tests on the last 30 percent, and its lightweight test partition holds 65 events against 3 here. Workings: references/satellite-event-count-investigation.md.

Training data

Dataset: ESA Anomaly Dataset (ESA-ADB), Mission1 lightweight channels 41-46
Licence: CC-BY-3.0-IGO
Source: https://zenodo.org/records/12528696 (licence read 2026-08-03)
Attribution (CC-BY-3.0-IGO): Kotowski et al., 'European Space Agency Benchmark for Anomaly Detection in Satellite Telemetry', arXiv:2406.17826. Data: ESA Anomaly Dataset, Zenodo record 12528696, CC BY 3.0 IGO.

Missions are anonymised Mission1/2/3 — they are not named Sentinel anything. Multi-year telemetry; the exact span is NOT stated by the Zenodo record, the paper or the upstream README, so it is withheld here rather than asserted - see DATA_LICENSES.md. 1.19% annotated. Mission3 is excluded upstream for triviality. The CC BY 3.0 IGO term appears on the Zenodo landing page but NOT in its REST metadata, which returns no licence at all - an API-only check reads this corpus as unlicensed.

Notes and limitations

Report event-wise F0.5, never point-adjusted F1. Always run a trivial baseline alongside: windowed isolation forest wins the ESA-ADB Mission2 lightweight subset outright. RETRACTION, Round 19: a Round 18 note here claimed this forecaster was about 22x WORSE than persistence. That was a measurement error of ours, not a property of the model -- the probe resampled at 60s while the model is trained at 300s (see resample in the achieved record). Re-measured at the trained cadence, next-step MAE is 0.0314 on the evaluation split against 0.0897 for persistence, i.e. the model is about 2.9x BETTER; 0.40x / 0.35x / 0.38x on train / test / whole series. Preprocessing for any evaluation must be read from this record rather than hardcoded, which is now what scripts/prep_model_inputs.py does. SEPARATE AND STILL OPEN: event-wise F0.5 does not reproduce across identical-config retrains -- 0.4545 here against the 0.714286 recorded -- while point-adjusted F1 does (0.9937 vs 0.9945) and the isolation-forest baseline reproduces exactly (0.000248). Treat the event-wise figure as a single draw from a wide distribution, which is the instability arXiv:2607.11969 (2026-07) describes for overlap-based metrics.

Honest scope

This model is published as part of a research proof of concept, not as a production system. Numbers above are what this repository measured on the split described, with the code in scripts/train/. Where a figure is carried from the literature rather than measured here, it is labelled as such.

Daemons-Q/qsh-satellite-telemanom-esa

Telemanom LSTM forecaster (per channel)