What is "Colossus"?

Colossus is the program which was run on the Apollo Guidance Computer (AGC) installed in the Command Module (CM) for Apollo 8 through 17.  In order to be loaded into the AGC—actually, to be converted to the "core ropes" within the AGC — it was orginally necessary to "assemble" Colossus's source code into binary machine language, using a computer program called YUL or its successor program, GAP.  In the context of the Virtual AGC project, of course, the process is somewhat different:  Colossus source code is assembled by the yaYUL program, and then loaded into the yaAGC simulation.

It's Always Sunny in Outer Space

AGC software family treePrograms with names like Sundisk, Solarium, and Sunspot were pre-Colossus versions of the Command Module AGC's software.  They were not forms of Colossus as such, but provided some code for Colossus.  If you click on the image to the right, you can see a simplified family tree for Colossus.

The Colossus program eventually had three major versions, not surprisingly known as Colossus 1 (COLOSSUS), Colossus 2 (COMANCHE), and Colossus 3 (ARTEMIS).  Additionally, there were minor versions such as Colossus 1A, Colossus 2C, and so on.

Finally, SKYLARK was post-Colossus Command-Module software, but it was simply adapted from ARTEMIS.

And that doesn't even begin to cover the varieties of software developed for the AGC installed in the Lunar Module (LM), which you can read about on our Luminary page.

If you'd like to see a more comprehensive but less CM-centric overview of all the available AGC software source code and the manufactured core ropes than presented on this page, you might want to look in the section of our Document Library dedicated to AGC software listings.  The diagram to the right (click to expand) attempts to show the interrelationships between every AGC software version (Luminary, Colossus, or otherwise) and every AGS software version, along with the documentation (Program Change Requests, memos, etc.) documenting the changes.  It covers literally thousands of versions and is therefore very big!

Of course, the command-modules themselves are still in existence, though that doesn't imply they still contain their AGCs ... and most likely none of them do.  Zach Greene has compiled the following partial list:

Available and Particularly-Desired Command Module Software Versions

Mission CSM number
Mission Type
CM Program Revision Source Code
Other Mission-Specific Documentation
Syntax-highlighted, hyperlinked HTML

Scanned page images
This is an unusual program, in that its sole purpose was to demonstrate usage of the original AGC assembler, YUL.  In fact, its very short program listing (only one page of source code!) that was attached to a 1963 manual for YUL.  The document itself was obtained from the personal papers of the late Russel Larson.

Syntax-highlighted, hyperlinked HTML
Apollo Mission 204 Reference Cards

Keyboard and Display System Program for AGC (Program SUNRISE)

The Compleat SUNRISE

AGCIS #12A: Job Control and Task Control

AGCIS #15: Block I Apollo Guidance Computer Subsystem

AGCIS #16: Progress Control and Fresh Start and Restart


AGCIS #22: Prelaunch Alignment

SUNRISE 69 and Erasable Memory Test Program
SUNRISE is a program used for Block I system tests, providing a similar service, I guess, to what the SUNDIAL program later provided for Block II, and which AURORA provided for the LM.  We acquired SUNRISE 45 by dumping the contents of three physical rope-memory modules to which an anonymous collector gave us access.

Now, it often happens that physical rope-memory modules are physically defective in one way or another, either as a result of sitting around for 50-60 years unused, or perhaps due to defective modules being more likely to have been discarded and thus more likely to have found their way into collectors' collections in the first place.  Usually, due to fact that these memory modules have a 16th (parity) bit in addition to the 15 data bits at each memory location, these flaws are relatively easy to correct.  The SUNRISE 45 rope is notable in that one of the three memory modules was rather more severely damaged than usual, but that Mike Stewart was able to deduce corrections for it anyway.  We had a high degree of confidence in those corrections ... but still, they were unusually difficult to make, and we'd have even more confidence in the dumped data if those corrections hadn't needed to be made in the first place.

Syntax-highlighted, hyperlinked HTML The same comments I made for SUNRISE 45 above basically hold equally well for SUNRISE 69, with the difference that the that SUNRISE 69 spans four rope-memory modules rather than just three.  What's really interesting, however, is that the four memory modules are the same three from SUNRISE 45, plus one extra.

If you think about it, that's pretty remarkable.  After all, if the three SUNRISE 45 modules never access a 4th rope module, then how would the code in the 4th rope module ever be reached when you're running SUNRISE 69?  The answer is apparently (and obviously, I admit) that there are verb/noun combinations that you can use in SUNRISE 45 to transfer control to an address in the extra rope-memory module, which is designated B22.  (The first three modules are B28, B29, and B21.)  For example, the SUNRISE 69 DSKY electroluminescent test is kind of fun, and it's started by the rather cumbersome command sequence,
Whereas if you were to attempt the same thing in SUNRISE 45, it doesn't initiate the test.  Rather, control is lost and all further key entries on the DSKY seem to be ignored.  I'd expect that memory-parity errors would be encountered if this happened on a physical AGC, and that it would cause a system reset; since that doesn't happen virtually, I must not support parity checking in the CPU emulator.  It sounds like the kind of thing I'd do.

Apollo 3
Syntax-highlighted, hyperlinked HTML Document Library
Corona 261 is the Block I AGC software for mission AS-202.  Quoting from the wikipedia article:
"AS-202 (also referred to as SA-202) was the second unmanned, suborbital test flight of a production Block I Apollo Command/Service Module launched with the Saturn IB launch vehicle. It was launched on August 25, 1966 and was the first flight which included the spacecraft Guidance and Navigation Control system and fuel cells. The success of this flight enabled the Apollo program to judge the Block I spacecraft and Saturn IB ready to carry men into orbit on the next mission, AS-204."
In other words, this is the first AGC software actually flown, and we can "fly" it too, at least in the sense of using it in the Orbiter spaceflight simulation system with the NASSP add-on.

Apollo 1
On our wish list
Document Library
Yes, Apollo 1 never flew, but it certainly deserves a place of honor out of respect for its crew.  I'd sure like to know where this program can be found!

This was a Block I spacecraft.

Apollo 4

Apollo 6



Scanned page images
Document Library (Apollo 4)

Document Library (Apollo 6)
Apollo 4 and Apollo 6 were unmanned missions, but it they did have working CSMs and working AGCs in the CMs.  Eldon C. Hall's copy of the AGC program listing is in the American Computer Museum, which has graciously allowed us to digitize the program listing for use in Virtual AGC. 

These were Block I spacecraft.

The "Programer"—yes, that's the real spelling, though some sources do refer to it as the "programmer"—was a gadget that was the stand-in for the crew.

Core rope dump

Syntax-highlighted, hyperlinked HTML
Document Library
There's a nice online writeup of the 2TV-1 mission that I won't paraphrase here, except to say that it comprised four substantial (177 hours), crewed missions that occurred entirely on the ground in June 1968.

"2TV-1" stands for Block II Thermal Vacuum no. 1, and more-or-less describes the missions, which were to subject the spacecraft to vacuum and to thermal shock whilst the astronauts did astronauty things.
Apollo 7
On our wish list
Document Library
Unfortunately, we do not presently have this software, which flew on Apollo 7.

For further insights, note document E-2150, "Guidance, Navigation, and Control Block II Command and Service Module Functional Description and Operation Using Flight Program SUNDISK (Rev. 282)", whose preface reads: "The purpose of this document is twofold.  The first is to provide a functional description (operationally oriented) of the CSM GNCS hardware and software and the interfaces with other SC systems.  The level of detail is that required to identify and define telemetry outputs. Also included are functional flow diagrams of the Sundisk 282 programs and routines together with lists of verbs, nouns, option codes, and checklist codes for this flow. The second purpose is to provide the operational procedures for this hardware and software including malfunction procedures, and program notes.  The nominal airborne expanded and condensed checklist for Sundisk appear in the Guidance System Operation Plan, R547."
Apollo 8
Colossus 1

Scanned page images
Document Library
Colossus 237 flew on Apollo 8.

The Colossus 237 program listing was made available by original AGC developer Fred Martin.  (Thanks, Fred!)  The listing cuts off abruptly after page 1557—whereas you'd normally anticipate that there are 1700+ pages—so some of the assembler-generated tables at the end of the listing are missing.  Of particular inconvenience, because it presents special problems for verifying the correctness of the simulation, is the absence of the octal listing and of the memory-bank checksums.  However, all of the source code is present, and that's enough to work with.

Note that we have a contemporary digital simulation of Colossus 237, though it only covers a bailout during burn (Colossus Software Anomaly Report #45, of which we unfortunately don't presently have a copy).  By "contemporary", I mean that it's an Apollo-era simulation, and not one produced by the Virtual AGC Project.  It is the only Colossus digital simulation available so far, although there are a number of Luminary digital simulations in our library, and as such it has a unique value.  Specifically, it has pad-loads, though those pad-loads are for March 1969 rather than for December 1968.
Apollo 9
Colossus 1A
Syntax-highlighted, hyperlinked,  HTML

The original raw scans, unprocessed
Document Library
Two separate sets of scanned images of the Colossus 249 program, which was the CM software that flew on Apollo 9, are available.  They are from different reproductions of the same original 1968 printout, as owned by different AGC developers.  Scanned image set #1 is in far better condition than set #2 and is more legible, and is the only one linked here.  But only the less-legible set #2 was actually available to me when I originally implemented Colossus 249.  Any hand-written notes on the listings are from the original developers (as far as I know), and so those notes differ on the two image sets.

Apollo 10
Colossus 2

Syntax-highlighted, hyperlinked,  HTML

Document Library


This is the first software release of Comanche targeted for Apollo 10 whose rope-memory modules were manufactured.  It is not the version that was eventually flown in the mission.

Syntax-highlighted, hyperlinked,  HTML This is the second revision of Comanche targeted for Apollo 10 with rope-memory modules actually manufactured.  It ended up not being flown in the mission.

Syntax-highlighted, hyperlinked,  HTML This is the third and final revision of Comanche for Apollo 10, and is the one actually flown in the mission.  I should clarify that strictly speaking this is not a revision of "COMANCHE".  Instead, it is a side development branch that split off from COMANCHE at COMANCHE 45, and is actually called "MANCHE45", thus if we had the printout of the original assembly listing (which we do not), at the tops of the pages it would say "REVISION 002 OF PROGRAM MANCHE45".  The reason for this is that it allowed Apollo 10 development to continue with MANCHE45 rev 1 and rev 2 while Apollo 11 development was independently going on with COMANCHE 46, COMANCHE 47, etc.

Apollo 11
Colossus 2A
Syntax-highlighted, hyperlinked,  HTML
Document Library
Comanche 51 was the software release initially intended to be used in the Apollo 11 Command Module.  Its rope-memory modules were manufactured.  However, there were subsequent revisions to the software before the mission occurred, so Comanche 51 never flew, and Comanche 55 (see below) did instead.

Syntax-highlighted, hyperlinked,  HTML

Scanned page images
This is the 2nd and final revision of Colossus for Apollo 11, and is the one actually flown in the mission.

Apollo 12
Colossus 2C
On our wish list
Document Library
This is the CM AGC software flown on Apollo 12.

We don't have a copy of this software, but we do know what its memory-bank checksums must be, as well as other documentation about the differences from other revisions.  I'm working on reconstructing it, and while I've had some successes (mostly due to help I've received with it), the challenges are tough and the road is long.  You can read the endlessly-detailed description of how I've been going about it ... but I wouldn't recommend it.  The good news is that we do have some prospects that are much better than my abortive reconstruction attempts for obtaining core dumps from physical memory modules, and to reconstruct the source code via that path, so stay tuned.

Apollo 13
Colossus 2D
On our wish list
Document Library
Comanche 72 was the first revision of the CM AGC software targeted for Apollo 13 that was actually manufactured into core ropes.  It wasn't flown, however, with that honor instead going to revision 72/3 (see next entry).

The good news is that we do have some prospects for obtaining core dumps from physical memory modules, and of reconstructing the source code for Comanche 72 via those dumps, so stay tuned.
On our wish list
This is the CM AGC software flown on Apollo 13.  Strictly speaking, it's not actually COMANCHE at all, but rather is a development branch called MANCHE72 that split off from COMANCHE at COMANCHE 72.  The idea is that bug fixes or other changes strictly for Apollo 13 could be made in MANCHE72 rev 1, rev 2, and rev 3 while changes strictly for Apollo 14 could continue to be made independently in COMANCHE 73, COMANCHE 74, COMANCHE 75, and so on.  This is a technique you use in strictly-controlled software to avoid unintentionally mixing in changes for one mission with those from another.  You see, all changes have to be approved by a so-called "control board", and changes are typically approved for specific missions rather than willy-nilly giving blanket approvals that affect lots of missions at the same time.  Thus if we were lucky enough to have a printout of the assembly listing (which we do not), at the tops of the pages it would say that this is "REVISION 003 OF PROGRAM MANCHE72", and would not mention COMANCHE at all.

The good news is that if Comanche 72 can be reconstructed (see the entry above), then reconstruction of Comanche 72/3 is highly likely.  But it hasn't happened yet, so stay tuned.
Apollo 14
Colossus 2E
On our wish list
Document Library
This is the CM AGC software flown on Apollo 14.

Although we don't presently have the Apollo 14 CM software, Niklas Beug, a developer and user of NASSP, the Apollo-mission add-on for the Orbiter spaceflight simulator, tells us that it's actually possible to use a modified form of Artemis (the Apollo 15-17 software) to fly satisfactory Apollo 14 missions.  Let me simply quote Nik:
We currently don't have any AGC version available that was flown on Apollo 14, which causes a few, mostly procedural issues for NASSP. Apollo 14 was the only mission flown in the time period from July 1st, 1970 to July 1st 1971 and as such used a slightly different coordinate system than the AGC versions before and after that. And using a different coordinate system means that a few hardcoded constants had to be changed every year: star unit vectors, solar and lunar ephemeris in the LGC and a few other numbers. These numbers for the yearly coordinate system used by Apollo 14 are available or can be calculated. So we did just that and created our own AGC versions for Apollo 14. Both CMC and LGC are derived from the AGC versions flown on Apollo 15, Luminary 1E and Artemis (Colossus 3). We determined that these most closely match the programs and capabilities of the AGC versions actually flown on Apollo 14. One regular NASSP user (Alex Bart, ...) tested Apollo 14 with these modified AGC versions and flew the complete simulated mission with them. We are very confident now that these AGC versions are working as they should. As I said, they are essentially still the AGC versions flown on Apollo 15, only slightly modified to properly work with Apollo 14. ... They are available here, as Artemis072NBY71.bin and Luminary210NBY71.bin, [i.e.] Artemis072 and Luminary210 modified to work for the Nearest Besselian Year (NBY) coordinate system of 1971.
Unfortunately, "reconstruction" of the source code for Comanche 108 in a similar fashion to some of the other software versions you see in this table isn't very likely.  The reason is that unlike the software versions which have been (or may yet be) reconstructed, we do not know the memory-bank checksums for Comanche 108; thus even if by some chance the code were to be reconstructed perfectly, we still would not have any assurance whatever that the reconstruction was correct.  And without that assurance, why would anyone use it rather than using Artemis?  So there's little point even attempting a reconstruction other than as a pure intellectual puzzle.

On the other hand, for all I know, a software listing or core-memory-module dump for Comanche 108 could suddenly appear, thus changing everything.  That would be lovely.  In lieu of that, with apologies to the Baroness Orczy,
We seek it here, we seek it there,
We seek Comanche everywhere.
Is it in Heaven?Is it in Hell?
It's a damned elusive ne'er-do-well!
Apollo 15
Colossus 3
Syntax-highlighted, hyperlinked,  HTML Document Library
Artemis 71 was the first revision of Colossus targeted for Apollo 15.  Its rope-memory modules were manufactured, but the program was subsequently revised and Artemis 72 instead flew in Apollo 15 through 17.

We do not have a copy of one of the original Artemis 71 program listings, but it turns out to be possible to reconstruct its source code with confidence from the source code for Artemis 72.  Mike Stewart (thanks, Mike!) has done so, and has provided an instructive full writeup of how he figured it out.


Syntax-highlighted, hyperlinked,  HTML

Scanned page images
Document Library This is the 2nd and final revision of Colossus, flown in Apollo 15, Apollo 16, and Apollo 17.

Also known as Artemis build 072, released (Fabrizio Bernardini tells me) January 31, 1971.  A private collector has graciously allowed us to digitize this program listing for use in Virtual AGC.

Apollo 16
Document Library
Apollo 17
Document Library
Skylab 2

Soon ...

Document Library
Click to enlargeThe Skylab 2, Skylab 3, Skylab 4, and ASTP missions all used the same Skylark 48 software for the Command Module's AGC.  While we do not have a contemporary listing of Skylark 48's source code, we do have dumps of the physical core-rope memory modules actually flown on the Skylab 2 mission from the New Mexico Museum of Space History; many thanks to the museum and Executive Director Chris Orwoll for allowing Mike to dump these modules, and to Larry McGlynn who facilitated!  You can see a photo (click to enlarge) of Mike's setup for dumping the contents of the modules at the right.

For some reason — I guess you use what you have! — the museum stores and/or displays the core-rope memory modules within the actually-flown AGC from the Apollo 15 mission:
Click to enlarge

It's likely that Mike will succeed in recovering the Skylark 48 source code, or most of it, in the near future, via the usual process of adapting source code from elsewhere (such as Artemis 72) until reaching a point where the recovered source code assembles identically to the module dumps, at which point we'll present that code here.  Of course, there are no advance guarantees of success or scheduling!  Until then, you can still run Skylark in our AGC emulator; in the VirtualAGC GUI wrapper of the AGC emulator, for instance, just select the Skylab2, 3, or 4 mission, or the Apollo-Soyuz mission.  Of course, the dumps of each module individually are also, as usual, available in our Rope Module Dump Library.

See also the Programmed Guidance Equations document (part 1, part 2, part 3).
Skylab 3
Document Library
Skylab 4
Document Library
Document Library

Source Code and Binary

You can obtain all of the software mentioned above as being in our collection by installing Virtual AGC on your computer.  The files are contained within a subdirectory named after the software version (such as "Colossus249" or "Artemis072").  The more important files supplied are these:

Source code for major subdivisions of the Colossus program.
Organizer which treats all of the other assembly-language files (*.agc) as include-files, to form the complete program.
Human-readable form of the binary executable as an octal listing.
Binary executable created from binsource (octal listing) file.

In other words, to create a Colossus binary executable rather than using the one provided with yaAGC (such as Colossus249.bin), one simply needs to assemble the file MAIN.agc.  Typically, if all files remain organized the way they are in the yaAGC distribution tarball, the sequence of steps for doing so (from a command-line prompt) would be something like this:

cd Colossus249
../yaYUL/yaYUL MAIN.agc >Colossus249.lst

The binary so produced, MAIN.agc.bin, should be byte-for-byte identical to the binary (Colossus249.bin) provided with the yaAGC distribution.  Therefore, the following Linux command should reveal no differences between the two

diff MAIN.agc.bin Colossus249.bin

(Replace "diff" with "fc /b" in Windows.) 

Technically speaking....

A point which may not be completely appreciated is that Colossus249.bin was not created from the assembly-language source files.  Therefore, the byte-for-byte equivalence mentioned above actually has some significance.  In fact, both the assembly-language source code and Colossus249.bin (or Colossus249.binsource) come from separate readings of the original Colossus assembly listing scan, so their equivalence provides an important check on validity.  (See below.)  The file Colossus249.bin was created from the human-readable/editable ASCII file Colossus249.binsource by means of the program Oct2Bin, with the following steps:

cd Colossus249
../Luminary131/Oct2Bin <Colossus249.binsource
mv Oct2Bin.bin Colossus249.bin

Admittedly, few people are likely to perform any processing of this kind unless contributing a new version of the Colossus code to the Virtual AGC project.


I generally put a lot of effort into insuring that the executable binaries being run in the simulated AGC are byte-for-byte (or in the case of the AGC, word-for-word) identical to the original binaries used in the spacecraft.  Most people won't be interested in this validation effort at all—just as we're not usually interested in how food gets from plants and animals onto the plates in front of us as long as we get our dinners on time; we're happy to trust others to do the work for us—but for those few of you who wonder if the simulated AGC really is running the correct code, the following sections provide explanations.  The different program versions are presented in the order in which they were worked on.

Validity of the Colossus 249 Code (Apollo 9)

The following description of validation was written prior to having the Colossus249 software source code in machine-readable form, and therefore prior to being able to the core-rope image against the source code as described above.  Not only that, but there were seven errors in the core-rope not detected by the procedure described below!

So the description is now obsolete.  But I like it, and it's my website, so it stays.  (Actually, it continues to have value, since we don't yet have all versions of the AGC source, and since it describes how to proceed in the case of recovering poor-quality assembly listings.)

A Tour of the Source Materials

Let's look at a pages from the source materials (actually, a greatly improved version) I've used in reconstructing Colossus 1A (build 249).  In appearance, what we have is a scan of a photocopy of an assembly listing of the program.  Significant portions of the assembly listing have been corrupted (or even destroyed) after scanning, seemingly by processing with the optical-character-recognition capabilities of Adobe Acrobat.  (The source materials used for Luminary 1C, build 131, are generally similar, except that the original photocopy seems to be of better quality, the scans seem to have been performed with much more care, and no destructive post-processing has been performed.  On the other hand, the original Luminary printout seems to have been on computer paper with green and white color bars, and the remnants of the green bars make portions of the listing very difficult to read with certainty.  Most of what I have to say will apply to Luminary as well as to Colossus.)  With some luck, I'll be able to find (or will be given) better source materials at some point in the future, but I've had no luck so far in obtaining such materials and therefore have had to make do with what's available.

An assembly listing is a type of report created when a program's source code is processed ("assembled"), in this case by the assembler program called YUL.  (Actually, by GAP, the successor to YUL, but when I say "YUL" I really mean "either YUL or GAP".)  Such an assembly listing contains not only a listing of the source code, but also other helpful information such as the octal values which the source code assembles to.  The octal codes are what is needed to run the program, while the source code is what is needed to modify the program.  The general layout of the Colossus (or Luminary) assembly listing has a series of sections, of which some of the more interesting are:

Proofing in the Usual Case

From the description above, you will have noted that source code is listed once within the assembly listing, but that octal values created from the source code are listed twice, and are therefore redundant.  Furthermore, knowing the source code, one can recreate the octal values from it.  (The reverse is not as easy.)  Finally, the octal listing is much shorter than the source-code listing.  Therefore, it shouldn't be surprising that it is much easier to create a file containing a valid octal listing of the program than it is to create a file containing a valid source-code listing.

It is also much easier to know that an octal listing is valid than it is to know that a source-code listing is valid.  This is true because the YUL assembler has added "checksums" to the octal values, which can be used to verify (or at least give a high degree of confidence) that the octal values are correct.  The way this works is that the so-called "fixed memory" of the AGC has been divided into 36 banks, numbered 0-43 in octal notation, each of which has a 15-bit checksum created by adding up all of the 15-bit words within the bank.  (Actually, the rule for computing the checksum is more complicated, but that's the basic idea.)  Therefore, if we compute the checksum of a bank and see that it is the same as the checksum provided by YUL, we can be fairly confident that the contents of the bank are correct.  (I should mention that I've provided a utility called Oct2Bin which is helpful in this connection.  Octal codes can be placed in a simple ASCII file, and then processed with Oct2Bin to create a file containing the actual octal/binary codes and to verify all bank checksums.)

Another helpful touch added by YUL is that each octal value is accompanied by a parity bit, contrived to make the number of 1-bits within the octal value (including the parity bit) an odd number.  This is useful in detecting single digit errors of some kinds.  (The digits 0, 3, 5, and 6 have even parity, while the digits 1, 2, 4, and 7 have odd parity.  Therefore, examination of the parity bit could detect substitution of a '2' for a '3', for example, but not substitution of a '5' for a '6'.)

In almost all cases, the following simple procedure suffices to produce a valid octal listing:
  1. Manually proof-read a memory bank of the octal listing against the original assembly listing, using the redundancy of the octal values or manually assembling source code as necessary to account for corrupted areas of the assembly listing.
  2. Compute and verify the checksum.
  3. On failure, go back to step 1.  On success do the next memory bank.
Given a valid octal listing, determining the validity of a source-code listing is rather trivial:  simply run the source code through the yaYUL assembler, and verify that the same octal listing is produced.

This procedure in fact suffices for the entire Luminary octal listing, and for all memory banks of Colossus except banks 35 and 36 (octal).  If you want to learn more about proofing Colossus memory banks 35 and 36, read onward.  Otherwise, quit now!

Postscript:  After writing all of the material below, I was fortunate enough to receive some much-improved page scans of relevant portions of bank 36 from Mr. Gary Neff.  Therefore, bank 36 is now capable of being proofed by the simple method explained above.  However, it was previously proofed using the much more complex methods explained below, and therefore I haven't bothered to change the explanation that follows.

Proofing in Extraordinary Cases

The simple proofing method outlined above fails when the scan of the assembly listing for some memory bank is so corrupted that the bank checksum or some value(s) within the memory bank are completely unknown.  Sadly, this condition occurs within Colossus 1C (build 249) banks 35 and 36.  Click here to see the culprit page of the octal listing.  This corrupted page from the octal listing actually wads together the end of bank 35 and the beginning of bank 36 in such a way that there is no chance of figuring out any but a small part of them.  In this case, one has no choice but to use the redundant listing of the octal values that appears side by side with the source code.

Unfortunately, though, the redundant octal values in the source code listing fail us here, though in different ways for bank 35 and bank 36.  For bank 35, we encounter the problem that the bank checksum appears only within the octal listing and is not redundant.  Therefore, we can theoretically know all of the octal values in the memory bank—if our proofing is 100% accurate—but cannot use the bank checksum to give us the degree of confidence we would like.  For bank 36, the checksum is intact but some of the octal values cannot be directly determined because of corruption within the source-code listing.  (In other words, both the octal listing and the source-code listing are corrupted in the corresponding place.)  As far as proofing is concerned, however, the conclusion is the same:  We need some method of proofing which gives us a high level of confidence in the result, in the absence of a checksum.

Here is the alternate method of proofing used in this cases, which seems to offer a fairly high level of confidence:
  1. Note first that each memory bank actually occupies 4 pages within the octal listing.  For this method, we proof each memory page separately from the others.
  2. Manually proof-read the memory page, using either the octal listing or the source-code listing as desired or appropriate.  While doing so, record the number of digits which have been corrected.
  3. Repeat step 2 until zero corrections have been made in two successive proofings.
  4. To the extent feasible, one of the 0-error proofings should use the octal listing, and the other should use the source-code listing.
The theory behind this procedure is that in any given proofing a certain percentage of errors is corrected, and that by monitoring the number of errors that have been corrected we can get a good sense of the number of errors remaining.  To take a "typical" case, suppose that on the first proofing 100 errors are detected and corrected, on the second proofing 1 error is detected and corrected, and on the third and fourth proofings 0 errors are detected.  It would be reasonable to suppose that each proofing pass detects 99% of errors, so the expected number of errors remaining after the 4th pass is 100*0.01*0.01*0.01*0.01, or about a millionth of an error.  Of course, this reasoning assumes that our ability to detect and correct different errors has some quality of statistical independence, and that errors are not remaining undetected through some systematic problem.  We try to overcome that objection by making sure that we use both the octal listing and the source-code listing in the proofing process, but how successful we have been is always debatable.

As far as proofing is concerned, this technique suffices for Colossus banks 35 and 36, but does not actually give us a 100% knowledge of the octal values within the banks.  For bank 35, we can simply construct a checksum and be done with it.  But for bank 36, we need to do a little more work.  If you want to know more about that, read onward!

When All Else Fails—Digital Archaeology

In a case like bank 36, the technique above may give us confidence that all octal values that are legible in the page images are keyed accurately, but the proofing technique can't give us any insight into the octal values that are simply missing from both the source-code listing and the octal listing.  One trick we can try is to compare the Colossus and Luminary source-code, which actually have a high degree of overlap in general.  Unfortunately, this is not a case in which there is any such overlap, so the comparison is of no help.

Click here (p. 846) and here (p. 847) to see the pages of the source-code listing that are relevant to the problems in bank 36, namely:
  1. Address 36,2634:  This source-code line is simply missing.  We have no clue as to what may have been in it, except that it is seemingly the first part of a double-precision constant whose second part is the word 00000 octal.
  2. Address 36,2734:  Appears to be interpretive code reading "something1 something2", but only a couple of pixel rows are present, and so neither the octal code nor the assembly-language code can be read.
  3. Address 36,2742:  Highly corrupted, but may have the octal code 00021 and may have interpretive source code reading "16D".
  4. Address 36,2747:  Pretty corrupted, but seems likely to be the interpretive source code "CALL" with octal code 77624.
As far as points 3 and 4 are concerned, I think it's pretty easy to convince ourselves that the proposed interpretations are probably correct.  If you want a challenge, you may like to try solving problems 1 and 2 for yourself, before continuing to read my proposed solution below.

Point 2 is somewhat trickier than problems 3 and 4, because the corruption of the page-image is so much greater.  But with some effort, we can still figure it out.  Within the offending page of the source-code listing, notice the notations "REF" and "LAST".  These notations appear only on source-code lines referencing variables.  "REF" gives the number of times the variable has been referred to so far in the assembly listing, which "LAST" gives the page number on which the preceeding reference occurred.  The problematic source line clearly contains REF and LAST notations (even though we cannot read them), and so it is clear that something2 is the name of a variable.  You should be able to convince yourself fairly easily that the variable in question is RCON.  In other words, something2="RCON".  (In case it's still not obvious how you would convince yourself of this, note that "REF 6" to RCON is below the missing source line, while "REF 4" is on the preceding page, but that "REF 5" is nowhere to be found.)  Next, because this section of code is AGC interpreter language (rather than assembly language), we can narrow down the choices for something1; theoretically, the only possibilities are STORE, STODL, or STOVL.  Simply in terms of appearances, it seems likely that something1=STODL.  Unfortunately, the interpreter documentation (Users Guide to the Block II AGC/LGC Interpreter, by Charles A. Muntz) was last updated in early 1965, and contains an incorrect numerical opcode for STODL.  Not to be deterred, though, if we look at the other source-code lines near the missing line, we see intact lines like "STODL RPRE," (the comma after "RPRE", by the way, is actually part of the variable name), from which we can deduce that the octal code for STODL is 16000.  Noting that RCON has the address 1635, therefore, "STODL RCON" must assemble to the octal code 16000+1635+1=17636.  But—and why, I'd like to know, does there always seem to be another "but"?—the missing line of code is preceded by a line of source code reading simply "STADR".  The STADR opcode has the odd effect of complementing the code that follows it.  Therefore—finally!—the missing octal code for address 36,2734 is probably 60141.  This closely matches the remaining pixel fragments (and expected parity bit) of the missing line, and so we can feel good about it.

Having now determined all of the octal values within the bank except at address 36,2634, and noting the checksum for bank 36 is known, we can consider the possibility of recreating the octal value at 36,2634 simply by putting a value there that correctly reproduces the bank's checksum.  This is actually trickier than you may suppose, since the checksum is unique only up to a sign.  (In other words, we don't know whether to insert a value that produces the positive checksum or one that produces the negative checksum.)  Often, either the positive or the negative checksum will be impossible to produce, thus removing the uncertainty.  Where it is possible to produce both the positive and the negative checksum, we note that 90% of the memory banks have positive checksums, so we have a much better chance of being correct if we choose the positive checksum.   All of this turns out to be an anti-climax, however:  Prior to any of the reasoning listed above, our inclination was simply to put 00000 at address 36,2634.  Remarkably, this produces the correct bank checksum without any additional work on our part!  Perhaps this is the first time in history that a gut reaction has been confirmed by checksum.

Happy Epilogue!

After the all of the intricate reasoning above was worked out—and the description above written—Mr. Gary Neff was able to send me some terrific replacement scans of the problematic assembly-listing pages from bank 36, namely pp. 846-847.  If you examine locations 36,2634, 36, 2734, 36,2742, and 36,2747 on the new scan, you'll see that they indeed contain exactly the values deduced above!  I still think the exploration I went through before receiving the new scans is a valuable illustration of technique, not to mention being soothing to my ego, so I'll leave it in place for future generations to enjoy.  :-)

Validity of the Comanche 055 Code (Apollo 11)

The Comanche 055 page images became available after the Colossus 249 and Luminary 131page images had already been converted to source-code files, and prior to any other missions becoming available.  The conversion technique was very abbreviated compared to that of Colossus 249, as follows:
The binary thus produced by yaYUL is supplied in the source tree and used for regression testing.

Validity of Artemis 072 Code (Apollo 15-17)

The Artemis 072 executable (Artemis072.bin) was created by manually entering binary data from scans of an Artemis 072 assembly listing.  The data was proofed and reproofed until the bank checksums, as described above, were correct, and the binary was made available to the public on 2006-01-10.  On 2010-02-20 (more than 4 years later), source code was finally available and assembled.  A single error which did not affect the memory-bank checksum was found and corrected in the binary.  So we are now confident that what is being provided is both complete and correct.

This is an interesting object lesson.  My previous notes in this section, prior to the availability of source code, made it clear that I wouldn't be surprised if additional pairs or even triplets of errors were found in the binary ... but the notion that an unpaired error might still exist which did not affect the checksum never occurred to me.  The problem is a weakness in the checksumming algorithm used in the AGC.

If the AGC had used a "normal" checksumming algorithm (from the modern point of view) with 2's-complement arithmetic, any single error in data would have caused the checksum to change, and the error to be detected.

But in the 1's-complement arithmetic of the AGC, simple checksumming does not have the same desirable property.  For example, since there is both a +0 and a -0, which produce the same result in an addition, a memory error in which +0 is replaced by -0 (or vice versa) produces the same checksum and would be undetectable.  But algorithmically the problem is more complicated than that, since a simple checksum algorithm was not used.  The actual algorithm used on a data-word by data-word basis was this:
For example, if the running checksum was 030000 (a positive number), and the next data word was 020000 (another positive number), then adding them would give 050000 (arithmetic overflow!), which would be decremented (due to the overflow) by 040000 to give 010000, and then incremented by 1 to give a running checksum of 010001.

If you are amused by such stuff, I invite you to consider the ways in which such an algorithm might fail to detect error.  In the case of the Artemis 72 binary, the now-corrected error was a single bit (the most significant bit) of a single memory location.

Validity of Colossus 237 (Apollo 8)

The Colossus 237 program listing presented a special challenge, because the only physical copy we've had access to is truncated.  It contains all of the source code, as well as the binary words associated with the source code on a line-by-line basis.  But it is missing the octal listing that condenses all of the binaries into an easily accessible form.  Because the octal listing is missing, the "bugger words" that are used to make the memory-bank checksums compute correctly are also missing.  Thus, important redundancies and checksums that normally would be used to provide confidence of correctness are missing.

Therefore, the following unique validation procedure was applied instead:

Thus while we can't have as much confidence in the validity of the Colossus 237 transcription, in comparison to almost all of the other Colossus and Luminary transcriptions, the code has nevertheless been double-checked both manually (visually) and in an automated fashion as well.

Validity of Solarium 055 (Apollo 6)

Solarium 055 presented a big challenge because it is the first Block I code processed, and no Block I support was previously provided by the yaYUL assembler.  However, support for Block 1 was in fact added to yaYUL, and after that validation of Solarium 055 differed little from any other AGC program.  Unfortunately there is no way as of yet to run Solarium in yaAGC, so we have not had the additional pleasure (and assurance!) of witnessing it run.

This page is available under the Creative Commons No Rights Reserved License
Last modified by Ronald Burkey on 2024-01-25.

Virtual AGC is hosted