The canonical dataset

One row per minute of a SoLEXS observing day. Source: .lc (light curve) + .gti, SDD2.

Column	Type	Units	Source / derivation	Allowed values	Missing semantics
timestamp	datetime64[ns,UTC]	—	minute index of the observation day	1,440/day, monotonic, unique	never missing
counts_total	float64	counts	Σ of .lc COUNTS over finite seconds in the minute	≥ 0	NaN ⟺ no finite second (q_no_data)
live_time_s	float64	s	GTI-covered seconds in the minute (inclusive endpoints, STOP−START+1)	0 … 60	never missing; 0 = fully excluded
rate_total	float64	counts/s	counts_total / live_time_s	≥ 0	NaN where live_time_s == 0 — not imputed
gti_fraction	float64	—	live_time_s / 60	0.0 … 1.0	never missing
n_seconds_present	int16	count	number of finite .lc seconds in the minute	0 … 60	never missing
q_no_data	bool	—	n_seconds_present == 0	True/False	flag
q_partial	bool	—	0 < gti_fraction < 1	True/False	flag
detector	category	—	SoLEXS detector	SDD2 only	never missing
src_file	object	—	source .lc path	—	never missing
src_sha256	object	—	SHA-256 of the source archive (Phase 0.5.1 manifest)	64 hex	never missing
archive_version	object	—	ISSDC archive variant	v1.0, v1.1	never missing

counts_total (light curve) is band-limited and is not Σ(T2 counts) — see SCIENTIFIC_FINDINGS.md F-1. The two are carried independently; neither is derived from the other.

One row per minute; the full SoLEXS spectrum. Source: .pi (OGIP Type II PHA) + .gti, SDD2.

Column	Type	Units	Source / derivation	Allowed values	Missing semantics
timestamp	datetime64[ns,UTC]	—	minute index	1,440/day	never missing
counts	object → array[340] float64	counts	per-channel Σ over finite seconds in the minute	≥ 0 per channel	NaN per channel where no finite second
live_time_s	float64	s	GTI-covered seconds (inclusive)	0 … 60	never missing
gti_fraction	float64	—	live_time_s / 60	0.0 … 1.0	never missing
n_seconds_present	int16	count	finite spectra contributing	0 … 60	never missing
q_no_data	bool	—	no finite spectrum in the minute	True/False	flag
q_partial	bool	—	partial GTI coverage	True/False	flag
detector	category	—	SoLEXS detector	SDD2	never missing
chantype	category	—	channel type from CHANTYPE	PI	never missing
src_file, src_sha256, archive_version	object	—	provenance	—	never missing

Channel axis: counts[i] is PI channel i, i ∈ [0, 339] — an ordinal index. No channel_energy_keV column exists, by design: the archive contains no response matrix, so no channel→energy mapping is possible. Do not assume linearity in energy.

LONG form (spec r6): one row per (minute, detector), keyed uniquely by that pair. Source: HEL1OS lightcurve_*.fits, ownership from the Version Resolution Engine.

Column	Type	Units	Source / derivation	Allowed values	Missing semantics
timestamp	datetime64[ns,UTC]	—	minute index	—	never missing
detector	object	—	the detector this row describes	CZT1,CZT2,CDTE1,CDTE2	never missing
{det}_{lo}_{hi}_rate	float64	counts/s	mean of CTR over the minute's samples	≥ 0	NaN when {det} ≠ this row's detector (structural)
{det}_{lo}_{hi}_stat_err	float64	counts/s	mean of STAT_ERR	≥ 0	as above
{det}_{lo}_{hi}_n_samples	float64	count	samples contributing to the minute	≥ 1	as above
orbit_id	object	—	owning orbit (version-resolved)	HLS_…_V###	never missing
orbit_version	object	—	orbit product version	V111,V211,V112,V311	never missing
src_file, src_sha256, archive_version	object	—	provenance	—	never missing

Band columns present (edges parsed from EXTNAME, allowlist-validated):

• CZT1/CZT2: 20_40, 40_60, 60_80, 80_150, 18_160 (keV)
• CDTE1/CDTE2: 5_20, 20_30, 30_40, 40_60, 1p8_90 (keV; 1p8 = 1.8)

Structural NaN: because the table is long-form, each row populates only its own detector's band columns; the other three detectors' columns are NaN by construction, not by missing data. Filter by detector before using band columns. A wide view may be derived, but the canonical form is long (spec r6) because ownership is per (minute, detector).

One row per minute of HEL1OS housekeeping. Source: aux/hk.fits.

Column	Type	Units	Derivation	Notes
timestamp	datetime64[ns,UTC]	—	minute index	never missing
czt1temp, czt2temp, cdte1temp, cdte2temp	float64	°C	minute mean	detector temperatures
czthvmon, cdtehvmon	float64	V	minute mean	high-voltage monitors
czt1ctr, czt2ctr, cdte1ctr, cdte2ctr	float64	counts/s	minute mean	per-detector rate monitors
czt1enth	float64	keV	minute mean	energy threshold
cdte1enerthr, cdte2enerthr	float64	(undeclared)	minute mean	energy thresholds; unit not declared in the archive
cdte1pilectr, cdte2pilectr	float64	count	minute max	pile-up counters
czt1satctr1, czt2satctr1	float64	count	minute max	saturation counters
czt1hotpixcnt, czt2hotpixcnt	float64	count	minute max	hot-pixel counters
sunradeg, sundecdeg	float64	degrees	minute mean	Sun pointing
fehkstat	float64	—	minute mean	front-end housekeeping status word
suninfov	bool	—	minute min (conservative: True only if every sample is True)	Sun-in-field-of-view
n_samples	int64	count	HK records in the minute	≥ 1
orbit_id, src_file, src_sha256, archive_version	object	—	provenance	never missing

Ordering: the parser preserves archive (telemetry-arrival) order and never sorts. Source records contain sub-second out-of-order steps (SCIENTIFIC_FINDINGS.md F-3). All aggregations used here are order-independent. A consumer needing chronological order must call app.v2.utils.timeseries.chronological_sort() explicitly.

One row per (minute, detector); HEL1OS spectra. Source: hel1os_*_spectra_*.fits.

Column	Type	Units	Derivation	Allowed values	Missing semantics
timestamp	datetime64[ns,UTC]	—	minute index	—	never missing
detector	object	—	spectra detector	CZT1,CZT2,CDTE1,CDTE2	never missing
counts	object → array[detchans] float64	counts	per-channel Σ over finite spectra in the minute	≥ 0	NaN per channel where absent
stat_err	object → array[detchans] float64	counts	per-channel Σ of STAT_ERR	≥ 0	as above
live_time_s	float64	s	Σ EXPOSURE of contributing spectra	≥ 0	never missing
n_spectra	int64	count	spectra contributing to the minute	≥ 1	never missing
chantype	category	—	from CHANTYPE	PHA	never missing
detchans	int16	count	channel-space size — carried explicitly (r5)	341 (CZT) or 511 (CdTe)	never missing
orbit_id, src_file, src_sha256, archive_version	object	—	provenance	—	never missing

Three incommensurable channel spaces (F-4): SoLEXS PI 340, HEL1OS CZT PHA 341, HEL1OS CdTe PHA 511. Never concatenate counts arrays across families or instruments — always filter by detchans/detector first. Channel indices are ordinal; no response matrix exists.

Long-form good-time intervals, one row per interval.

Column	Type	Units	Derivation	Notes
instrument	object	—	source instrument	solexs
detector	object	—	detector	SDD2 (SDD1 contributes no intervals — F-12 inactive)
start_utc	datetime64[ns,UTC]	—	interval start	inclusive
stop_utc	datetime64[ns,UTC]	—	interval stop	inclusive
duration_s	float64	s	stop − start + 1 (inclusive convention)	> 0 always
src_file, src_sha256	object	—	provenance	never missing

A GTI-excluded second may carry a finite count (spec r5). The invariant is one-directional: NaN(counts) ⇒ GTI-excluded. The excess is unexplained (A-14 / F-2).

One row per parsed source file. Every canonical row traces here via src_file.

Column	Type	Derivation	Notes
src_file	object	source path	join key; unique with (product,detector)
src_sha256	object	archive SHA-256 (Phase 0.5.1 manifest)	64 hex
instrument	object	solexs / hel1os	—
detector	object	e.g. SDD1,SDD2,CZT1,CDTE2	None for whole-orbit products
product	object	lc,pi,gti,hk,spectra,lightcurve,events	—
archive_version	object	v1.0,v1.1,V111,…	—
obs_date	object	YYYYMMDD	—
orbit_id	object	HEL1OS orbit, else None	—
creator	object	producing pipeline, from the FITS header	e.g. solexs_pipeline-1.4
processing_date	object	header DATE	None for HEL1OS (no DATE header)
rows_in / rows_out	int64	source rows / emitted rows	0/0 legal for F-12 inactive detectors
parser_version	object	parser identity	v2-0.5.2
parsed_at_utc	object	build wall-clock time	the one non-reproducible field — excluded from content hashes
time_epoch_resolution	object	resolved time convention	unix_seconds (SoLEXS) / relative_seconds (HEL1OS spectra) / mjd_days
assumptions_applied	object → array[str]	assumptions recorded at parse	audit trail

---

1. Every canonical row carries src_file + src_sha256; no row lacks provenance (verified: 0 orphans, 0 duplicates, 0 missing).
2. src_sha256 is the archive SHA-256 from the Phase 0.5.1 manifest — it pins the input, not the output.
3. HEL1OS tables additionally carry orbit_id, resolved by the Version Resolution Engine per (minute, detector). Two detectors in the same minute may legitimately have different orbit_ids at orbit boundaries.
4. Provenance columns are not measurements; they must not be used as features.

Anomaly class	Rule	Count	Products affected	Disposition
GTI interval STOP ≤ START	F-19	12	SoLEXS daily archives	Archive defect (CONTRADICTION-005 Defect C) — excluded, logged
Duplicate HK mjd timestamp	F-16	2	HEL1OS orbits	Archive defect (CONTRADICTION-006 Defect B) — excluded, logged
Inactive detector (empty GTI)	F-12	426	SoLEXS SDD1 GTI files	Legal state, not an error — SDD1 is GTI-only, SDD2 carries science
HK time inversions (jitter)	—	see A-12	HEL1OS orbits	Characterised in MILESTONE_VIII_VALIDATION_REPORT.md (A-12)
Two PHA channel families	F-07→resolved	0 (resolved)	HEL1OS CZT/CdTe	Contract corrected (CONTRADICTION-005 Defect A, r5) — 0 residual failures
AppleDouble ._* sidecars	F-18	(excluded)	filesystem artifacts	exFAT/macOS artifacts, never data — excluded by allowlist

Total products excluded from the canonical dataset: 14 (12 SoLEXS + 2 HEL1OS), of 436 + 391 = 827 available. Coverage: SoLEXS 424/436, HEL1OS 389/391.

A GTI interval whose stop time is at or before its start time. STOP ≤ START is not a convention question; it is an internal inconsistency in the archived GTI table. F-19 terminates on parse.

Affected (all SoLEXS daily archives): AL1_SLX_L1_20240422_v1.0, 20240501, 20250729, 20251205, 20260406, 20260407, 20260408, 20260409, 20260411, 20260413, 20260414, 20260415.

Disposition: archive defect. Excluded from the canonical dataset, individually logged. No amendment (F-19 behaves as designed). Note the temporal clustering (8 of 12 in April 2026) — recorded as an observation, not interpreted.

Two housekeeping rows sharing an identical mjd timestamp. §2.8 (r4) requires HK mjd to be unique; a duplicate terminates via F-16.

Affected (HEL1OS orbits):

• HLS_20260201_120005_43198sec_lev1_V111 — 2 duplicated values (of 63,298 rows)
• HLS_20260202_000005_43183sec_lev1_V111 — 1 duplicated value

Disposition: archive defect (CONTRADICTION-006 Defect B, owner ruled not approved for amendment — F-16 continues to terminate). Excluded, logged. 389 of 391 orbits have fully unique HK mjd.

A GTI table with zero rows (NAXIS2 = 0), which the contract treats as a legal inactive-detector state, not an error (the single deliberately non-terminating rule). OBSERVED: SoLEXS SDD1 supplies GTI only (no .lc/.pi), and its GTI is empty across the archive; SDD2 carries all science.

Count: 426 F-12 detections among 850 parsed SoLEXS GTI files (SDD1 across the built days). Disposition: expected structural property of the SoLEXS archive; not a defect.

OBSERVED during the archive-wide build: HEL1OS has two detector families with different PHA channel counts — CZT = 341, CdTe = 511. The r0 contract asserted a single DETCHANS = 341, which blocked all 391 orbits at F-07.

Disposition: contract corrected (CONTRADICTION-005 Defect A → spec r5: per-family allowlist; F-11 restated over three incommensurable channel spaces — SoLEXS PI 340, HEL1OS CZT PHA 341, HEL1OS CdTe PHA 511). After the fix: 0 F-07 DETCHANS failures archive-wide. Not an archive defect — an archive property the contract now encodes.

macOS writes ._<name> metadata sidecars on the exFAT volume. These are filesystem artifacts, never instrument data. F-18 (allowlist + ._* guard) excludes them. Disposition: not data; excluded structurally. (Note: the withdrawn "10 F-01 unreadable days" figure from CONTRADICTION-005 originated in an ad-hoc scan that globbed these sidecars; the authoritative build has no F-01 skips.)

From Phase 0.5.1: exactly one genuinely lost SoLEXS date (2023-12-13, an interrupted download in downloads/corrupted/), and the HEL1OS archive begins only 2025-12-07 (combined SoLEXS+HEL1OS coverage = 179 days). These are acquisition gaps, not quality defects of the data on hand; carried to Phase 0.5.3.

Product(s)	Rule	Permanent disposition
12 SoLEXS F-19 archives (§2)	F-19	Archive defect — permanently excluded
2 HEL1OS F-16 orbits (§3)	F-16	Archive defect — permanently excluded
426 SDD1 empty GTIs (§4)	F-12	Legal inactive — not excluded, not a defect
CZT/CdTe channel counts (§5)	F-07→r5	Archive property — contract corrected, 0 residual
._* sidecars (§6)	F-18	Not data — structurally excluded
2023-12-13, pre-2025-12 HEL1OS (§7)	—	Acquisition gap — Phase 0.5.3

These are the permanent archive findings of Phase 0.5.

Measurement. 564,160 usable minutes contain 581 independent M/X events (4,065 ≥C; 47 X-class; 179 M/X on combined days).

Constrains model choice. Model capacity must be sized to ~581 events, not to half a million rows. A model with thousands of effective parameters will fit event-specific noise. Recommendation: ≤ ~15 features and a low-capacity learner for the first benchmark; deep sequence models only if Stage 3 evidence justifies them.

Constrains evaluation. Minute-level metrics overstate confidence by roughly 564,160/581 ≈ 970×. All metrics must be event-level, with confidence intervals from a block bootstrap over events. Reporting minute-level AUC as the headline would be a statistical error, not a stylistic choice.

Constrains claims. Any subgroup smaller than ~50 events cannot support a quantitative claim. This rules out standalone X-class conclusions (47 events; ~15–20 in a test split, where one event moves recall by ~5 points).

Measurement. Prediction AUC on quiet minutes only (in-progress flares excluded): 30 min 0.8119, 60 min 0.8023, 120 min 0.7951, 360 min 0.7884. A 12× horizon change costs 0.024 AUC.

Constrains claims — severely. Genuine precursor information must decay with horizon. Near-flat skill across 30 min to 6 h is the signature of a slowly varying activity state that persists for days. A raw forecasting AUC of ~0.80 must never be reported as flare-prediction skill. It must be reported as improvement over a persistence baseline, with the skill-vs-horizon curve published alongside.

Constrains validation. Persistence and climatology baselines are mandatory, not optional — without them the number is uninterpretable. This is v1's forecast-vs-persistence finding reproduced on real data, now with the mechanism identified.

Constrains loss functions. Optimising a horizon-agnostic loss will find the persistence solution, because that is where the gradient is. If precursor detection is the goal, the objective must explicitly reward short-horizon discrimination over long — otherwise the model will correctly learn "the Sun is active" and the result will be misread.

Measurement. No RMF/ARF anywhere in the archive. Channels are ordinal. The light curve is a band-limited integral of the spectrum with LC/Σ(PI) ratio varying 1.76–20.4 (F-1).

Constrains model choice. No regression target in physical units (W/m², GOES class as a continuous quantity) can be built or validated. Calibrated severity regression is excluded — this is a data limitation no architecture resolves.

Constrains claims. Permitted: "the model ranks events by observed intensity," "the model classifies against GOES-defined labels." Not permitted: "the model estimates GOES flux," "the model measures plasma temperature," or any statement in keV.

Constrains features. No physically-calibrated spectral feature (temperature, emission measure, spectral index) is derivable. Spectral features are ordinal-channel statistics only.

Measurement. SoLEXS-only: 424 days, 581 M/X events. Combined SoLEXS+HEL1OS: 171 days (2025-12-07 → 2026-06-15), 179 M/X events, 13 X-class.

Constrains validation. A chronological split within 171 days leaves a test period of ~2 months — too short for a stable estimate and containing no activity-phase diversity.

Constrains claims. Combined-instrument results cannot be generalised beyond a single ~6-month solar-maximum window. The ISRO brief's combined requirement is best served by an ablation on the shared window, not by a primary benchmark, since only a matched-window comparison attributes a difference to the instrument rather than the epoch.

Measurement (55,710 minutes, 40 days): band sums add ≤ 0.012 AUC over the total rate (hard 0.8911 vs total 0.8794). Hardness ratios: −0.030 for nowcast, and 0.5075 — indistinguishable from random — for 60-min prediction.

Constrains features. Elaborate spectral engineering is not justified in advance. Hardness-ratio features for prediction are directly refuted. The 340-channel spectrum must earn inclusion through a measured ablation.

Constrains claims. Until the Stage-4 ablation runs, no claim may be made that spectral resolution improves flare detection or prediction. The univariate evidence points the other way.

Scope of this limitation: it establishes that simple band-sums and ratios add little. It does not prove the full 340-dim spectrum is uninformative under a multivariate model — which is exactly why the ablation exists.

Measurement. Base rates: M/X nowcast 2.37 %, M/X ≤60 min 4.58 %, ≥C nowcast 12.25 %, ≥C ≤60 min 29.93 %.

Constrains loss functions. At 2.4 % positives, unweighted accuracy is meaningless and plain cross-entropy under-weights positives. Class weighting or focal loss is appropriate — but with 581 events, aggressive re-weighting increases variance on the rare class. Recommendation: mild class weighting, and select on a threshold-free ranking metric (AUC/AP) rather than on accuracy.

Constrains evaluation. Report the precision–recall frontier, not a single operating point. v1's most important operational lesson was that a fixed false-alarm budget can bury a working detector; the frontier makes the trade-off visible instead of pre-judged.

Measurement. T1 counts_total NaN on 4.20 % of minutes; rate_total non-finite on 7.6 %; every NaN flagged (q_no_data, q_partial). T3's 75 % NaN is structural (long form), not missing data.

Constrains model choice. Architectures requiring dense complete input need an explicit mask channel; imputation is prohibited by dataset policy and would fabricate observations. If an architecture cannot accept masks, prefer a different architecture rather than imputing.

Constrains evaluation. The dropped-window fraction is an experimental parameter and must be reported — dropping differs systematically between quiet and active periods, so silently dropping could bias the base rate.

Measurement. On 70 of 414 days, GTI excludes seconds that nevertheless carry finite counts — 266,919 s total, up to 43,199 s in one day, concentrated (top-10 days = 80.2 %). Mechanism unknown (A-14 / F-2).

Constrains features. live_time_s and gti_fraction are usable as observability covariates but their exact semantics are open. Do not build a physical rate normalisation that depends on GTI meaning more than "seconds the archive marks good."

Constrains claims. Exposure-normalised quantities carry an unquantified systematic on those 70 days. Any result sensitive to exposure normalisation should be checked with and without them.

Measurement. All labels come from the GOES catalogue, defined on GOES 1–8 Å flux. SoLEXS has no cross-calibration to that scale. v1 measured C-class detection at 0.244 recall versus 0.909 for M — i.e. many catalogued C flares are below SoLEXS's effective sensitivity.

Constrains claims. A "missed" flare may be a genuinely unobservable one, not a model failure. Recall against the GOES catalogue conflates model performance with instrument sensitivity. Report per-class recall with event counts always attached, and treat ≥C results as sensitivity-limited.

Constrains loss functions. Training on ≥C labels injects label noise concentrated in the positive class — the most damaging kind. Recommendation: train the primary benchmark on M/X labels (0.954 vs 0.800 univariate AUC quantifies the difference) and use ≥C as a robustness check.

Measurement. SoLEXS science is SDD2 only (SDD1 F-12 inactive across 426 GTI files). The span sits near solar maximum; M/X events per quarter range from 3 (2025Q2) to 181 (2024Q2).

Constrains validation. No detector-redundancy check is possible for SoLEXS. The 60× quarterly variation means a naive chronological split can produce train/test periods with wildly different base rates — the split must be chosen with event counts in view (hence the recommended 2024Q1–2025Q4 / 2026Q1–Q2 split, ~423 / ~155 events).

Constrains claims. No cross-cycle generalisation claim is supportable. Results apply to a solar-maximum regime observed by one detector.

---

Supportable with this dataset

• M/X flare nowcast/detection performance, event-level, with CIs — the signal is strong and verified (AUC 0.954 from one raw column).
• Improvement over persistence for short-horizon prediction, if demonstrated against the mandatory baselines.
• Ablation results: does spectral resolution help? does HEL1OS help? — as measured deltas with CIs.
• Operational characterisation: recall, false-alarm attribution, latency, the full frontier.

Not supportable

• Calibrated flux or severity regression (no RMF).
• Raw forecasting AUC presented as precursor skill (horizon-flat).
• Standalone X-class conclusions (47 events).
• Cross-cycle or multi-detector generalisation (one phase, one detector).
• Any claim in keV or physical spectral units (ordinal channels only).

The through-line: this dataset's strength is detection, and its strength there is genuine and large. Its forecasting signal is real but is an activity-state effect, and its severity axis is physically unavailable. A programme that leads with detection, controls forecasting against persistence, and treats spectra and HEL1OS as measured ablations will produce defensible results. One that leads with forecasting AUC or severity regression will produce claims this dataset cannot support.