Saturday evening I finally got down to watching Sam's talk . It's excellent . Thank you Sam . I think there are few others who could have gone thru the code you did as clearly as you did .

I have clipped it and uploaded it to my Coherent Systems YouTube channel . I'll be annotating it anon . For the nonce , here are some thoughts at timestamps relevant to the beginning .

• 02:11 | It is relative .
  
• 03:30 | Once an object is allocated , I don't see why performance should not be as good as any static allocation . Which is not to say my implementations are optimal .
• 05:05 | I chose Reva because it was "light-weight" compared to Win32Forth and , for that matter Stephen Pelc's VFX . And had aspects I liked from my path in APL and K .
  The symmetric vertical bar " | " comment symbol is one . As is case-sensitivity .
  I didn't know when I defined  s"  to be a CoSy string that it was used by some standards like G-Forth .
• 07:00 | CoSy's size .  Not super important , but diverted me into seeing just how many lines actual are in the CoSy.f file . As Sam notes , there are lots of comments because comments are included with definitions and the help function ?? simply searches all source files for the occurrence of any phrase and returns all text containing it delimited by totally empty lines . I have recently , when I define or modify words , started including dates referring to the days in my log on which I worked on them . Thus additional info on motivation of development and examples it available in those logs .
  
  But here's the scrubbing I spent more time on than warranted
  

   s" ./CoSy/CoSy.f" F> "nl VM >t0> rho     |>| 1733    
  | Read file , split on "cr "lf
   t0 dae >t0> rho     |>| 1375     | delete all empty lines |
  | eliminate full line comments .
   t0 t0 { 0 _at s" |" <>i } 'm ,/ & at >t0> rho         |>| 1176
  | delete all lines which are totally whitespace
   t0 t0 { c>i 32 _i >i ['] maxi ./ } 'm ,/ & at rho     |>| 1060
  

• 08:20 | I've moved the meta pointer ( 4th cell in header ) referencing words to more where they belong in sequence . Sam was right to totally skip over them . I know they are extremely useful from my experience with K's attributes , but have have only now , working on multiple windows , started finishing implementing them .
  
• 10:15 | Modulo indexing is one innovation I'm particularly happy with . It eliminates , generalizes APL's scalar extension . ( But singletons still gets some special treatment , disclosing results of some functions . )
  Interestingly it actually makes CoSy's data structure loops of loops rather than lists of lists .
• 12:20 | Basing reference counting at 0 rather than 1 may optimize some testing , but is a minor matter . I would like to see the alternative .
  
• 13:50 | vdup is actually used in only 2 places : eachMcr where it could be easily replaced , and lower which converts a string to lower case and , again simply does what essentially needs to be done .
  More important and far more recent is rep which replicates entire nested lists , and , quite recently  : .. dup rep ;  which performs the function of dup in making totally independent copies of ref counted objects . It's a great simplifer .
• 16:50 | Thanks for recognizing the subtly recursive nature of ref- . It , and the save and restore functions in  SaveRestore.f which convert the allocated nested list structures back and forth to a linear file format , are the crux vocabulary .
• 18:45 | Starting with the each functions , we get into the area which is the result of bushwhacking which needs significant pruning .  The top level each functions are
  'm 'd 'R 'L for monadic dyadic eachRight eachLeft . Note that because of Forth's simple RPN syntax and and ' which returns the address of a word , whether noun or verb , it is quite easy to implement what in APL are called operators . I found it interesting that what APL implemented as commute , ⍨ , originally , I believe from John Scholes , is simply swap in Forth .
  
• 21:40 | If there is a better way to allocate space for initializing lists of unknown length , I'm all ears . I was really interested to see Sam's use of here as the base for temporaries in his over the shoulder demo . I think much of my code might be different if I really knew Forth as a Forther before I began .
  With a 2^n slot inventory memory allocation , a concatenating method may be efficient .
  
• 22:48 | Yes the abandonment of the notion of rectangularity which K introduced , and with it the default application of functions along the most nested dimension greatly reduces complexity . Interestingly , the primacy of the leading axis is infiltrating APL with the implementation of tally , ≢ .
• 24:15 I had and and or associated with min and max backward -- as I found out in the all whitespace line deletion above . I don't know where a colon followed by a number is .
  
• 26:00 Symbols may be one type to far .  at the Minnowbrook 2011 APL implementers conference Morton Kromberg who runs Dyalog APL asked me what is a symbol other than a string . Functionally it's just a word . I made a separate type out of it with a count of 1 rather than the number of characters . I find that more commonly a pain than useful .  I may discard the type .
  Similarly , Arthur has a dictionary type in K . In CoSy I have almost exactly copied his structure as a list of 2 lists of symbols and associated values , and a vocabulary essentially supplying name indexed lists ( which Sam didn't get to ) . I haven't made them a separate type and have never run into a problem . Back in 1998 I wrote about types just being shortcuts for determining the ding an sich .
  But they are a powerful to the point of essential shortcut in many cases .
• 27:30 | I think of iota as generating the index of an address space . When I first met APL back in the '70s , I thought ⍳ was the strangest thing in a computer language . And a language evolved from matrix algebra which didn't even have a verb for creating an identity matrix .
  
• 28:10 | A bunch of that Type stuff needs to be thought out better and cleaned up . The most fundamental are essential , but I've laid out some which are yet , if ever , to be used .
  One important point which did not get mentioned is that a general list , ie: a list of lists , ie:  a list of pointers to CoSy objects , is type 0 . So , since type is the first cell in the header , you can simply fetch & test it .
• 30:45 | Yes . Good description of apply operators .  They are what allow for instance a dot product to be applied to data in , eg: each voxel of a map of the atmosphere .
  I would phrase it as leafs of the lists because the leafs themselves are generally simple lists of numbers or characters .
• 31:35 | I'm cleaning up scan right now . I plan to be presenting a bit of APL and Forth and CoSy to math class when I'm at my 55th Culver Military Academy reunion next month . Arthur Whitney , probably following Iverson , has no compunctions simply calling scan integration .
• I would strongly suggest a type field in the header . Without it , you cannot have any sort of generic verbs , nor , for that matter , any sort of general , ie: nested , lists . But with 64 bit cells a whole cell is rather overkill -- as it is for most of the items . Looks like you are providing for max 256 bit items . I'm definitely interested in looking at it in detail .

Thank you Sam . And for your kind comments . I think your presentation provides a great resource for anyone interested in CoSy and branching it in the direction of their interests .

The distillation of the dynamic array vocabulary of APL in the open simplicity of Forth , I am continually discovering to be more powerful , and far more flexible than either .