The Japanese clock stopped after 26 days.
Update: The new replacement mainspring showed a similar curve, reaching a top on the 18th day, comparing lines 2 and 3 in the graph below, suggesting that the variation was more likely caused by the escapement instead, possibly in a longer term trend line. Whereas the old mainspring was 0.015" thick and 170" long, the new mainspring was 0.012" thick, with only half the strength, and was only 70" long. I decided to try it because my spreadsheet for mainsprings calculated that I would get almost 26 turns with the old mainspring, and 23 turns with the new one, so the difference would be small. I polished the pivots of the escape wheel and the fourth wheel, replaced the mainsprings and one bushing. The Japanese clock ran for 34 days with the new mainsprings, (though the strike stopped after 25 days).
 |
With the new mainsprings, the Japanese clock had a maximum variable error of 35 seconds on the 18th day, shown in line 3, or a variable error of about 2 seconds per day. Between the 18th day and the 30th day, the clock gained about 3 seconds per day. Line 4 shows the performance for the Korean clock with new 0.012" mainsprings, which raises more questions than answers because it is more similar to the lines for the Japanese clock rather than Line 1. The data show that by lowering the mainspring strength by 50% will lower the variable error in the timekeeping by at least 50%. Furthermore, the duration of up to 36 days is not affected by the mainspring length by much, shown by replacing a 170" long spring with a 70" long spring. The duration is much more affected by the size of the mainspring barrel, or the amount of space the spring has available to expand in. The photo below shows the Korean clock with the new 0.012" mainsprings. The strike mainspring is fully wound, and barely visible. I left the time mainspring unwound for the photo, so that you could see it.
 |
The semi-deadbeat escapement in the Korean and Japanese clocks was more accurate than all the others because there was a small amount of recoil to offset the variation in the Graham escapement, and the small amount of recoil was symmetrical (the same) for both the entry and exit pallets. While the semi-deadbeat escapement in the Seth Thomas 89 was more accurate than a recoil escapement, it was less accurate than the Graham escapement in the Hermle clock. The recoil in the escapement of the Seth Thomas 89 was not the same for both pallets, resulting in greater variation in timekeeping. Everything in the Korean and Japanese clocks was designed to keep manufacturing costs as low as possible, and keeping costs to a minimum is evident in the design of the pallets. Finding nothing in the clock or the pendulum that suggests an effort to make a superior timekeeper, the conclusion has to be that the superior accuracy was coincidental.
Studying the Korean and Japanese clocks on this page revealed more information than any previous experiment on my website. The only mechanical clock I ever had which kept more accurate time than this Japanese clock was a Self-Winding Clock (ca. 1910) with a 39" mercury-compensated pendulum and a total error of 20 seconds per month. Another clock which performed similarly was my Hermle 1161 grandfather clock with its 39" non-compensated pendulum, but this clock did not have mainsprings. My Herschede never came close.
Update:
After using the Japanese clock for two years, I decided to repeat the test and compare the results. The mainspring would be expected to lose some strength over time, and the effect on timekeeping could be observed. The effect was small. The biggest change was in duration: the clock stopped after 28 days in 2020, compared to 34 days in 2018. The data from two years ago were taken from the graph above (look for "3 Japanese (0.012)") and modified to bring the variation in timekeeping, on the vertical axis, to zero on the 28th day instead of the 34th day, so that the two curves could be compared. The days are shown on the horizontal axis. For a mechanical clock to have a maximum cumulative error under 25 seconds over a 28 day period is really good.
Another update (July 2022):
Looking at the graph for the Korean clock, because it has the most data points, we could add a parabola with the following quadratic equation.
Y=-0.2863(X-18)2+92.58
If we expand and differentiate the quadratic equation, we get the rate of change of error in seconds per day, which is linear, and which is proportional to the mainspring torque.
A new arrival caught my attention, a 1975 Hermle 1051-020/15cm which also proved to be unusually accurate. Below are graphs comparing this Hermle to the Korean clock over 7 days.
After adjusting the data:
In the second graph, the Hermle gained a maximum error of 6.2 seconds in 4 days (or an average of 1.55 seconds per day), whereas the Korean clock gained 6.6 seconds in 3 days (or 2.2 seconds per day), during a test run of 7 days. The results would have been closer if I had been able to use the same mainspring in both clocks: the mainspring in the Korean clock was about a third stronger than the mainspring I installed in the Hermle. A clock with a short pendulum is not expected to keep time as accurately as a clock with a long pendulum, like a grandfather clock, but this Hermle with its very short pendulum, only 15 cm., proved to be one of the most accurate, among clocks with mainsprings, that I have seen in over 30 years at the bench. In fact, both clocks had superb results.
Two Extraordinary Hermle Clocks
Clock Repair Main Page
Escapements in Motion
Links Page