Here is an idea to speed up the automation process. We use very commonplace machine recognition ballots to train an algorithm to determine if a ballot was for which Presidential candidate, or flagged as indeterminate if uncertain. Off the top of my head, we could do a CNN (convolutional neuaral network -- not fake news) based model in tensorflow, with an open source training code. This is incredibly simple, and could be trained in one day (for each unique type of ballot in different states and perhaps counties).

This would speed up ability to check and highlight problematic areas. We can run the checks against ballot counters in the same piles, then look into discrepancies. This program should be more flexible and robust in counting a selection. It should also easily resolve some errors that arise from things like duplicating a single ballot many times, sharpies or whatever, and resulted in incorrect removal of ballots due to error. A separate thing to do is see if there were any systematic adjustments to the ballot images in certain areas (CodeMonkeyZ on OAN said if you adjusted the gamma systematically, you could get all ballots flag and then have power to manually determine whether to keep all ballots there). The program could make the correct reading even if these happened, only count unique ballots, and also give a report of where any "errors" happened with statistically unlikely frequencies. It will be accurate, save time (even granting you need to check flags later), and will be unbiased and open source -- so should be largely trusted and persuasive. (I look forward to seeing fake news try to explain why a simple machine learning program with open source is not trustable.)

A separate program can check the likelihood of ballot issues by cross referencing registration data. It would be flagging things mentioned before, like impossible return times for mail-ins, suspicions for voting where you did not live, etc.

One of the things this can show, besides individual instances of fraud to look at in criminal investigations, is a case comparison between counties. We would get an output that shows the % of similarity between machine-learning counts and the ones counted in different precincts that had various counting programs, election officials, and status for being a critical swing-vote area. Unless fraud were equally applied everywhere, we would see huge anomalies in certain districts. That would be reinforcing some of the decisions and be an unbiased measurement.

To implement the first stage and make the program, we first need a few thousand images of ballots where the correct candidate is filled out and passes audits and this is used to build the model, and we need a test set that can just be a few thousand ballot images that are stored from the scanned dominion system output. I don't think these ballots are currently available, but Powell and Trump's lawyers have the capability to get access through a public records act (as do others, but I think they'd be more successful). In the future, this could be kept as a safeguard to detect fraud. I would say it could replace dominion and be a worthwhile profitable venture, but we should aim to not rely on electronic voting machines because there will be vulnerabilities to manipulating the output, even if the program was faultless.