I had a long 1 hour+ with Googles AI about why NCM Pro gives out different answers on a close game if I’m not confident of the answers given to me because I’m having to rerun your 16 x CPU Core
4.5GHz AMD Ryzen 7950X Twice , sometimes 3+ times as it spits a different answer out each time and swaps the best move around etc bast on score by pieces (+ for white & - for black) , the Google AI pointed out how it’s calculated on the score of pieces down to 0.10 of a pawn + it gave a good explanation of how re-running the same game over and over on your AI engine will have different branches each time leading to slightly different variations creating different best moves,
I’ll quote part of its many answers below:

“ 1. Parallel Processing and Timing Fluctuations
The NCM service uses a high-performance 16-core AMD Ryzen 7950X server. Chess engines like Stockfish achieve their immense strength by running complex search algorithms across multiple CPU cores simultaneously.
Race Conditions: In a multi-threaded (multi-core) environment, the precise order in which different threads explore parts of the search tree can vary slightly from run to run due to minor operating system scheduling differences, memory access timings, and other system-level fluctuations. This creates minor "race conditions."
Different Search Depths/Paths: These tiny initial variations can cause the engine to explore slightly different branches of the complex chess position search tree. Over the course of thousands or millions of positions evaluated per second, these initial differences compound.
Time Control: NCM typically runs for a specific duration (e.g., 10 seconds per move). Because the exact number of positions (nodes) evaluated within a fixed time frame can fluctuate slightly (e.g., one run might complete 26 million nodes, the next 24 million in the same 10 seconds), the final "best" move found might differ if the variation in search depth leads to a change in the top choice.
2. Engine Randomization (if applicable)
While the default Stockfish engine used by NCM is largely deterministic under identical conditions (same nodes searched, single thread), many modern engines or analysis platforms incorporate a minor degree of randomness, especially in the opening phase or when multiple moves are equally strong, to encourage variety in games and prevent endless repetition of opening books in engine vs. engine matches.
3. Server Load and Resource Allocation (less likely for NCM Pro)
Next Chess Move states that for NCM Pro users, each calculation gets 100% of the server's resources and the server handles only one calculation at a time. However, if you are not using NCM Pro, or in edge cases, background processes on the server might introduce minor timing variations that influence the final output.
In essence, while the NCM server is consistent in its power, the complex and massive scale of the engine's search tree means that minor, unavoidable micro-fluctuations in timing and parallel processing cause the engine to consider a slightly different sequence of moves and nodes on each run, which can result in different move recommendations or evaluation scores.“

Anyways I was telling Google AI that near the end of a hard game I’d still have to work out myself what I thought was the best move ,

Eventually I said to its AI that it would be really handy near the end of a game if yourselves added in an extra AI that told us the Pros and Cons of each move instead of the score based on having a pawn extra etc ..
as human players running NCM Pro and getting different answers if running more than once. Because we’re not sure of the first answer and then getting multiple close different answers on your server, we can only see so far ahead ,
It’s like trying to figure out the best and fastest way to climb 100 different trees with different branches, it’s a lot of thinking time for a human , a good computer with good AI can figure that out much faster, Chess is like 1000+ fast growing trees with many branches..

So yeah, when a chess game gets down to the last 6 - 8 pieces, there’s not as many different branches and so if different moves are given out on re-runs based on 0.10 of a pawn or more in difference,
It would be great if you had an extra AI eventually that gave out some pros and cons for each move, even telling us what pieces are strengthened and weekend plus telling us the long game pros, how we will benefit 10 moves on etc ..

I hope I’ve explained that well enough.
Thank you

& Best regards

Thomas Rutherford