🌞 Solar Activity Summary

The quietest week since the W32–W33 lull. Sep 7–13 produced no M- or X-class flares; the strongest event all week was a C5.9 from AR 4521 on Sep 12. (Verified against the NOAA edited_events record, which covered through Sep 13 at fetch time.)

The disk's four active regions of note all declined or stalled. AR 4524 — the M3.0/M1.2 producer of W36 — fell from 7.5% to the low single digits without further eruption. AR 4521 delivered the week's strongest C-class event but never exceeded 4.7% on the Transformer. AR 4528 and AR 4529 plateaued below 3% after brief early-week activity. Several new HARPs (13980–13989) emerged in the east but no numbered successor to AR 4513 has yet taken shape.

The one region that matters for next week is AR 4527, which spent the entire week pressed against the alert line without crossing it — see the performance section below.

📊 SWF System Performance

Performance is measured at each model's operating threshold (Transformer 7.9%, LightGBM 4.7%). Zero events makes POD undefined and FAR non-computable; the week's signal is entirely in the near-miss behaviour of the alert line.

For the second consecutive week the Transformer issued zero AR-day alerts — and this time the miss was measured in tenths of a percentage point:

• AR 4527 sat at 7.8% (Sep 11), 7.5% (Sep 12), and 7.6% (Sep 13) — a fraction of a point under the 7.9% operating line, three days running. This is the closest sustained near-miss in the bulletin record. The region never produced an M-class flare, so the threshold held — but it is the clearest illustration yet that a hard operating threshold makes a binary distinction out of a continuum. A region at 7.8% is not meaningfully "safer" than one at 8.0%.
• The signal was rising, not flat. AR 4527's Transformer probability climbed from 4.0% (early week) to 7.8% by Sep 11 and held there. Both models were reading a region that was building — LightGBM also rose from 2.1% to 3.1% over the same span, though from a much lower base.
• LightGBM remained far more conservative throughout (3.1% peak vs the Transformer's 7.8%), consistent with the calibration-gap pattern documented since W27: the two models agree on direction, disagree on magnitude at the top end.
• AR 4524's decay was correctly tracked. The W36 headline region fell from 7.5% to 2.3% (LightGBM Sep 7) and stayed low — and it produced nothing further. As in W35, the model's decay was faithful to the region's actual behaviour.

Two consecutive alert-free weeks (W36 with events, W37 without) mark the quietest sustained stretch in the bulletin's working record. The system is behaving correctly in the sense that no high-tier false alarms were raised — but the Transformer's total alert silence for two weeks running is worth flagging as a calibration observation: with the operating threshold at 7.9% and the disk's most active region hovering at 7.5–7.8%, the model has effectively been muted. Whether that is appropriate depends entirely on whether AR 4527 erupts in W38.

🔬 Notes

AR 4527 is the single region to watch into W38: it closed the week at 7.8% (Transformer), pressed against the alert line for three days without crossing, with LightGBM climbing alongside (3.1%). If it erupts, this week becomes a textbook near-miss validation; if it decays, the threshold's silence was the right call. AR 4521 produced the week's C5.9 and remains the disk's most magnetically active region by C-class output. New HARPs (13980–13989) are emerging but unnumbered. No X-class has occurred since July 4, and no region on the disk currently resembles the AR 4513 profile that produced W34–W35's M8.2/M7.1. The two-week alert silence is the system's calmest stretch since before W27 — a state that has historically preceded abrupt change.

🌞 Solar Activity Summary

A transitional week following AR 4513's decay. Aug 31–Sep 6 produced 3 M-class flares, 0 X-class, all in the M1–M3 range. AR 4524, a new region that emerged early in the week, contributed an M3.0 on Sep 2 and an M1.2 on Sep 4; AR 4520, a late-life region, produced an M1.0 on Sep 5. AR 4513, the M8.2/M7.1 producer of W34–W35, faded to 6.6% (Transformer) by Aug 31 and exited the alert stage entirely — its epic run is over.

The C-class background tells the real story of AR 4524: it ran hot all week below the M threshold — C7.9 on Sep 3, C7.2 on Sep 4, and a C8.6 on Sep 5, just 0.4 magnitude under M1.0. This was a region repeatedly knocking at the door of the M-class range.

📊 SWF System Performance

Performance is measured at each model's operating threshold (Transformer 7.9%, LightGBM 4.7%). This week must be reported plainly: both models missed all three events — the first 0% POD for either model in an event week since W25's LightGBM, and the first ever for the Transformer.

The raw numbers are the worst of the record, but the three misses each fall into a weakness pattern the bulletin has already documented — and none involved an M5+ event:

• The M3.0 was a new-emergence first-shot. AR 4524 does not appear in either model's tables on Sep 2 (the event day). It entered the Transformer table on Sep 3 at 2.2% and LightGBM's on Sep 4 at 2.3% — the region produced its strongest flare before either model had accumulated enough data to score it. This is the AR 4500 (W31) pattern: not attribution lag, but zero pre-flare data. By the end of the week both models had climbed — Transformer to 7.5%, LightGBM to 6.5% on Sep 6 — but the M3.0 had already happened.
• The M1.2 was a sub-threshold weak event. On Sep 4 AR 4524 was scored at 4.0% (Transformer) and 2.3% (LightGBM) — below both operating thresholds. An M1.2 with 4% probability is not a model error in the statistical sense; it is the tail of a calibrated distribution.
• The M1.0 was a post-decay delayed producer. AR 4520 had been in both tables since Aug 31 (Transformer 4.0% early-week, LightGBM 1.9%) but its probability decayed to 0.1% (Transformer) / 0.0% (LightGBM) by Sep 5 — the day it fired. This is exactly the AR 4492 (W30–W31) pattern: the region delivers only after the models have correctly downgraded it.

The structural takeaway is the threshold cliff. The Transformer issued zero AR-day alerts all week — the first alert-free event week in the bulletin record — yet its highest probability of the week was 7.5% (AR 4524 on Sep 6), just 0.4 points under the 7.9% operating line. LightGBM's single alert (AR 4524, 6.5%, Sep 6) came the day after the region's last M-class event. Both models were climbing toward the threshold as the week ended — a forward-looking signal, not a post-mortem. The C8.6 on Sep 5, in particular, occurred when LightGBM had the region at 3.7% and the Transformer at 6.6%: the models were reading the building stress correctly, one escalation step behind.

• Zero alerts on the Transformer does not mean the disk looked dangerous and was ignored. The highest-scoring region all week (AR 4524) never crossed 7.5%, and the W35 star AR 4513 decayed below 7% by Aug 31. The alert-free posture was a reasonable read of a mostly quiet disk — the problem is the events that did occur all landed in the small-M range where the operating-threshold design has known blind spots.
• All three events were M1–M3. The system has not missed an M5+ event since W25. The record's worst POD weeks (W25, W31, W36) share the same profile: small events from newly emerged or late-life regions. This is a consistent, quantified weakness, not random variation.
• AR 4524's probability trajectory (2.2% → 4.0% → 6.6% → 7.5% Transformer) matches its escalating C-class run (C7.9 → C7.2 → C8.6). The model was tracking the region correctly; it just never crossed the line in time. If 4524 continues to strengthen in W37, the models are positioned.
• HARP 13967 reached 7.4% on Sep 6 (Transformer) — another region knocking at the threshold. The week ended with two regions near the line and none above it.

🔬 Notes

AR 4524 is the clear region to watch heading into W37: it closed the week at 7.5% (Transformer) and 6.5% (LightGBM) with a C8.6 already on the books — the same pre-M signature seen before its M3.0. HARP 13967 (7.4%) is approaching the threshold unnumbered. This week's three misses strengthen the case, first made in W31, for a pipeline-side watchlist that treats newly numbered regions with elevated C-class activity as de-facto alerts even before they cross the probability threshold. With two consecutive quiet-ish weeks now behind us, the next M5+ event remains the key unknown; no region currently on the disk is close to the W34–W35 AR 4513 profile.

🌞 Solar Activity Summary

A single-region week of exceptional density. Every one of the 10 M-class flares (as listed from the NOAA feed — dual-sensor records make some bursts appear in pairs; ~7 independent eruptions) came from AR 4513, the region that announced itself with last week's M8.2. The week's largest event was an M7.1 on Aug 25 at 10:02 UTC, one of five M-class flares AR 4513 delivered within a nine-hour window that morning. No X-class appeared — last week's bulletin flagged this region as an X-class candidate, and while it produced the strongest week of any single region since W27, it stopped one step short.

The sequence: an opening M1.7 at 19:36 on Aug 24, then the Aug 25 cluster (M2.0 at 01:26, M1.0 at 04:02, M1.8 at 06:40, M7.1 at 10:02), then a final pair on Aug 26 (M2.2 at 06:34, M1.0 at 12:14). From Aug 27 onward the region went quiet. No other region produced an M-class event all week — the most concentrated output distribution the bulletin has recorded. AR 4516 (6.6% Transformer peak) and a few unnumbered HARPs were the only other features of note on the disk.

📊 SWF System Performance

Performance is measured at each model's operating threshold (Transformer 7.9%, LightGBM 4.7%). For the first time in an event week, both models achieved 100% POD simultaneously — and LightGBM, which a week ago scored 0.0% on this same region due to a NOAA-mapping lag, now matches the Transformer exactly.

Both models kept AR 4513 at alert for the entire active window and began decaying in perfect step with the region's actual output:

• Every event was caught. The Transformer held 4513 at 72.3% (CRITICAL) on Aug 24–25 — covering the M7.1 day — then 42.6% through Aug 27, catching both Aug 26 events. LightGBM held 65.7% → 76.4% across the two peak days and 39.9% on Aug 26. The 08:05–16:11-style gaps of W34 were absent: this week the region was in both tables from day one.
• LightGBM's 0% → 100% swing is the cleanest possible demonstration that W34's failure was purely the NOAA-mapping lag. Same region, same model, same threshold — the only change is that AR 4513 entered LightGBM's table on Aug 24 (65.7% on its first day) instead of Aug 21 for a region that had been producing since Aug 19. When the pipeline is healthy, the two models agree exactly.
• FAR reached a bulletin low for an event week (57.1% Transformer / 40.0% LightGBM) because all alerts were placed on the one genuinely active region. The false alerts are the post-peak days (Aug 27–30 for the Transformer, Aug 27–28 for LightGBM) when the models were still above threshold while the region had gone quiet — a lag of 24–48 hours between the last flare and the last alert, by far the shortest in the record.
• Lead time rebounded to 14.0 h — the best since W27 — because the alert was already fully in place when the first flare of the week arrived (carry-over from W34's CRITICAL rating).

The most instructive single number of the week is the calibration decay curve. Transformer: 72.3% → 42.6% → 42.6% → 39.5% → 12.8% → 8.7% across Aug 24–30. LightGBM: 76.4% → 39.9% → 26.5% → 15.7% → 3.3% → 2.6%. Both models stopped warning before the region stopped producing — LightGBM's last alert day was Aug 28 and the region's last M-class flare was Aug 26 — and this week, unlike AR 4492 in W30–W31, the decay was correct: the region genuinely quieted. The system's alert posture is now demonstrably tracking regional magnetic activity on a 24–48 hour scale rather than "sticking" to stale risk.

🔬 Notes

AR 4513 remains on the disk but is decaying: both models are at or near their thresholds (Transformer 8.7% on Aug 30, LightGBM 2.6%). Its documented lifecycle — first appearance Aug 20 at 42.6% (same-day alert), M8.2 that same day, ten M-class flares this week including M7.1, then synchronized decay — is the most complete rapid-emergence-to-decay record in bulletin history. A candid revisit of last week's X-class call: the models reached 72–77% on AR 4513, the same territory as AR 4479 in W27 (which produced X1.1), but this region stopped at M7.1. High probability is not a guarantee of X-class — worth noting as a calibration observation, not a failure. AR 4516 (6.6% Transformer) is the only secondary region of note. The NOAA feed call succeeded this week after four consecutive failures — if it holds, the weekly flow is back to fully automatic. No X-class has occurred since July 4; with AR 4513 fading, the next X-class will need a genuinely new major complex region.

🌞 Solar Activity Summary

The three-week calm is over. After the longest zero-event streak in bulletin history (W31–W33), Aug 17–23 produced 6 M-class flares, headlined by an M8.2 from AR 4513 — the strongest event since July 4's X1.3, and the strongest M-class since July 1's M8.7. (The bulletin's NOAA feed call failed a fourth consecutive week; the event list below was hand-verified against NOAA edited_events data and entered via --flares.)

The week's story is AR 4513, a region that emerged with almost no preamble and produced a three-flare run in under eight hours on Aug 20: an M1.9 at 08:05, the M8.2 at 11:42, and an M3.0 at 16:11 — preceded by an M1.0 at 18:51 on Aug 19, the region's opening shot. AR 4507, which both models had rated HIGH for the better part of two weeks, finally delivered with an M1.6 at 09:57 and an M1.7 at 10:18 on Aug 21.

📊 SWF System Performance

Performance is measured at each model's operating threshold (Transformer 7.9%, LightGBM 4.7%). This week produced the sharpest model divergence of the bulletin record — and it is entirely a data-pipeline story, not a model-quality story.

Transformer: 83.3% POD with the one miss fully explained. The single missed event was AR 4513's opening M1.0 on Aug 19 — the region was not yet in any model's table (first appearance: Aug 20). Every event from Aug 20 onward was caught, including the M8.2: the Transformer had AR 4513 at 42.6% on the day it fired, and its AR 4507 alert (five consecutive days at 18–22%) paid off on Aug 21. The high FAR (81.8%) is the small-N effect of a two-region week — 11 alert-days vs 2 confirmed — and most of the "false" alerts were, in hindsight, correct vigilance on regions that did eventually produce.

LightGBM's 0.0% requires the full explanation, because it is the most misleading statistic the bulletin has ever produced. The model's own output shows no loss of discrimination — the failures are entirely in the SHARP→NOAA mapping:

• AR 4513 entered LightGBM's table only on Aug 21 — two days after its first flare and one day after its M8.2. All four AR 4513 events (M1.0, M1.9, M8.2, M3.0) therefore register as missed even though the model, once it saw the region, scored it at 34.2% (Aug 21) and 77.3% (Aug 23) — deep in very-high territory. This is the NOAA-numbering latency documented in W25, W27, and W30, now catching an M8.2.
• AR 4507 vanished from LightGBM's table exactly when it fired. The model had held 4507 at HIGH (19.5–27.1%) for four consecutive days (Aug 17–20), then the region dropped out of the table entirely on Aug 21 — the day its M1.6/M1.7 occurred. The alert was in place for four days; the events landed in a window where the region's data had disappeared from the pipeline.
• Meanwhile the model's alert posture was not wrong: 9 alert-days, all on regions that were genuinely active (4507, 4506, 4513) — none on quiet regions. Every alert-day that "fails" this week is a region that either had already produced or produced within days.

The actionable conclusion is unchanged from W30 but now carries an M8.2: the HARP-fallback attribution fix in rerun_bulletin_stats.py is no longer a convenience — it is the difference between reporting LightGBM's true performance and a 0.0% that contradicts the model's own output. Until it is implemented, every bulletin that involves a newly numbered region will systematically under-report LightGBM.

• AR 4513 is now the most dangerous-looking region of the quiet period's end. Both models closed the week at their highest joint rating since W27: Transformer 65.6%, LightGBM 77.3% on Aug 23. The M8.2 has not resolved the magnetic stress; the region is still climbing.
• AR 4507's W33–W34 arc validated the "delayed producer" pattern. Two weeks of HIGH alerts across both models (W33: up to 43.1% LightGBM; W34: 18–22% Transformer, 19.5–27.1% LightGBM) finally paid off with M1.6/M1.7 — modest events, but the alerts were substantively correct.
• W33's end-of-week warning was directionally right. The bulletin flagged "a quiet week closing with synchronized escalation on freshly emerged regions — W34 will tell whether this is a repeat or a false spring." It was a repeat: the escalation cluster produced 6 M-class events, though the strongest producer (AR 4513) was numbered only after W33 ended.
• FAR remains structurally inflated in small-event weeks: 11 alert-days vs 2 confirmed produces 81.8% regardless of alert quality. The meaningful read is that no alert was placed on an inactive region.

🔬 Notes

AR 4513 is the single most important object on the disk heading into W35. An M8.2 producer whose model probability is rising after the flare (42.6% → 65.6% Transformer; 34.2% → 77.3% LightGBM across Aug 20–23) has historically been a precursor to an X-class event — compare AR 4479 in W27, which reached 72–79% and produced X1.1. The similarity in both magnitude and trajectory is the strongest forward signal the models have given since W27. AR 4507 has delivered and may quiet; watch it for continued post-flare probability. The NOAA-fetch failure that has now occurred four weeks running should be treated as a systemic server-side issue: add a fallback (cached copy or manual entry) to the weekly workflow before the next bulletin.

🌞 Solar Activity Summary

The third consecutive zero-event week — the longest calm stretch in the bulletin record. Aug 10–16 produced no M- or X-class flares, but the week ended with the strongest C-class event in over a month: a C8.2 from AR 4506 on Aug 16 at 19:28 UTC, just 0.2 magnitude below the M1.0 threshold. (Event list verified against NOAA edited_events data fetched locally after the script's NOAA call failed again; the zero-M/X result is confirmed.)

The quiet early week gave way to a build-up on the weekend. AR 4507 and AR 4506 — both new regions emerging in the east — climbed in lockstep across both models: AR 4507 produced a C3.8 on Aug 16 and reached 43.1% on LightGBM (very high) and 18.2% on the Transformer (high) by Sunday; AR 4506 delivered the week's C8.2 and reached 18.9% on LightGBM the same day. The older W32-era regions (AR 4503, AR 4504, AR 4508) remained quiet, with only AR 4503 briefly rising to 6.5% on LightGBM (Aug 13) before decaying.

📊 SWF System Performance

Performance is measured at each model's operating threshold (Transformer 7.9%, LightGBM 4.7%). Zero events again makes POD undefined and FAR 100% by construction; the meaningful signal this week is where the alerts went and how the models ended the week.

All of this week's alerts concentrated on two new regions — and the models' conviction built through the weekend rather than decaying. This is the W26 pattern again:

• AR 4507: four consecutive alert days on both models, rising into the weekend. Transformer: 8.7% (Aug 13) → 9.3% → 9.3% → 18.2% (Aug 16). LightGBM: 29.0% (Aug 13) → 16.3% → 16.6% → 43.1% (Aug 16) — deep in very-high territory. The models did not see this region fade; they saw it strengthen.
• AR 4506 was flagged two days before its C8.2. Transformer alerted Aug 15–16 (8.7% both days); LightGBM alerted Aug 15 (12.3%) and pushed to 18.9% on Aug 16 — the same day the C8.2 occurred. A near-M event was preceded by model escalation, at the C-level stage.
• LightGBM led the Transformer in magnitude again (43.1% vs 18.2% on AR 4507), consistent with the W27-W28 pattern where LightGBM's top-end probabilities ran higher on escalating regions despite its lower operating threshold.
• No alert went anywhere else. AR 4503's one-day LightGBM alert (6.5%, Aug 13) was a single-day excursion. The rest of the 19-region disk was silent. The alert load's focus on the two weekend climbers is exactly what a calibrated system should do.

The parallel with W26 is worth stating explicitly. W26 was also a zero-event week that ended with both models pushing two new regions deep into high territory (AR 4478 at 72.3% Transformer); W27 then produced 67 M/X events. W33 ends with AR 4507 at 43.1% (LightGBM) and AR 4506 at 18.9% with a C8.2 already in hand. The magnitudes are lower than W26's, but the shape is the same: a quiet week closing with synchronized escalation on freshly emerged regions. W34 will tell whether this is a repeat or a false spring.

🔬 Notes

AR 4507 and AR 4506 are the clear regions to watch heading into W34: both are rising on both models, both are new, and AR 4506 has already produced a C8.2 — the strongest C-class since the last M-class flare of July 30. If either crosses M1.0 in the coming days, it will break the longest zero-event streak in bulletin history at exactly the point where both models had already begun warning. AR 4508 (C4.3 on Aug 14) is a tertiary watch. The pipeline caveat from W31 applies with renewed force: if these regions produce M-class flares within their first 24–48 hours of NOAA numbering, the SHARP-to-NOAA attribution lag will again depress LightGBM's apparent POD — the HARP-fallback fix in rerun_bulletin_stats.py remains the single most important outstanding correction.

🌞 Solar Activity Summary

The fourth zero-event week in the bulletin record, and the longest stretch of low activity since late June. Aug 3–9 produced no M- or X-class flares — the largest event all week was a C1.8 from AR 4505 on Aug 9. (Note: this bulletin's event list was verified against NOAA edited_events data fetched locally after the script's NOAA call failed once; the zero-event result was confirmed against the full record — the last M-class flare remains July 30's M1.9.)

The disk grew quieter throughout the week. The four W31-era regions that had closed the previous week — AR 4496, AR 4499, AR 4501, AR 4502 — all faded below 5% and stayed there. AR 4498, the week's only alerted region, decayed steadily (22.5% → 2.7% on the Transformer across the week) without producing. New regions appeared at the very end of the week: AR 4505 on Aug 9 with a cluster of C-class flares (C1.8 peak), plus fresh unnumbered HARPs (13848, 13852). Nothing on the disk currently resembles the complex configurations that produced W27's outbreak.

📊 SWF System Performance

Performance is measured at each model's operating threshold (Transformer 7.9%, LightGBM 4.7%). With zero events, POD is undefined and FAR is 100% by construction — the meaningful quantities this week are the alert count and how quickly it collapsed.

The models behaved in near-lockstep, and both were exhaustively quiet: every single alert this week — 4 of 4 on the Transformer, 2 of 2 on LightGBM — came from the same region, AR 4498.

• AR 4498 is now the first region in bulletin history to sit at HIGH for two consecutive weeks without producing an M-class event. It peaked at 42.6% (Transformer) in W31 and 27.7% (LightGBM) across this week's opening days, then decayed monotonically: Transformer 22.5% (Aug 3) → 18.2% → 11.6% → 8.0% → below threshold (Aug 7) → 2.7% (Aug 9); LightGBM 27.7% (Aug 3) → 17.7% (Aug 4) → 3.7% (Aug 5). Both models correctly released the region as its magnetic configuration failed to escalate — the calibrated decay pattern we first documented in W24, working as designed.
• Both models reached full silence faster than in any prior zero-event week. LightGBM's last alert was Aug 4 (5 full days before week's end); the Transformer's was Aug 6 (3 days). Compare W29's final alert days (Jul 14 LightGBM / Jul 17 Transformer). The system is learning nothing new here — it is simply reading an increasingly benign disk.
• The week's FAR (100%) is a small-N artifact, as in W29. With zero events and only 1–2 alerted regions, the metric carries no information; the useful signal is the near-total absence of alerts on a 13-region disk.

What makes this week notable is not what happened but what did not: the first consecutive zero-event weeks (W31–W32) in the bulletin record, during which the alert load collapsed from W30's 15–17 AR-days to W31's 4–2 to this week's 4–2, all attributable to a single decaying region. If AR 4498 stays quiet into W33, the system will have registered a complete low-activity regime transition without a single high-tier false alarm.

🔬 Notes

The disk is quiet but not empty. AR 4498 — the W31–W32 consensus HIGH that never delivered — has decayed below threshold on both models and should be considered a closed (or dormant) chapter; if it were to re-emerge from the east in a later rotation, its magnetogram history will be instructive. AR 4505 is the week's only new numbered region and its C1.8 is the strongest C-class event of the week — it warrants watching but is not yet flagged. New HARPs (13848, 13852) are forming in the east. As of this writing (Aug 10), solar activity is at its lowest sustained level since before W27's outbreak, and the models' baseline alert load has dropped to near-zero. The next M5+ event will require a genuinely new complex region — currently none is visible.

🌞 Solar Activity Summary

A third consecutive week of low activity, and the weakest week since W24. July 27–Aug 2 produced only 3 M-class flares, all M1-level, with no X-class events. All three events are M1s — the boundary class where flare prediction is hardest.

The week opened with M1.0 from AR 4494 at 02:43 UTC on July 27 — a continuation of the region's W30 activity (it had been at HIGH, 22.5%, when W30 ended). This was the week's only event both models caught, and it was the final event of AR 4494's documented lifecycle; the region decayed steadily afterwards (22.5% → 0.5% across the week).

The remaining activity clustered at a single moment: an M1.9 at 17:00 UTC on July 30 appears in the flare list against both AR 4500 and AR 4492 — almost certainly a dual attribution of the same physical flare. AR 4500 is a newly emerged region that appears in neither model's probability tables all week; AR 4492 had been the week's unresolved question in W30 (five days of Transformer alerts with no payoff), and had by then decayed below both models' alert thresholds.

The structural story of the week is AR 4498: a new region that both models flagged consistently and independently — the Transformer at 15.3–42.6% over six consecutive days, LightGBM at 13.1–21.9% over five — yet which produced no M-class event. AR 4499 was similarly flagged (LightGBM peaked at 27.7%) without producing.

📊 SWF System Performance

Performance is measured at each model's operating threshold (Transformer 7.9%, LightGBM 4.7%). The two models behaved identically this week: both caught exactly the same single event and missed exactly the same two.

The 33.3% POD is the lowest for both models since the bulletin record began, and it needs careful interpretation. The two "missed" events were both M1-class — the hardest events to forecast by construction — and one of them came from a region the models had never seen:

• AR 4500's M1.9 was unforecastable with the current pipeline. The region appears in neither model's table on any day — it emerged and fired within the week without ever being scored. Unlike W30's AR 4493 (which WAS scored, as "AR 0", but failed attribution), AR 4500 never entered the scoring stream at all. This is the hardest variant of the new-emergence problem: not a NOAA-lag accounting issue, but zero pre-flare data.
• AR 4492's M1.9 fired after its risk had decayed. The Transformer had it at 6.6% on July 30 — below the 7.9% threshold; LightGBM at 2.7%. W30's five days of alerts on AR 4492 never saw a payoff while elevated, and the region eventually delivered only after both models had correctly downgraded it. The alert was right in W30, quiet in W31, and the event arrived in the quiet window — a fundamentally hard case for any threshold-based system.
• Both "missed" events are listed at M1.9 at the same minute (17:00, July 30). If this is a single flare dual-attributed to two regions — which the identical timing strongly suggests — the true event count for the week is 2 (M1.0 + M1.9), and POD would be 50% for both models.

The extreme FAR (93.3% Transformer, 88.9% LightGBM) is again dominated by persistent alerts on regions that did not deliver: AR 4498 accounted for 6 of the Transformer's 15 alert-days and 5 of LightGBM's 9; AR 4499 for 3–4 more. With only one confirmed AR-day in the week, any alert load above ~2 days pushes FAR toward 100% by construction. The signal value of the week is forward-looking: both models spent the week building conviction on AR 4498, the region most likely to erupt next.

• AR 4498 is a consensus HIGH from both models with no event — the week's defining call. Transformer peaked at 42.6% (July 31), LightGBM at 21.9%. Both models escalated it in the middle of the week and held through Sunday. Either this becomes the W31 version of "AR 4492" (eventual payoff at low probability) or a genuine false-alarm streak; W32 will tell.
• W30's AR 4492 question was partially resolved. The region did eventually produce (M1.9 on July 30) — but only after both models had decayed it below threshold. The "delayed producer" hypothesis holds, with the caveat that the attribution to AR 4492 may itself be wrong (dual-listed with AR 4500).
• AR 4494's lifecycle is complete: W30 build (4.0% → 22.5%), M3.2 on July 26, M1.0 on July 27, then decay to 0.5%. Both models captured every phase, and the July 27 event was caught at full alert.
• Lead times collapsed this week (2.7 h / 1.7 h) — a direct consequence of forecasting M1-class events, whose precursor signatures are faint and short. The W27-era lead times (16.7 h) correspond to M5–X1 events; weak-event weeks compress lead time by construction.
• Rapidly emerged regions that fire before entering the scoring stream (AR 4500) are distinct from the NOAA-lag attribution issue (W30's AR 4493). The former needs a pipeline-side watchlist for newly numbered regions; the latter needs the HARP-fallback matching fix in rerun_bulletin_stats.py already recommended in W30.

🔬 Notes

AR 4498 is the clear region to watch heading into W32: both models held it at HIGH for the second half of the week and it has not yet produced. AR 4499 (LightGBM 27.7% peak) and AR 4496 (4.0%) are secondary. AR 4500, having announced itself with an M1.9 on day one, should now be entering the scoring stream — its first full week of model coverage will be W32. The Sun continues to operate at the low end of the range seen since late June; a return to M5+ activity would require a new complex region to form, and AR 4498 currently looks like the only candidate building toward that.

🌞 Solar Activity Summary

After two consecutive zero-event weeks (W28–W29), solar activity returned to moderate levels. July 20–26 produced 7 M-class and 0 X-class flares, all from two regions: the newly emerged AR 4493 (six events) and AR 4494 (one event).

AR 4493 was the week's dominant producer. It appeared abruptly on July 20 and immediately delivered: an M2.5 at 14:36 UTC opened the run, followed by an M1.1 at 19:20 UTC, and a tight M3.5/M3.4 pair at 22:25 UTC — four M-class flares in a single day, all within 8 hours. An M1.9 on July 21 and an M3.6 on July 22 (the week's largest) completed its run before the region quieted on July 23–26. The Transformer had AR 4493 at 72.3% CRITICAL on July 21 — a rapid escalation that matched the region's fast emergence.

AR 4494 built more slowly. It first appeared in the Transformer table on July 21 at 4.0%, climbed to 22.5% (HIGH) by July 23, and held at that level for four consecutive days before producing a single M3.2 on July 26 at 15:28 UTC. A textbook slow-build case with days of advance warning.

AR 4492 was flagged repeatedly by the Transformer (8.6–15.9% across July 20–25) but produced no M-class flares — the main source of the week's false alerts. It remains on the disk and warrants continued monitoring.

📊 SWF System Performance

Performance is measured at each model's operating threshold (Transformer 7.9%, LightGBM 4.7%). This week produced a dramatic split between the two models that requires careful interpretation.

Transformer: perfect recall, elevated FAR. The 100% POD is the third time in four event weeks that the Transformer has caught every M/X event (W26-W27 at 98.5% with 1 miss to a new-emergence X-class). The 76.5% FAR reflects a structural pattern seen in most weeks with ≤2 source regions: 17 AR-day alerts distributed across 4 confirmed AR-days inevitably produces a high ratio. Most of the "false" alerts came from AR 4492 (Transformer flagged it for 5 consecutive days at 8.6–15.9% without an M-class event) and AR 4494 (flagged July 23–26 but only produced an event on the final day).

LightGBM: POD=14.3% is a pipeline artifact, not a model failure. This is the single most important number in the report, and it needs to be read correctly:

• AR 4493's six M-class events (July 20–22) all occurred before the NOAA number was assigned in LightGBM's pipeline. On July 20 and 21, the physical region appeared in LightGBM's table as "AR 0" (a HARP without a NOAA number), at 39.4% and 19.7% respectively — both deep in alert territory. The model did see the risk: it had the HARP at HIGH/CRITICAL on the very days the flares occurred. But the bulletin statistics pipeline matches alerts to events by NOAA number. When the NOAA number isn't assigned, "AR 0" alerts cannot be attributed to AR 4493 events — and all four July 20 events and the July 21 event are counted as "missed."
• On July 22, the NOAA number was assigned and LightGBM immediately scored AR 4493 at 76.9% — the single "caught" event. Every other "missed" event is attributable to the NOAA-number lag, not to model insensitivity.
• AR 4494's M3.2 (July 26) appears to be missed in the stats despite LightGBM scoring AR 4494 at 13.6% on that day — likely a cadence-boundary mismatch at the hourly level.
• Evidence from the probability table is unambiguous. AR 0 scored 39.4% → 19.7% on July 20–21 (the flare days). AR 4493 scored 76.9% → 30.9% → 8.2% on July 22–24. The model's risk assessment was correct and timely; the NOAA-number latency broke the accounting.

This is the third confirmed SHARP→NOAA latency incident (W25 AR 4472/4473, W27 AR 4482, now W30 AR 4493). The pattern is consistent: rapidly emerging regions that produce flares within 24–48 hours of appearance are systematically undercounted in LightGBM's POD because the NOAA numbering pipeline lags behind the physical emergence. The fix is not in the model — it is in the attribution logic of the bulletin statistics script, which should match events to the closest HARP when a NOAA number is absent.

• Transformer's alert on AR 4493 was remarkably fast. The region first appeared at 42.6% on its debut day (July 20) — no ramp-up, immediate HIGH. This mirrors the pattern seen with AR 4473 (W25) and AR 4482 (W27): the Transformer is sensitive to rapid-emergence signatures and escalates within hours.
• AR 4494 is a clean early-warning success. The Transformer flagged it 3 days before its M3.2, and LightGBM (once NOAA-linked) had it at 18.7–20.4% for two days beforehand. This is what the system is designed to do.
• AR 4492 is the week's unresolved question. The Transformer maintained alerts on it for 5 days (8.6–15.9%) with no M-class event. It remains on the disk heading into W31 — either a false-alarm streak or a delayed producer.
• FAR across both models is inflated by the same small-N effect: 7 events from 2 source regions gives only 4 confirmed AR-days against 17 (Transformer) and 9 (LightGBM) alert-days.

🔬 Notes

AR 4494 is the clearest threat heading into next week: HIGH on both models, Earth-facing, and its July 26 M3.2 may not be its last word. AR 4493 has quieted but remains on the disk. AR 4492 is the wildcard — 5 days of Transformer alerts with no payoff yet. AR 4496 (7.5% Transformer) is a new region worth watching. The NOAA-number latency issue documented above is the single highest-priority pipeline fix: it has now artificially depressed LightGBM's POD in W25, W27, and W30. A HARP-based fallback matching rule in rerun_bulletin_stats.py would resolve it.

🌞 Solar Activity Summary

The second quiet week in a row — and a clean one. After the historic W27 outburst (67 M/X) and W28's modest six M-class events, July 13–19 produced zero M- or X-class flares. This is the third zero-event week in the bulletin record (after W24 and W26), and unlike those, it came with no emerging region that looked poised to break the streak.

The last remnants of the W27 powerhouse regions faded out entirely. AR 4482 — which had been the most continuously tracked region since W27 (X1.3 → M4.1 → M1.5 → M1.1) — finally dropped below both models' alert thresholds after July 17. AR 4485, the W28 newcomer that produced July 12's M1.1/M1.2 pair, slipped below threshold even earlier (July 13 was its last alert day on Transformer, at 9.3%). New regions appeared — AR 4489, AR 4490, AR 4491, and a crop of unnumbered HARPs (13774, 13777, 13778) — but none showed significant magnetic complexity or rising probability. The Sun appears to be in a genuine interlude between active nests.

📊 SWF System Performance

Performance is measured at each model's operating threshold (Transformer 7.9%, LightGBM 4.7%). With zero events, POD is undefined and FAR is 100% by construction — but the FAR numbers themselves are misleading because the absolute alert count was very low. This is a composure check: did the models go quiet, and how quickly?

The headline is not the FAR (100% — meaningless with 0 events) but the speed and completeness of the quieting. Both models shed their alert posture rapidly after W27→W28's activity faded:

• Transformer issued only 4 AR-day alerts across the entire week (cf. 24 in W27, 15 in W28). All four were concentrated on July 13–14 and July 17, all from AR 4482 (the last W27 remnant) and AR 4485 (the W28 newcomer). After July 17, the model went completely silent — no region above 7.9% for the final two days.
• LightGBM issued only 2 AR-day alerts (cf. 20 in W27, 10 in W28), both from AR 4482 on July 13–14. After July 14 — five full days before the end of the week — LightGBM had zero alerts. Its more conservative calibration meant it disengaged even faster than the Transformer.
• No new region triggered either model. AR 4489 appeared in both tables but never crossed 4.0% on either model. The new HARPs (13774 at 7.5% Transformer, 13777 at 6.6%) are approaching but not yet at the threshold — the models are watching but not shouting.

This is exactly the behaviour expected from a well-calibrated system: alert density tracked activity level downward (24→15→4 Transformer AR-day alerts across W27→W28→W29), and the models did not invent risk where none existed. The 4 Transformer and 2 LightGBM alerts this week are the tail of known active regions decaying — not spurious alarms on quiet regions.

• Three-week trend in alert counts tells the story. Transformer: 24 (W27) → 15 (W28) → 4 (W29). LightGBM: 20 → 10 → 2. The system's alert posture scales with genuine solar activity, not noise.
• AR 4482's full lifecycle is now documented: emerged late W27 → X1.3 (W27) → M4.1/M1.5/M1.1 (W28) → declining alerts (W29) → off the table. This is the most completely tracked region in bulletin history, from first emergence through peak activity to decay, and both models captured every phase.
• W26 vs W29 comparison: both had zero events, but W26 had two regions at CRITICAL (72.3%) that correctly anticipated W27's explosion. W29 has no region above 14.4% — the models see nothing brewing. Whether this calm holds or breaks next week will be an informative test of the current low-alert baseline.

🔬 Notes

The disk is populated but magnetically unremarkable. AR 4482 is finally decaying past the west limb. Several new HARPs (13774, 13777, 13778) are emerging in the east — none yet at alert threshold, but 13774 (7.5% Transformer) is the closest. The system is in a genuine quiet phase for the first time since late June. If this persists into W30, it would be the longest calm stretch in the bulletin record.

🌞 Solar Activity Summary

A dramatic comedown after W27's historic outburst. The Sun returned to modest activity levels over July 6–12, producing 6 M-class and 0 X-class flares — a factor-of-ten drop from the previous week's 67 events. The powerhouse regions of W27, AR 4479 and AR 4478, declined sharply: AR 4479 disappeared from both models' tables after July 6, and AR 4478 followed by July 8. Neither produced another M-class flare after W27's barrage.

The baton passed to AR 4482, which continued the activity it started with W27's X1.3. It contributed four of the week's six M-class events — an M4.1 on July 7 (the week's largest), an M1.5 on July 8, and an M1.1 on July 9 — all caught by both models at their operating thresholds. The remaining two events came from newly emerged AR 4485, which produced a pair of M1.1/M1.2 flares on July 12 after spending several days building.

📊 SWF System Performance

Performance is measured at each model's operating threshold (Transformer 7.9%, LightGBM 4.7%). Both models achieved 100% POD — every M-class event was caught. However, both models also posted elevated FAR (73.3% Transformer, 60.0% LightGBM), which requires interpretation.

The high FAR this week is a structural artifact of the transition from W27's extreme activity, not a model degradation. In detail:

• Carryover alerts from W27's CRITICAL regions. Both AR 4478 and AR 4479 entered the week at CRITICAL (72.3% Transformer). They remained above the alert line on July 6 (42.6% each) and AR 4478 on July 7 (21.9%), despite producing zero M-class events. These 3 AR-day alerts account for a large slice of the false count — they are the tail of a correct W27–W28 assessment (both regions had just produced dozens of M- and X-class flares) decaying naturally.
• Six events across only four "confirmed" AR-days. The M4.1 + M1.5 + M1.1 from AR 4482 clustered across July 7–9, and the M1.1/M1.2 pair from AR 4485 landed on July 12. That gives at most 4 confirmed AR-days out of 15–10 total alert-days — producing the high FAR automatically.
• Both models maintained alerts on AR 4482 and AR 4485 on days with no event (e.g. Transformer kept AR 4482 on alert July 10–12; LightGBM kept AR 4485 on alert July 8–12). These are not "wrong" in a forecasting sense — the regions were active and the model correctly stayed vigilant — but they count as false alarms in the strict binary metric.

The takeaway: 100% POD with modest lead time (10.7 h) on a quiet week is the right headline. The FAR number is inflated by the small-N problem (6 events, 2 source regions) and the W27 carryover tail. On a normal active week with more source regions, both models would be expected to return to the ~25–30% FAR range seen in W27.

• AR 4479, the most prolific region in bulletin history (65 M + 2 X in W27), went quiet this week. Both models recognised the decline: Transformer dropped it from 72.3% to 42.6% on July 6, then off the table entirely by July 7. The calibration decayed risk promptly rather than "sticking" to a past-prolific region.
• AR 4482 is now the most tracked region by continuity: X1.3 in W27 → M4.1 in W28. Both models caught every one of its four M-class events this week.
• AR 4485 emerged east and was flagged days before it fired: Transformer first alerted on July 8 (8.7%), three days before the July 12 flares; LightGBM alerted July 8 (8.8%) and peaked at 27.7% on July 10. A clean early-warning case for a newly emerged region.
• Lead time is shorter this week (10.7 h vs W27's 16.7 h) because this week's events were weaker (M1–M4 vs M5–X1 last week) and weaker events typically have shorter detectable precursor windows.

🔬 Notes

AR 4482 remains on the disk and is the clear region to watch heading into next week, though its activity level is moderate. AR 4485 is the rising threat: both models now have it at HIGH (Transformer 22.5%, LightGBM 27.7% peak) and it just delivered its first M-class pair. Several new HARPs (13744, 13752) are emerging in the east. AR 4478 and AR 4479 appear to have rotated off or decayed below scoring threshold. After two weeks of contrasting activity (explosive W27, quiet W28), the system's ability to scale its alert posture appropriately — high during the storm, low during the lull — is the clearest signal of calibration quality.

🌞 Solar Activity Summary

A quiet week on the flare front — but the models started seeing trouble build. No M- or X-class flares were produced over June 22–28, with activity staying in the C-class background. The story of the week was a rapid run of new regions: AR 4478 emerged on June 23–24 and grew rapidly in magnetic complexity, climbing to 72.3% on the Transformer (CRITICAL) and 35.3% on LightGBM by the end of the week. AR 4479 emerged late in the week and joined 4478 in CRITICAL territory by Sunday. AR 4475 (flagged HIGH out of W25) stayed elevated through Wednesday before easing — no M-class flare materialised from it.

📊 SWF System Performance

Performance is measured at each model's operating threshold (Transformer 7.9%, LightGBM 4.7%). With no M- or X-class events on the week, this is again a composure check — and notably the first week in which both models pushed multiple regions deep into the CRITICAL tier.

Both models independently identified AR 4478 as the most dangerous region on the disk: the Transformer reached 72.3% (CRITICAL) on June 25 and held it for the rest of the week; LightGBM crossed its CRITICAL line (18.76%) on June 26 and peaked at 35.3% on June 27. The two models agreed on direction and timing, differing only in magnitude — LightGBM's calibration remains more conservative at the top end.

• With zero events this week, every CRITICAL alert maps to "false" by construction — standard FAR is not informative here. The signal value is forward-looking: AR 4478 is the region we will be watching most closely next week.
• AR 4475, which both models had at HIGH coming out of W25, declined steadily without producing an M-class flare — similar to AR 4465 in W24. Calibration continues to decay risk on regions that fail to escalate.
• Transformer kept 10 AR-day CRITICAL alerts this week vs LightGBM's 6 — consistent with their relative thresholds and the Transformer's tendency to push earlier on rapidly emerging regions.

🔬 Notes

AR 4478 and AR 4479 are the clear regions to watch heading into next week. AR 4478 is Earth-facing and magnetically complex; both models now rate it CRITICAL. AR 4475 is declining but still on the disk. Several new HARPs (13680, 13683, 13704, 13709) are emerging in the east without a NOAA number yet. If a strong flare arrives next week, both models will have given it a clear advance warning.

🌞 Solar Activity Summary

A week of two halves. From Monday through Thursday the Sun stayed in C-class background, with no M-class activity. Then on Saturday and Sunday two newly numbered regions, AR 4472 and AR 4473, erupted in quick succession: a M1.3 from 4472 on June 20 opened the run, followed by an M1.0 from 4473 the same day, an M2.7 on June 21 early morning, and an M7.0 on June 21 evening — the largest event of the week.

📊 SWF System Performance

Performance is measured at each model's operating threshold (Transformer 7.9%, LightGBM 4.7%). Of the six M-class events, all came from two newly emerged regions on June 20–21 — events the models had to identify with little prior history.

The Transformer ensemble caught the M7.0 ahead of time: it flagged AR 4473 at 8.7% on June 20 — the day of the M1.0 and ~33 h before the M7.0 — then raised both 4473 (17.9%) and 4472 (21.9%) to HIGH on June 21, before that evening's event. The two missed events were both opening shots from newly emerged regions (the M1.3 from 4472 and the M1.0 from 4473); the model picked each region up immediately after its first flare and successfully called the subsequent stronger ones.

LightGBM did not catch any of the six events at its operating threshold — an outcome that warrants explanation. In its output for this week, AR 4472 and AR 4473 appear in the table only on June 21, with no scores during the days leading up to the events; the model's main weekday signal went to AR 4465 (43.1% on June 15), which produced no M-class activity. The most likely cause is upstream data: the SHARP → NOAA region mapping for the new regions appears not to have been linked in time for the LightGBM pipeline (a number of HARPs are listed without a NOAA number this week). We are investigating; this looks like a feed-side latency issue rather than a model-side miss.

• AR 4473 emerged rapidly on June 19–20 and accounted for five of the six M-class events.
• AR 4465, which both models had flagged in W24, declined steadily without producing an M-class flare; the Transformer's confidence in 4465 decayed from 18.2% on Monday to 2.4% by Friday.
• Going forward we will treat new-emergence days as a known weakness of the operating-threshold view and report the first-shot/follow-on split explicitly.

🔬 Notes

AR 4472 and AR 4473 remain Earth-facing into next week and are the clear regions to watch. Two CMEs from the June 21 events are under analysis; a glancing geomagnetic impact is possible mid-week. We are also auditing the LightGBM SHARP-to-NOAA linkage to ensure newly emerged regions are scored from day one.

🌞 Solar Activity Summary

A calm week after last week's fireworks. Following the M-to-X sequence from AR 4455 on June 1–3, the Sun quieted to mostly C-class activity, with no M- or X-class flares produced over June 8–14. The largest event was a long-duration C6.7 from AR 4465 early on June 11, accompanied by Type II and Type IV radio emissions and a CME of roughly 850 km/s. AR 4455 rotated past the western limb and quieted down; AR 4465 inherited the role of the most magnetically complex region (beta-gamma) but did not produce a strong flare. Mild G1–G2 geomagnetic activity was forecast for the weekend from glancing CME impacts on June 11 and 13.

📊 SWF System Performance

Performance is measured at each model's operating threshold (Transformer 7.9%, LightGBM 4.7%). With no M- or X-class events on the week, this is a check on the system's composure: did the models avoid raising spurious top-tier alarms, and did they correctly flag the most complex region present?

Both models behaved as intended on a quiet week. Neither issued a CRITICAL-tier alert. Both correctly identified AR 4465 as the region to watch: the Transformer peaked at 22.5% (HIGH) early in the week, LightGBM at 13.1% (HIGH). As 4465 failed to escalate beyond C-class, both models' confidence in it decayed naturally — the Transformer dropped to 7.9% by Saturday, LightGBM to 2.3% — showing the calibration is not "sticky" on stale risk.

• Standard false-alarm rate is not informative this week: a ~20% probability is, by definition, expected to verify only about one day in five, and only seven such days occurred. With zero events, every alert maps to "false" by construction.
• AR 4455 (the X1.0 source) is now beyond the west limb — off the disk — and no longer scored.
• No CME-driven significant geomagnetic storm materialised; G1–G2 conditions were briefly forecast for the weekend.

🔬 Notes

AR 4465 remains the most complex region on the disk going into next week, but is declining. Several new regions (HARP 13636, 13637, 13642–44) are emerging in the east but have not yet been numbered or shown significant flux. We will continue to track them.

🌞 Solar Activity Summary

The Sun came alive this week, and it was almost entirely the work of AR 4455 — the region we flagged to watch last week. After a quiet start, 4455 produced two M-class flares on June 2 (M1.3 and M3.3), then erupted in a rapid-fire sequence on June 3: an M9.3 at 01:36 UTC, an M7.7 at 07:00 UTC, and an X1.0 at 11:28 UTC. The X1.0 caused an R3 strong radio blackout across the sunlit hemisphere; the two M9/M7 events each drove an Earth-directed CME, prompting a G3 strong geomagnetic storm watch for June 4–5. A second region, AR 4461, added two smaller M-class flares (M1.2 on June 2, M1.8 on June 6).

📊 SWF System Performance

Performance is measured at each model's operating threshold (Transformer 7.9%, LightGBM 4.7%), the same line behind our reported skill scores. This was an active week dominated by one region, so the statistics largely reflect how each model handled AR 4455.

Last week's call paid off. Both models had held AR 4455 at HIGH since late May and kept it elevated (18–28%) all through June 1–3, so both caught the region's entire M-to-X sequence, including the X1.0, well in advance. LightGBM pushed 4455 into its top tier (27.7%); the Transformer ensemble peaked at 22.5% (HIGH, below its 30% CRITICAL line). Both models missed the two small M-class flares from AR 4461 (M1.2, M1.8), which neither raised above its operating threshold.

• AR 4455 was flagged HIGH for roughly a week before the X1.0 — the elevated-risk signal we noted in last week's bulletin.
• Late in the week both models also raised AR 4458 and AR 4462 to HIGH as they grew — the regions to watch next.
• The two missed events were small (M1.2, M1.8) from a separate, less-developed region (AR 4461).

🔬 Notes

The June 3 CMEs drove a G3 strong geomagnetic storm watch and visible aurora at mid-latitudes. AR 4455 remains the dominant region, with AR 4458 and AR 4462 now also active. Solar activity is likely to stay elevated into the coming week.

🌞 Solar Activity Summary

A quiet week. Nine NOAA-numbered active regions were on the disk and flaring stayed at low levels — mostly C-class. The strongest event early in the week was a C9.7 from AR 4446 on May 26. A large far-sided partial-halo CME (first seen late on May 25) drove a brief S1 minor solar radiation storm on May 26, and a coronal-hole high-speed stream left the geomagnetic field unsettled-to-active over May 30–31.

📊 SWF System Performance

Performance is measured at each model's operating threshold — the decision threshold set during training (Transformer 7.9%, LightGBM 4.7%), the same line behind our reported skill scores. With only one M-class event this week, these figures describe behaviour, not skill: a single low-level flare cannot rank the models, and false-alarm rates on probabilistic forecasts are not meaningful over one week.

Both models tracked AR 4455 as it grew. The Transformer caught the week's only M-class flare: it flagged AR 4455 above its operating threshold (12.4%) on May 28 — the day before the M1.1 — rising to 21.9% (HIGH) by May 30–31. LightGBM crossed its own threshold on AR 4455 roughly a day later (15.7% → 17.7%, HIGH on May 30–31), so it did not flag the region before the May 29 flare. Neither model issued a CRITICAL-tier alert this week.

• AR 4446 stayed elevated all week (Transformer ~18–22%, LightGBM ~5–7%) but produced only C-class — expected for a ~20% forecast, which usually does not verify on a given day.
• One low-level M-class event is too little to separate the two models; we report both for transparency and will compare them over longer windows.

🔬 Notes

AR 4455 is the clear region to watch: magnetically complex and still developing, and both models now rate it HIGH. We will be tracking whether its risk crosses into CRITICAL in the coming days.