• Re: A much shorter proof that the Halting Problem is a category error.

    From Mikko@mikko.levanto@iki.fi to comp.theory on Thu Oct 30 13:06:29 2025
    From Newsgroup: comp.theory

    On 2025-10-29 16:32:16 +0000, olcott said:

    On 10/29/2025 5:04 AM, Mikko wrote:
    On 2025-10-28 14:56:16 +0000, olcott said:

    On 10/28/2025 4:18 AM, Mikko wrote:
    On 2025-10-27 13:47:17 +0000, olcott said:

    On 10/27/2025 4:39 AM, Mikko wrote:
    On 2025-10-26 14:46:33 +0000, olcott said:

    Summary of the key point:

    The halting problem's self-referential construction creates two >>>>>>> distinct computational entities:

    The halting problem does refer itself nor requires that any entity >>>>>> should refer to itself. At most the descriptions of the program and >>>>>> can be seen as references to the program and input asked about, and >>>>>> some formulations of the problem don't mention even that. The program >>>>>> asked about does not refer to anything outside itself and the input >>>>>> asked about. No semantics of the input asked about is relevant fo the >>>>>> halting problem so in that context nothing in the input refers to
    anything.

    Input (DD-as-simulated-by-HHH): Shows non-halting behavior - recursive >>>>>>> pattern that HHH correctly identifies
    Non-input (DD-as-directly-executed): Halts because HHH returns 0

    Not relevant to the halting problem,

    Saying that the halting problem proof counter-example
    input is not relevant to the halting problem is dishonest.

    More importantly, an attempt of a straw man deception dishonest.
    Less importantly, an attempt to distract with an irrelevancy is
    dishonest, too.


    int D()
    {
       int Halt_Status = H(D);
       if (Halt_Status)
         HERE: goto HERE;
       return Halt_Status;
    }

    H simulates D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    until H sees this repeating pattern.

    When simulating halt decider H is reporting on the
    behavior that its input specifies then H is correct
    to reject D as non-halting.

    Deciders only compute a mapping from their actual
    inputs. Computing the mapping from non-inputs is
    outside of the scope of Turing machines.

    Nice to see that you don't disagree.

    So you have agreed that I proved the halting problem
    itself is a category error on the basis that it requires
    halt deciders to report on behavior other than the behavior
    specified by their inputs.

    No, it is not my style to agree with obvious falsehoods. But I do find
    it nice that you don't disagree with my opinion that you are dishones
    when you try to dstract reader's attention from truth with an irrevancy.
    --
    Mikko

    --- Synchronet 3.21a-Linux NewsLink 1.2
  • From olcott@polcott333@gmail.com to comp.theory on Thu Oct 30 08:38:17 2025
    From Newsgroup: comp.theory

    On 10/30/2025 6:06 AM, Mikko wrote:
    On 2025-10-29 16:32:16 +0000, olcott said:

    On 10/29/2025 5:04 AM, Mikko wrote:
    On 2025-10-28 14:56:16 +0000, olcott said:

    On 10/28/2025 4:18 AM, Mikko wrote:
    On 2025-10-27 13:47:17 +0000, olcott said:

    On 10/27/2025 4:39 AM, Mikko wrote:
    On 2025-10-26 14:46:33 +0000, olcott said:

    Summary of the key point:

    The halting problem's self-referential construction creates two >>>>>>>> distinct computational entities:

    The halting problem does refer itself nor requires that any entity >>>>>>> should refer to itself. At most the descriptions of the program and >>>>>>> can be seen as references to the program and input asked about, and >>>>>>> some formulations of the problem don't mention even that. The
    program
    asked about does not refer to anything outside itself and the input >>>>>>> asked about. No semantics of the input asked about is relevant fo >>>>>>> the
    halting problem so in that context nothing in the input refers to >>>>>>> anything.

    Input (DD-as-simulated-by-HHH): Shows non-halting behavior -
    recursive pattern that HHH correctly identifies
    Non-input (DD-as-directly-executed): Halts because HHH returns 0 >>>>>>>
    Not relevant to the halting problem,

    Saying that the halting problem proof counter-example
    input is not relevant to the halting problem is dishonest.

    More importantly, an attempt of a straw man deception dishonest.
    Less importantly, an attempt to distract with an irrelevancy is
    dishonest, too.


    int D()
    {
       int Halt_Status = H(D);
       if (Halt_Status)
         HERE: goto HERE;
       return Halt_Status;
    }

    H simulates D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    until H sees this repeating pattern.

    When simulating halt decider H is reporting on the
    behavior that its input specifies then H is correct
    to reject D as non-halting.

    Deciders only compute a mapping from their actual
    inputs. Computing the mapping from non-inputs is
    outside of the scope of Turing machines.

    Nice to see that you don't disagree.

    So you have agreed that I proved the halting problem
    itself is a category error on the basis that it requires
    halt deciders to report on behavior other than the behavior
    specified by their inputs.

    No, it is not my style to agree with obvious falsehoods. But I do find
    it nice that you don't disagree with my opinion that you are dishones
    when you try to dstract reader's attention from truth with an irrevancy.



    D simulated by H measures the semantic property of
    the actual input as opposed to and contrast with
    the semantic property of a non-input.

    We can tell an input from an input because an
    input is an argument to the function H.

    D.input_to_H
    specifies different behavior than
    D.input_to_H1.

    int D()
    {
    int Halt_Status = H(D);
    if (Halt_Status)
    HERE: goto HERE;
    return Halt_Status;
    }

    H simulates D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    until H sees this repeating pattern
    Then H returns 0;

    H1 simulates D
    that calls H(D) to simulate D
    then H(D) returns 0 to caller D
    then H1 returns 1;

    int sum(int x, int y){ return x + y;}
    int product(int x, int y){ return x * y;}

    // same input different results
    sum(3,4) != product(3,4)
    --
    Copyright 2025 Olcott "Talent hits a target no one else can hit; Genius
    hits a target no one else can see." Arthur Schopenhauer
    --- Synchronet 3.21a-Linux NewsLink 1.2
  • From Mikko@mikko.levanto@iki.fi to comp.theory on Fri Oct 31 13:40:33 2025
    From Newsgroup: comp.theory

    On 2025-10-30 13:38:17 +0000, olcott said:

    On 10/30/2025 6:06 AM, Mikko wrote:
    On 2025-10-29 16:32:16 +0000, olcott said:

    On 10/29/2025 5:04 AM, Mikko wrote:
    On 2025-10-28 14:56:16 +0000, olcott said:

    On 10/28/2025 4:18 AM, Mikko wrote:
    On 2025-10-27 13:47:17 +0000, olcott said:

    On 10/27/2025 4:39 AM, Mikko wrote:
    On 2025-10-26 14:46:33 +0000, olcott said:

    Summary of the key point:

    The halting problem's self-referential construction creates two >>>>>>>>> distinct computational entities:

    The halting problem does refer itself nor requires that any entity >>>>>>>> should refer to itself. At most the descriptions of the program and >>>>>>>> can be seen as references to the program and input asked about, and >>>>>>>> some formulations of the problem don't mention even that. The program >>>>>>>> asked about does not refer to anything outside itself and the input >>>>>>>> asked about. No semantics of the input asked about is relevant fo the >>>>>>>> halting problem so in that context nothing in the input refers to >>>>>>>> anything.

    Input (DD-as-simulated-by-HHH): Shows non-halting behavior - recursive
    pattern that HHH correctly identifies
    Non-input (DD-as-directly-executed): Halts because HHH returns 0 >>>>>>>>
    Not relevant to the halting problem,

    Saying that the halting problem proof counter-example
    input is not relevant to the halting problem is dishonest.

    More importantly, an attempt of a straw man deception dishonest.
    Less importantly, an attempt to distract with an irrelevancy is
    dishonest, too.


    int D()
    {
       int Halt_Status = H(D);
       if (Halt_Status)
         HERE: goto HERE;
       return Halt_Status;
    }

    H simulates D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    until H sees this repeating pattern.

    When simulating halt decider H is reporting on the
    behavior that its input specifies then H is correct
    to reject D as non-halting.

    Deciders only compute a mapping from their actual
    inputs. Computing the mapping from non-inputs is
    outside of the scope of Turing machines.

    Nice to see that you don't disagree.

    So you have agreed that I proved the halting problem
    itself is a category error on the basis that it requires
    halt deciders to report on behavior other than the behavior
    specified by their inputs.

    No, it is not my style to agree with obvious falsehoods. But I do find
    it nice that you don't disagree with my opinion that you are dishones
    when you try to dstract reader's attention from truth with an irrevancy.

    D simulated by H measures the semantic property of
    the actual input as opposed to and contrast with
    the semantic property of a non-input.

    We can tell an input from an input because an
    input is an argument to the function H.

    D.input_to_H
    specifies different behavior than
    D.input_to_H1.

    int D()
    {
    int Halt_Status = H(D);
    if (Halt_Status)
    HERE: goto HERE;
    return Halt_Status;
    }

    H simulates D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    until H sees this repeating pattern
    Then H returns 0;

    H1 simulates D
    that calls H(D) to simulate D
    then H(D) returns 0 to caller D
    then H1 returns 1;

    int sum(int x, int y){ return x + y;}
    int product(int x, int y){ return x * y;}

    // same input different results
    sum(3,4) != product(3,4)

    Thank you for an example of an attempt to dever attention in a way
    that I called dishonest and you din't disagree.
    --
    Mikko

    --- Synchronet 3.21a-Linux NewsLink 1.2
  • From olcott@polcott333@gmail.com to comp.theory on Fri Oct 31 07:50:30 2025
    From Newsgroup: comp.theory

    On 10/31/2025 6:40 AM, Mikko wrote:
    On 2025-10-30 13:38:17 +0000, olcott said:

    On 10/30/2025 6:06 AM, Mikko wrote:
    On 2025-10-29 16:32:16 +0000, olcott said:

    On 10/29/2025 5:04 AM, Mikko wrote:
    On 2025-10-28 14:56:16 +0000, olcott said:

    On 10/28/2025 4:18 AM, Mikko wrote:
    On 2025-10-27 13:47:17 +0000, olcott said:

    On 10/27/2025 4:39 AM, Mikko wrote:
    On 2025-10-26 14:46:33 +0000, olcott said:

    Summary of the key point:

    The halting problem's self-referential construction creates >>>>>>>>>> two distinct computational entities:

    The halting problem does refer itself nor requires that any entity >>>>>>>>> should refer to itself. At most the descriptions of the program >>>>>>>>> and
    can be seen as references to the program and input asked about, >>>>>>>>> and
    some formulations of the problem don't mention even that. The >>>>>>>>> program
    asked about does not refer to anything outside itself and the >>>>>>>>> input
    asked about. No semantics of the input asked about is relevant >>>>>>>>> fo the
    halting problem so in that context nothing in the input refers to >>>>>>>>> anything.

    Input (DD-as-simulated-by-HHH): Shows non-halting behavior - >>>>>>>>>> recursive pattern that HHH correctly identifies
    Non-input (DD-as-directly-executed): Halts because HHH returns 0 >>>>>>>>>
    Not relevant to the halting problem,

    Saying that the halting problem proof counter-example
    input is not relevant to the halting problem is dishonest.

    More importantly, an attempt of a straw man deception dishonest. >>>>>>> Less importantly, an attempt to distract with an irrelevancy is
    dishonest, too.


    int D()
    {
       int Halt_Status = H(D);
       if (Halt_Status)
         HERE: goto HERE;
       return Halt_Status;
    }

    H simulates D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    until H sees this repeating pattern.

    When simulating halt decider H is reporting on the
    behavior that its input specifies then H is correct
    to reject D as non-halting.

    Deciders only compute a mapping from their actual
    inputs. Computing the mapping from non-inputs is
    outside of the scope of Turing machines.

    Nice to see that you don't disagree.

    So you have agreed that I proved the halting problem
    itself is a category error on the basis that it requires
    halt deciders to report on behavior other than the behavior
    specified by their inputs.

    No, it is not my style to agree with obvious falsehoods. But I do find
    it nice that you don't disagree with my opinion that you are dishones
    when you try to dstract reader's attention from truth with an irrevancy.

    D simulated by H measures the semantic property of
    the actual input as opposed to and contrast with
    the semantic property of a non-input.

    We can tell an input from an input because an
    input is an argument to the function H.

    D.input_to_H
    specifies different behavior than
    D.input_to_H1.

    int D()
    {
       int Halt_Status = H(D);
       if (Halt_Status)
         HERE: goto HERE;
       return Halt_Status;
    }

    H simulates D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    until H sees this repeating pattern
    Then H returns 0;

    H1 simulates D
    that calls H(D) to simulate D
    then H(D) returns 0 to caller D
    then H1 returns 1;

    int sum(int x, int y){ return x + y;}
    int product(int x, int y){ return x * y;}

    // same input different results
    sum(3,4) != product(3,4)

    Thank you for an example of an attempt to dever attention in a way
    that I called dishonest and you din't disagree.


    It has been a verified fact for three years.
    What you call dishonest is your own lack of
    comprehension. It looks like I will be able
    to rewrite it using a C interpreter.
    --
    Copyright 2025 Olcott "Talent hits a target no one else can hit; Genius
    hits a target no one else can see." Arthur Schopenhauer
    --- Synchronet 3.21a-Linux NewsLink 1.2
  • From Mikko@mikko.levanto@iki.fi to comp.theory on Sat Nov 1 11:56:49 2025
    From Newsgroup: comp.theory

    On 2025-10-31 12:50:30 +0000, olcott said:

    On 10/31/2025 6:40 AM, Mikko wrote:
    On 2025-10-30 13:38:17 +0000, olcott said:

    On 10/30/2025 6:06 AM, Mikko wrote:
    On 2025-10-29 16:32:16 +0000, olcott said:

    On 10/29/2025 5:04 AM, Mikko wrote:
    On 2025-10-28 14:56:16 +0000, olcott said:

    On 10/28/2025 4:18 AM, Mikko wrote:
    On 2025-10-27 13:47:17 +0000, olcott said:

    On 10/27/2025 4:39 AM, Mikko wrote:
    On 2025-10-26 14:46:33 +0000, olcott said:

    Summary of the key point:

    The halting problem's self-referential construction creates two >>>>>>>>>>> distinct computational entities:

    The halting problem does refer itself nor requires that any entity >>>>>>>>>> should refer to itself. At most the descriptions of the program and >>>>>>>>>> can be seen as references to the program and input asked about, and >>>>>>>>>> some formulations of the problem don't mention even that. The program
    asked about does not refer to anything outside itself and the input >>>>>>>>>> asked about. No semantics of the input asked about is relevant fo the
    halting problem so in that context nothing in the input refers to >>>>>>>>>> anything.

    Input (DD-as-simulated-by-HHH): Shows non-halting behavior - recursive
    pattern that HHH correctly identifies
    Non-input (DD-as-directly-executed): Halts because HHH returns 0 >>>>>>>>>>
    Not relevant to the halting problem,

    Saying that the halting problem proof counter-example
    input is not relevant to the halting problem is dishonest.

    More importantly, an attempt of a straw man deception dishonest. >>>>>>>> Less importantly, an attempt to distract with an irrelevancy is >>>>>>>> dishonest, too.


    int D()
    {
       int Halt_Status = H(D);
       if (Halt_Status)
         HERE: goto HERE;
       return Halt_Status;
    }

    H simulates D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    until H sees this repeating pattern.

    When simulating halt decider H is reporting on the
    behavior that its input specifies then H is correct
    to reject D as non-halting.

    Deciders only compute a mapping from their actual
    inputs. Computing the mapping from non-inputs is
    outside of the scope of Turing machines.

    Nice to see that you don't disagree.

    So you have agreed that I proved the halting problem
    itself is a category error on the basis that it requires
    halt deciders to report on behavior other than the behavior
    specified by their inputs.

    No, it is not my style to agree with obvious falsehoods. But I do find >>>> it nice that you don't disagree with my opinion that you are dishones
    when you try to dstract reader's attention from truth with an irrevancy. >>>
    D simulated by H measures the semantic property of
    the actual input as opposed to and contrast with
    the semantic property of a non-input.

    We can tell an input from an input because an
    input is an argument to the function H.

    D.input_to_H
    specifies different behavior than
    D.input_to_H1.

    int D()
    {
       int Halt_Status = H(D);
       if (Halt_Status)
         HERE: goto HERE;
       return Halt_Status;
    }

    H simulates D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    until H sees this repeating pattern
    Then H returns 0;

    H1 simulates D
    that calls H(D) to simulate D
    then H(D) returns 0 to caller D
    then H1 returns 1;

    int sum(int x, int y){ return x + y;}
    int product(int x, int y){ return x * y;}

    // same input different results
    sum(3,4) != product(3,4)

    Thank you for an example of an attempt to dever attention in a way
    that I called dishonest and you din't disagree.

    It has been a verified fact for three years.
    What you call dishonest is your own lack of
    comprehension. It looks like I will be able
    to rewrite it using a C interpreter.

    A Turing machine interpreter would be easier to do and would be
    more erelvant to the topic of Turing computability. In C there
    is too much unnecessary complesity and ambiguity.
    --
    Mikko

    --- Synchronet 3.21a-Linux NewsLink 1.2
  • From olcott@polcott333@gmail.com to comp.theory on Sat Nov 1 08:43:46 2025
    From Newsgroup: comp.theory

    On 11/1/2025 4:56 AM, Mikko wrote:
    On 2025-10-31 12:50:30 +0000, olcott said:

    On 10/31/2025 6:40 AM, Mikko wrote:
    On 2025-10-30 13:38:17 +0000, olcott said:

    On 10/30/2025 6:06 AM, Mikko wrote:
    On 2025-10-29 16:32:16 +0000, olcott said:

    On 10/29/2025 5:04 AM, Mikko wrote:
    On 2025-10-28 14:56:16 +0000, olcott said:

    On 10/28/2025 4:18 AM, Mikko wrote:
    On 2025-10-27 13:47:17 +0000, olcott said:

    On 10/27/2025 4:39 AM, Mikko wrote:
    On 2025-10-26 14:46:33 +0000, olcott said:

    Summary of the key point:

    The halting problem's self-referential construction creates >>>>>>>>>>>> two distinct computational entities:

    The halting problem does refer itself nor requires that any >>>>>>>>>>> entity
    should refer to itself. At most the descriptions of the >>>>>>>>>>> program and
    can be seen as references to the program and input asked >>>>>>>>>>> about, and
    some formulations of the problem don't mention even that. The >>>>>>>>>>> program
    asked about does not refer to anything outside itself and the >>>>>>>>>>> input
    asked about. No semantics of the input asked about is
    relevant fo the
    halting problem so in that context nothing in the input >>>>>>>>>>> refers to
    anything.

    Input (DD-as-simulated-by-HHH): Shows non-halting behavior - >>>>>>>>>>>> recursive pattern that HHH correctly identifies
    Non-input (DD-as-directly-executed): Halts because HHH >>>>>>>>>>>> returns 0

    Not relevant to the halting problem,

    Saying that the halting problem proof counter-example
    input is not relevant to the halting problem is dishonest.

    More importantly, an attempt of a straw man deception dishonest. >>>>>>>>> Less importantly, an attempt to distract with an irrelevancy is >>>>>>>>> dishonest, too.


    int D()
    {
       int Halt_Status = H(D);
       if (Halt_Status)
         HERE: goto HERE;
       return Halt_Status;
    }

    H simulates D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    until H sees this repeating pattern.

    When simulating halt decider H is reporting on the
    behavior that its input specifies then H is correct
    to reject D as non-halting.

    Deciders only compute a mapping from their actual
    inputs. Computing the mapping from non-inputs is
    outside of the scope of Turing machines.

    Nice to see that you don't disagree.

    So you have agreed that I proved the halting problem
    itself is a category error on the basis that it requires
    halt deciders to report on behavior other than the behavior
    specified by their inputs.

    No, it is not my style to agree with obvious falsehoods. But I do find >>>>> it nice that you don't disagree with my opinion that you are dishones >>>>> when you try to dstract reader's attention from truth with an
    irrevancy.

    D simulated by H measures the semantic property of
    the actual input as opposed to and contrast with
    the semantic property of a non-input.

    We can tell an input from an input because an
    input is an argument to the function H.

    D.input_to_H
    specifies different behavior than
    D.input_to_H1.

    int D()
    {
       int Halt_Status = H(D);
       if (Halt_Status)
         HERE: goto HERE;
       return Halt_Status;
    }

    H simulates D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    until H sees this repeating pattern
    Then H returns 0;

    H1 simulates D
    that calls H(D) to simulate D
    then H(D) returns 0 to caller D
    then H1 returns 1;

    int sum(int x, int y){ return x + y;}
    int product(int x, int y){ return x * y;}

    // same input different results
    sum(3,4) != product(3,4)

    Thank you for an example of an attempt to dever attention in a way
    that I called dishonest and you din't disagree.

    It has been a verified fact for three years.
    What you call dishonest is your own lack of
    comprehension. It looks like I will be able
    to rewrite it using a C interpreter.

    A Turing machine interpreter would

    Takes a million steps just to move data from one
    memory location to another.

    I could implement The Linz template at that same
    level of abstraction quite easily.

    be easier to do and would be
    more erelvant to the topic of Turing computability. In C there
    is too much unnecessary complesity and ambiguity.


    Looking for a bug in an operating system by carefully
    examining the billions of lines of microcode. Or we
    could look at the same thing at the C level.
    --
    Copyright 2025 Olcott "Talent hits a target no one else can hit; Genius
    hits a target no one else can see." Arthur Schopenhauer
    --- Synchronet 3.21a-Linux NewsLink 1.2
  • From Mikko@mikko.levanto@iki.fi to comp.theory on Sun Nov 2 14:48:07 2025
    From Newsgroup: comp.theory

    On 2025-11-01 13:43:46 +0000, olcott said:

    On 11/1/2025 4:56 AM, Mikko wrote:
    On 2025-10-31 12:50:30 +0000, olcott said:

    On 10/31/2025 6:40 AM, Mikko wrote:
    On 2025-10-30 13:38:17 +0000, olcott said:

    On 10/30/2025 6:06 AM, Mikko wrote:
    On 2025-10-29 16:32:16 +0000, olcott said:

    On 10/29/2025 5:04 AM, Mikko wrote:
    On 2025-10-28 14:56:16 +0000, olcott said:

    On 10/28/2025 4:18 AM, Mikko wrote:
    On 2025-10-27 13:47:17 +0000, olcott said:

    On 10/27/2025 4:39 AM, Mikko wrote:
    On 2025-10-26 14:46:33 +0000, olcott said:

    Summary of the key point:

    The halting problem's self-referential construction creates two >>>>>>>>>>>>> distinct computational entities:

    The halting problem does refer itself nor requires that any entity >>>>>>>>>>>> should refer to itself. At most the descriptions of the program and
    can be seen as references to the program and input asked about, and
    some formulations of the problem don't mention even that. The program
    asked about does not refer to anything outside itself and the input
    asked about. No semantics of the input asked about is relevant fo the
    halting problem so in that context nothing in the input refers to >>>>>>>>>>>> anything.

    Input (DD-as-simulated-by-HHH): Shows non-halting behavior - recursive
    pattern that HHH correctly identifies
    Non-input (DD-as-directly-executed): Halts because HHH returns 0 >>>>>>>>>>>>
    Not relevant to the halting problem,

    Saying that the halting problem proof counter-example
    input is not relevant to the halting problem is dishonest. >>>>>>>>>>
    More importantly, an attempt of a straw man deception dishonest. >>>>>>>>>> Less importantly, an attempt to distract with an irrelevancy is >>>>>>>>>> dishonest, too.


    int D()
    {
       int Halt_Status = H(D);
       if (Halt_Status)
         HERE: goto HERE;
       return Halt_Status;
    }

    H simulates D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    until H sees this repeating pattern.

    When simulating halt decider H is reporting on the
    behavior that its input specifies then H is correct
    to reject D as non-halting.

    Deciders only compute a mapping from their actual
    inputs. Computing the mapping from non-inputs is
    outside of the scope of Turing machines.

    Nice to see that you don't disagree.

    So you have agreed that I proved the halting problem
    itself is a category error on the basis that it requires
    halt deciders to report on behavior other than the behavior
    specified by their inputs.

    No, it is not my style to agree with obvious falsehoods. But I do find >>>>>> it nice that you don't disagree with my opinion that you are dishones >>>>>> when you try to dstract reader's attention from truth with an irrevancy. >>>>>
    D simulated by H measures the semantic property of
    the actual input as opposed to and contrast with
    the semantic property of a non-input.

    We can tell an input from an input because an
    input is an argument to the function H.

    D.input_to_H
    specifies different behavior than
    D.input_to_H1.

    int D()
    {
       int Halt_Status = H(D);
       if (Halt_Status)
         HERE: goto HERE;
       return Halt_Status;
    }

    H simulates D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    that calls H(D) to simulate D
    until H sees this repeating pattern
    Then H returns 0;

    H1 simulates D
    that calls H(D) to simulate D
    then H(D) returns 0 to caller D
    then H1 returns 1;

    int sum(int x, int y){ return x + y;}
    int product(int x, int y){ return x * y;}

    // same input different results
    sum(3,4) != product(3,4)

    Thank you for an example of an attempt to dever attention in a way
    that I called dishonest and you din't disagree.

    It has been a verified fact for three years.
    What you call dishonest is your own lack of
    comprehension. It looks like I will be able
    to rewrite it using a C interpreter.

    A Turing machine interpreter would

    Takes a million steps just to move data from one
    memory location to another.

    Which could be simulated in a few senconds with a modern computer,
    or even less with a fast one.
    --
    Mikko

    --- Synchronet 3.21a-Linux NewsLink 1.2