I proposed 250 candidates for good cells and Stephen put an x in col 1 of a spreadsheet where he disagreed. This note documents what we know about some examples of the cells that were marked with an "x".
 
The current situation is that we have two basic sources of information that tell us which FMS large cells are good and which are bad.  I have been using method 1) to make the call and Chris and Stephen have depended on method 2 for their judgement.  

More often than we like, the two approaches are giving different answers as to whether a cell is good or bad.  

One cell that I claim has a healthy base and Stephen rejects is cell (NSTB=1,ROW=0,COL=9).

In this note I am putting together all the information on that one cell for us to talk about.  I then follow up with basic data on several more.

Method 1 is based on all of the calibration led data taken in Run 12 and Run 13. The callibration data is supposed to be a continuous monitoring of an led trigger during regular data collection at a rate of about 1 hz. Sets of about 40 of these led events are taken every minute or so and an average led pulse height is calculated with an error based on the distribution of 40 events.  This average LED pulse height is plotted vs time, (typically about 1 point per minute).  

Data is plotted on a pseudo-day  time scale, based on run number and event number. Within a day, the fractional day time is obtained from an algorithm and is really based on run number/event number for the led events used.  When ploted over a 30 day period, all the points of a particular run occur very close together on the time axis but it is possible to zoom in to a shorter period of time and observe the minute by minute variation.

Data for Run 12 extends from Day 45 to Day 75 of Run 12.  The plot of the run 12 ave led points vs the pseudo-day number is shown here for (NSTB=1,ROW=0,COL=9).  

Data for Run 13 extends from Day 107-140 of Run 13. The plot of the run 13 ave led points vs the pseudo-day number is shown here for (NSTB=1,ROW=0, COL=9). Note that changes in the trigger significantly reduced the led rate in the Run 13 data so the time for 40 led events was significantly larger than the minute that had been planned. 

The Run 13 data showed significant reduction of pulse height with time. This was because of the darkening of the lead glass from radiation. The pulse height in each cell was fit to an exponential for pulse height vs. time in Run 13 data. The chi square of the fit is an important test of the stability of the cell with time. Note that plots like that below contain thousands of points at slightly different times. The points appear on top of each other and the spread of error bars is dominated by the tails of the distribution. The errors are difficult to read on the plot below. The chi square is calculated point by point and is not affected by the fact that the points appear so close together.

The figure below is a blowup of the very last part of "Day" distribution shown above. The individual points on a small part of day 140 are displayed. Again, the time per point is on the order of 1 minute. It is seen that the true errors on individual points are smaller than they appear in the plot above with many points plotted on top of each other.

I have zoomed in to the "Day" time distributions for (NSTB=1, ROW=0, COL=9) with various magnifications and various time regions in the attached pdf file

Based on method 1, I have ranked the top 250 large cells. The rankings are posted here.
The spreadsheet of my list of best cells is also found in that posting.  (NSTB=1,ROW=0,COL=9) has a very good score and survived a visual final cut based on these data to on the list of good bases.

Method 2 is based on a dedicated LED run taken at the end of Run 13.  In consisted of > 10,000 led events collected over some short period of time, I guess a few hours. Details of analysis of that data are shown on a drupal page by Chris Dilks.

My reading of this page is that the cell (NSTB=1, ROW=0,COL=0) is fully discussed on that page.

Chris posts a figure called pmt health. The cell in question is in the top row and is the leftmost cell in the top row.
The key for this cell says that "white" is a healthy cell, this one is colored white.

However later another plot on his same drupal page is labeled as indicating the   

"Number of peaks in ADC histogram, minus 1 (i.e. if a cell has nonzero value, there is likely jumping ADC rate)"

This time the cell (NSTB=1,ROW=0,COL=9) is marked with "yellow" indicating about 17 peaks in the adc distribution. 
Chris actaully uses this cell as an example and he explains that his cell is one with the shift set to create apparent 
peaks in the adc distrubution.

Looking at the actuall distributions for this particular channel, the distributions look fine but apparently the peak actually has gaps due to adc bit shifting. This has nothing to do with the phototube base and does not seem to be an issue. The question does remain as to why this channel was bit shifted and when.
 

----------------------------------------------------------------------
----------------------------------------------------------------------
Now lets look at some other cells on which we disagree.
------------------------------------------------------------------------
(NSTB=1,ROW=0,COL=9)
This one is white on the health chart. I don't know what the issue is.
------------------------------------------------------------------------
(NSTB=1,ROW=3,COL=6)
This is marked as varying on the health chart.
Based on the data distribution I do not understand this rating.
-----------------------------------------------------------------------
(NSTB=1,ROW=4,COL=12)

This is marked as varying. Here is the end of run led data distribution.
There is some variation here.

Lets zoom in on the 40 day RUN 13 led data and look for instability.

This looks like a case where during the few hours of calibration data at the end of the run 13 we see instability that did not show up during the 40 days that preceeded. I don't know how to reconcile this. I doubt this is about the photo-tube base.
 
-----------------------------------------------------------------------
(NSTB=1,ROW=5,COL=4)

Method 2: Marked as junping 

Method 1: Good

Lets zoom in on the 40 day RUN 13 

-----------------------------------------------------------------------
(NSTB=1,ROW=5,COL=11)

Method 2: Green varying

Method 1: Good 
Lets zoom in on the 40 day RUN 13

-----------------------------------------------------------------------
(NSTB=1,ROW=7,COL=10)
Method 2: Green varying

Method 1: Good 
Lets zoom in on the 40 day RUN 13

-----------------------------------------------------------------------
(NSTB=1,ROW=7,COL=14)

Method 2: Green varying

Method 1: Good 
Lets zoom in on the 40 day RUN 13

-----------------------------------------------------------------------
(NSTB=1,ROW=8,COL=9)

Method 2: Light Blue jumping 

Method 1: Good
Lets zoom in

-----------------------------------------------------------------------
(NSTB=2,ROW=27,COL=11)

Method 2: White (but there is an "x" on spreadsheet)
There is a lower band of adc hits but it does not look to me like the base or tube. 
Perhaps secondary crossing trigger?

Method 1: Good
Lets zoom in 

-----------------------------------------------------------------------
(NSTB=1,ROW=6,COL=14)
Method 2:  Blue jumping

Method 1: Good
Lets  zoom in

-----------------------------------------------------------------------
(NSTB=2,ROW=18,COL=15)
Method1:
Method2:

-----------------------------------------------------------------------
(NSTB=2,ROW=19,COL=14)
Method1:
Method2: