Wolf's Cave (vdB 152 / Barnard 175)

This is the full capture log for my Wolf's Cave image, in the order it happened, including everything that went wrong.

I build my N.I.N.A. sequences with an AI assistant that also reads the logs the next morning. The times and numbers below come straight from the NINA and PHD2 logs and plate solves of the actual frames — not from memory. Several of the mistakes were the assistant's. Several were mine. A couple were the hardware's. They are all here.

The day before (Mon Sept 7) — prep, and the first near-disaster

The first framing on paper was wrong twice. The catalogue positions put the dark nebula south of the reflection nebula; the actual dust column in photographs runs north. Lesson: frame on pixels, not on catalogue coordinates. The numbers from that first attempt were never used on the sky — but they lingered in a prep note and in the sequence builder's defaults, and nearly got used for night 2.

The assistant built the night's sequence from a saved template whose filename said "Wolf's Cave." I asked why we weren't copying a sequence that had actually run a full night. Good thing: it turned out eight of my saved templates had "Stop Guiding" at the end but no "Start Guiding" anywhere. Any of them would have shot the whole night unguided. Rule since: only build from a file that has actually run, and check for Start Guiding every time.

There was no known focus position for the 384 mm configuration. The assistant estimated one by interpolating between my other two configurations: about 11,900 steps. The real answer turned out to be about 4,750. Never interpolate a focuser position from focal length.

Night 1 — Mon Sept 7 / Tue Sept 8 — a good night with a bad start

Plan: R, G, B at 60 s, then L at 90 s until 05:00, then short 10 s frames in each filter.

• 21:22 First autofocus of the night. The camera reports "exposure failed" three times in a row.
• 21:24 N.I.N.A. prints "Aborting Sequence!" The whole night's plan stopped over one failed autofocus. Why: every Center and Autofocus instruction in my template was set to "abort on error." Found the next afternoon; every instruction is now "continue on error" and the builder refuses to write a file otherwise. Nothing is allowed to abort the night.
• 21:38–21:52 Camera reconnected, sequence restarted, guiding up, first R sub at 21:52. Cost: 25 minutes and 17 red subs. The suspect was the camera's 12 V power sagging under start-of-night cooler load. We wrote "unverified" and moved on. That was a mistake — the same fault came back ten days later and cost an hour before I traced it to the cigarette-lighter (car charger) plug in the camera's power lead. Those plugs make a lousy contact. It has been replaced with a proper straight plug and the fault has not returned. No car-charger plugs anywhere in the 12 V chain.
• 00:15–00:36 Meridian flip. Fifteen silent minutes waiting at the meridian, slew, refocus, recentre. It worked first try, but: after the flip my rotator cannot reach the requested angle, so N.I.N.A. quietly folds it by 180 degrees. Same sky, upside down. Red, green and half of blue are one way up; the other half of blue, all the luminance and all the short frames are the other way. We now tag every file as pre-flip or post-flip so nothing gets stacked blindly.
• The assistant then told me the rotator had spun 358 degrees and slipped its encoder. I asked whether that came from the headers or the data. It came from a log line that is arithmetic, not motion: the real move was 1.5 degrees and took under five seconds. Withdrawn. FITS headers and log arithmetic report intent; only a plate solve reports the sky.
• 01:03–05:00 Luminance, 148 subs, uneventful. Guiding 1.3–1.4″ after a loose first hour (2.98″). Remember that loose hour.
• 05:00–05:35 Short frames. The block ran twice and the 05:35 cutoff chopped blue off at 22 of 40.
• The R and B counts came out at 43 and 44 instead of the planned numbers. We did not understand why until night 2.
• Dawn: flats and flat-darks in all four filters, then 35 × 10 s darks, because it turned out no 10 s darks existed for the short frames we had just shot. Check the dark library against the exposure list before the night.

Night 1 haul: L 148, R 43, G 61, B 44, shorts 142. Median HFR 1.55–1.79 px. 423 of 438 kept.

Night 2 — Tue Sept 8 / Wed Sept 9 — the bad one

Clear sky, but hazy with poor seeing. Plan: a few R and B top-ups, then L all night. That afternoon the assistant also wrote and installed a N.I.N.A. plugin that pauses the sequence when clouds kill the star count.

• 20:04–20:30 Polar alignment. The last good reading was 13 arcminutes off and the next three solves failed, so I never confirmed it.
• 20:36 Pressed Start. The sequence had no "wait for darkness" at the top, so it slewed and started shooting with the sun 12 degrees down. The assistant argued the filter could tolerate twilight. I disagreed, loudly. Stopped and restarted at 20:47. Rule since: every sequence opens with a wait until astronomical dark minus 12 minutes (the 12 minutes covers slew, centre and focus). No exceptions, no "this filter tolerates twilight."
• 20:53–21:56 The red block was meant to be 17 subs. It shot 17, recentred, refocused, shot 17 more, and started a third pass. In N.I.N.A. a block's end time is not a cap — it means "keep repeating until this time." That is also what happened to the counts on night 1. I cut it off by hand at 46. Rule since: fixed-count blocks get no time condition at all; only the last "fill" block gets a dawn cutoff.
• 21:56 For blue, the assistant advised a live edit of the block's end time to 22:00. Centering and autofocus used up the four minutes. Blue got zero subs. Never give a block less than about 15 minutes.
• 22:00–05:30 Luminance, 252 subs. Meridian flip at 00:10, first try. Eight guider settle timeouts cost another 15 minutes.
• Guiding was 5 to 9 arcseconds RMS all night against a normal 1.3–1.9.
• 05:45 At dawn I checked the guide scope's focus. Guide stars were blobs 7 pixels wide. Refocused to 2.5. It had been like that on night 1 as well — that was the loose first hour.
• I then cut the power without shutting Windows down. Everything had already closed; no damage, but not a habit to keep.

The post-mortem was worse than the night felt. Quality control kept 126 of 298 subs — 58 % thrown out as trailed or soft. The guide log explained it: from 01:34 to 04:36 PHD2 had lost the guide star on 45 to 71 % of its frames. On a lost frame it sends no correction at all, so the mount was effectively unguided for most of three hours. And PHD2's RMS figure never showed it, because RMS only counts the frames where the star was found. The haze had thinned the field down to faint stars, and an out-of-focus guide scope did the rest. Hour by hour, the luminance subs with more than 30 % lost guide frames: 01h 14 of 34, 02h 37 of 38, 03h 35 of 35, 04h 22 of 37. We lost two-thirds of a clear night.

Fixes that came out of it:

• Guide scope refocused (the actual cure).
• PHD2 camera gain 44 % to 60 %, star-mass-change detection off.
• The assistant added a "Guide Star Watchdog" to the plugin: if two subs in a row had more than 30 % lost guide frames or a weak guide star, it stops guiding, picks a new star and restarts. I can't be awake to check which star PHD2 picks at 2 am.
• Judge a night's guiding by lost-frame percentage per hour, not RMS. At every connect: guide star 3 px or smaller, SNR over 30.

One more wart the next morning: the rebuilt sequence loaded into N.I.N.A. looking perfectly normal and was broken inside — 135 errors in the log, "failed to deserialize." The hand-inserted wait instruction had duplicated an internal ID. Fixed in the builder, and the habit now is to search the log for that phrase after the first load of any generated file. Silence means pass.

Night 2 haul: L 252, R 46, B 0. Median HFR 2.33 (night 1 was 1.79). 126 kept.

Night 3 — Wed Sept 9 / Thu Sept 10 — the one that worked

Sized from the cull: luminance was now 68 % of my kept time and the rule I work to is 50–60 %, so this night was mostly colour. And in case it was the last clear night, an hour of Ha first.

Plan: Ha 20 × 180 s, then R 64, G 90, B 94 once each, then L until 05:30.

• 20:21 Started. The wait held everything until 20:51, exactly as designed. First time that had existed, first time it worked.
• 21:05–21:09 The watchdog's first real save: guide star SNR 16 on the first two Ha subs, it re-picked a star, SNR 31 and zero lost frames afterwards.
• 21:27 The mini PC's Wi-Fi dropped and stayed down for eight hours. Remote desktop dead, phone notifications failing, weather logger erroring. The sequence, running locally, never noticed. I did not know that at the time.
• 22:10 R, 23:22 G, meridian flip 00:08–00:24 first try (green split 43 before / 47 after, so again two orientations), 01:22 B, 03:08 L.
• After the flip, the recentre turned the rotator about 1.5 degrees, so post-flip green sits 1.5 degrees off blue and luminance. Registration handles it; it is flagged in the data notes.
• Red was the soft channel (median HFR 2.11 against 1.60 for green): 68 minutes on a single autofocus while the evening was cooling fast.
• The forecast had the dew point spread at 0.0 °C from 04:00. No dew came; luminance stayed at HFR 1.65–1.76 to the end.
• 05:30 Mount parked. With remote desktop dead I hard-reset the PC seconds later — it landed between the park and the camera warm-up, so the camera went from −5 °C to off without warming. No damage. Lesson: a box that is off the network is not a box to reset; let the run finish. A Wi-Fi watchdog task now pings the router every two minutes and reconnects. The real fix is an Ethernet cable.
• 05:48 Ha flats and flat-darks with the rotator untouched.

Guiding all night: 0 lost frames out of 13,800, 1.1–1.7″ RMS. Same mount, same target, same settings as night 2. The only change was focusing the guide scope.

The watchdog fired five times; three were marginal (SNR 18.8–20.0 against a threshold of 20), so the threshold is now 15.

Night 3 haul: Ha 20, R 64, G 90, B 94, L 91 — all 359 planned subs. Median HFR 1.60–1.68 apart from red. The sharpest night of the three.

Afterwards

• Sept 10: luminance and colour were judged sufficient; the only thing left on the list was more Ha on a moonlit night.
• Sept 18: I scrapped the Ha plan. The processing side then asked for roughly ten more hours of colour, blue being the noisiest, and a colour-only run was saved for the October dark window.
• Sept 20: I reprocessed what I had, I am happy with it, and the target is closed. The saved run went to the archive unused.

No clouds on any of the three nights. Everything that cost me data was a file mistake, a focus ring, or a power plug.

What this target changed for good

1. Every sequence starts with a wait until astronomical dark minus 12 minutes.
2. Fixed-count blocks have no end time. One dawn cutoff, on the final fill block only.
3. No instruction is allowed to abort the night.
4. Sequences are built only from a file that has already run a full night, and checked for Start Guiding.
5. After the first load of a generated sequence, search the N.I.N.A. log for "Failed to deserialize."
6. Guide star 3 px or smaller at connect. Judge guiding by lost frames, not RMS. A watchdog re-picks the star when I am asleep.
7. After a meridian flip the frames are 180 degrees round on this rig — tag them. Headers report intent; plate-solve the pixels.
8. Check the dark library against the night's exposure list in daylight.
9. No car-charger plugs in the 12 V chain.
10. Wi-Fi watchdog on the PC, and don't reset a machine just because you can't see it.
11. Every issue gets written down with its fix the day it happens. The 12 V plug cost me twice because the first time we wrote "unverified" and moved on.