Please precisely define...

[lightvector]

There seems to be a disconnect between Daniel's problem and the proposed suggestions.

The metric needs to be very simple, to be possible to implement as an practical algorithm and applied by computer on the order of 100 million times. It is already difficult to program any sort of open-space tactical reading with concepts and goals such as stability, connection, etc. Gnugo is probably the best freely available attempt at this sort of thing (that is, hardcoding expert knowledge and judgement), and we all know how poor Gnugo is. Then on top of all that, getting it fast enough...

And with millions of positions, it is not feasible to do it all manually in any reasonable amount of time.

The approach perceval and Daniel mentioned he is using right now is pretty close to what I would try first, if I were trying to implement this. I don't see an obvious way to do much better, besides fiddling and tweaking the algorithm and inspecting the results to see where it goes wrong, to try to fix it.

As for averaging number of moves before playing again, I think it's a metric worth investigating. I don't know how well it will do, but I'd go ahead and implement it and look at some of the results to see if they give anything reasonable. Even though it does depend on the rest of the board, averaged over thousands of games, there's a good chance it will correlate with some notion of urgency.

You might also consider different ways of weighting the number of moves before another local play prior to averaging. For instance, the difference between "played immediately" to "played after 5 moves" might matter a lot more than the difference between "played after 45 moves" to "played after 50 moves".

[RobertJasiek]

I do not share the pessimism about programs and expert knowledge. If you can't implement tactics, then let the tactics be done by Monte Carlo and the like. It can't be that tough to describe as code things like stable, urgent, investment, unsettled, important (too important to be sacrificed) along the definitions for those terms in my book. Note that, unlike traditional expert system programs, we do not run into their difficulty of mutually competitive expert level terms because I invent them essentially hierarchically and the OP is interested only in tenuki considered alone.