There are fewer than 350 Cretan hieroglyphic inscriptions. They were cut into seals, clay bars, medallions, and a four-sided prism class that scholars label CHIC P. The script is logo-syllabic. It has not been read.

It ran from roughly 2100 BCE to 1700 BCE on Crete. For most of that span it was used alongside Linear A. The two systems were not sequential. They overlapped for around two centuries.

Linear A and Cretan hieroglyphic appear in the same buildings. Quartier Mu at Malia is the clearest case. Excavators found both scripts in the same workshop complex, in some cases in adjacent rooms, in at least one case in the same deposit. The working assumption is that the same scribes used both.

Two systems. Same desks. Different purposes that nobody can now name.

The CHIC corpus (Corpus Hieroglyphicarum Inscriptionum Cretae, Olivier and Godart, 1996) catalogues 314 documents. Seals and sealings dominate. There are also clay bars, cones, and medallions. The four-sided prisms carry signs on each face. Reading order on the prisms is unresolved. You can begin on any face. The text closes on itself.

I have been looking at the prism CHIC #294. It is steatite. Four sides. Each side carries between three and five signs. One side has been polished smooth by handling so that two signs are illegible. The catalogue records this as "abraded by use." The other two faces are sharp.

The signs that recur most often across the corpus are catalogued by number. Sign 044 is the so-called "gate." Sign 070 is the "eye." Sign 019 is a small bent figure that has been described as a left hand. The hand sign appears 38 times in the corpus. On 31 of those occurrences it is oriented with the thumb on the viewer's right, which would make it a left hand seen palm-out. The other 7 are reversed.

Linear A and Cretan hieroglyphic share some signs. Not many. Estimates run between 15 and 20 percent overlap depending on who is counting. The shared signs do not necessarily carry the same values. A sign that is syllabic in Linear A may be logographic in hieroglyphic. The systems are not transliterations of each other. They are parallel.

What I keep coming back to is the desk. A scribe at Malia, sometime around 1850 BCE, sitting in a room in Quartier Mu, with two scripts available. They chose one for some documents and the other for others. The choice was not arbitrary. It cannot have been arbitrary, because they maintained both for two hundred years. But the rule that governed the choice is not visible in what survives.

I made a list of the documents from Quartier Mu where both scripts appear in the same deposit. There are seven such deposits in the published record. I counted them twice.

One of the four-sided prisms in the corpus, CHIC #292, has signs on three faces only. The fourth face is blank. The blank face is not damaged. It was left blank when the prism was carved. The catalogue notes this without comment.

I have not found a published explanation for the blank face. I have looked.

Malbolge is an esoteric programming language. Ben Olmstead released it in 1998. It was designed to be as difficult to program in as possible. The first working Malbolge program was not written by a human. A program called the Lisp Beam-search generated it two years after the language existed.

That program is seven lines long. It prints "HEllO WORlD" with the casing exactly like that.

I copied it out:

(=<`#9]~6ZY32Vx/4Rs+0No-&Jk)"Fh}|Bcy?`=*z]Kw%oG4UUS0/@-ejc(:'8dc

That is one line. The published version wraps. I am counting the wraps as lines because that is how it appears in the original post on the language's page. Count them again. Seven.

The instruction set has three trits per instruction. Base three. Every instruction encrypts the next one according to a fixed table of 94 values. The program is not what it looks like. After the first cycle the memory has been rewritten.

I worked through the first six instructions by hand on Sunday. By the seventh I had lost the thread and had to start over. I started over four times.

There is a mailing list archive from 2000 where someone named Andrew Cooke describes solving it. He says he used a local beam search with a width of 2000. He says it took him "a weekend." The page is cached. The original is gone. The cached version has a gap where a code block should be. The surrounding text refers to the block.

The word KALENDA appears in the URL of one of the mirrors. Not in the content. In the path: /mirror/kalenda/malbolge/. The mirror is hosted at a university in Kraków. I cannot find what KALENDA is supposed to stand for there. The rest of the directory is empty.

I am going to try the program on the reference interpreter tonight. I want to see the casing come out wrong on purpose.

I bought this domain yesterday.

A few months ago I found a notebook in a box from my last move. I don't know whose it is. It's not mine. It has what looks like grammar rules for a language I can't identify, written in handwriting that isn't mine, in a notebook I apparently own.

I've been working through it since March.

This is where I'll put what I find.

Ten inscriptions on bronze. Two on stone. One on a spatula. One on a fragment whose original support is not identified. Fourteen in total. The Byblos syllabary, sometimes called pseudo-hieroglyphic Byblos, was excavated by Maurice Dunand between 1928 and 1932 at the site of ancient Gubla, modern Jbeil, Lebanon.

The signs number around 90 to 100, depending on who is counting. This is the count consistent with a syllabary rather than an alphabet or a logographic system. Some of the signs match later Phoenician letters in shape. The match is close enough that it cannot be coincidence and loose enough that it cannot be ancestry. Nobody has resolved this.

The corpus is too small to decipher by internal analysis. You need around 30 to 40 substantial inscriptions for a syllabary of this size to yield to combinatorial attack. There are 14. There will not be more. The site has been excavated to bedrock.

I list them by Dunand's designation.

• Bronze tablet a (the longest, 225 signs)
• Bronze tablet b
• Bronze tablet c
• Bronze tablet d
• Bronze spatula
• Stele a
• Stele b
• Stele c
• Stele d
• Fragment e
• Fragment f
• Fragment g
• Fragment h
• Object i

That is fourteen. Count them again. Fourteen.

Stele c is the one I keep coming back to. It is limestone. It was broken in antiquity. The break is along a roughly horizontal line that removes the central portion of the inscription. The top survives. The bottom survives. The middle is gone. The break surfaces are weathered, which means the stone was broken and the pieces separated before they were buried. They were not buried together. The top and bottom were recovered from different layers of the same deposit.

I have not found a published reconstruction of the missing section. Maurice Sznycer's 1977 analysis acknowledges the gap and proceeds around it. Garbini in 1979 estimates the missing portion at 11 to 13 signs based on the spacing of what remains. He does not propose what those signs were. He says the loss is "decisive."

I add Stele c to the list. M-346. CHIC #292. The cached page with the code block removed. Now this.

One more thing about the Byblos corpus. Bronze tablet a was found rolled. It was unrolled in the laboratory in 1930. The unrolling damaged the left edge along the full length of the tablet. Dunand notes this in the publication. The damage is described as "regrettable but unavoidable given the condition of the metal." The signs along the left edge are abraded or lost on every line. The right edge is intact.

The reading direction of the Byblos syllabary is not established. Some scholars argue right to left, by analogy with later Semitic scripts. Others argue left to right, by analogy with the Aegean systems that may have influenced it. The argument cannot be settled from the corpus. The corpus is too small.

If the script reads right to left, the damaged left edge contains the end of each line. If it reads left to right, the damaged left edge contains the beginning. Either way, what survives is fourteen documents from which a consistent portion has been removed.

I am keeping a list. I do not yet know what the list is for.

There is a numbers station designated S06s by enthusiasts. Russian-speaking woman, mechanical cadence, five-figure groups. It has been broadcasting since at least 1996. Most of its transmissions are unremarkable in the sense that all numbers stations are unremarkable. A header. A count. A message. A repeat. A null.

I am not interested in S06s.

I am interested in what someone recorded over the top of it on 4625 kHz on the night of 2014-11-09, between 23:47 and 23:54 UTC, while S06s was off the air. The recording is on the Priyom forums, uploaded by a user called sigint_b in 2015 and not commented on since. The file is a seven-minute capture. The first thirty seconds are static. The rest is a woman reading what is not numbers and not Russian.

The phonemes are consistent. I sat with the file and a notepad for two evenings and built an inventory. Fourteen consonants. Five vowels. No clusters longer than CC. Stress falls on the penult except when it doesn't, and when it doesn't it falls on a syllable that contains one specific vowel I have transcribed as /ɨ/. The syllable structure tolerates final nasals but not final stops.

She reads in groups. The groups are not five. The groups are seven. I counted them four times.

The intonation contour at the end of each group is flat. Not falling, which would mark a statement in most languages. Not rising. Flat. The same pitch as the rest of the group. She does not breathe between groups. She breathes between sets of groups, and the sets are also seven.

I tried to find a base form. There is a recurring trisyllable that opens roughly one group in nine. /ka.lɛn.da/. I wrote it down before I noticed I had written it down.

There is no header. There is no null. There is no repeat. At 23:54:11 the woman stops mid-group, the carrier holds for nine seconds, and S06s comes back on frequency at 23:54:20 with its standard test tone as if nothing had been there.

The Priyom database has no entry for a 2014-11-09 interruption. The logs for that night show S06s active from 23:00 to 00:00 with no gap.

sigint_b's account has one upload and no posts. The IP is logged as a university netblock in Kraków.

I downloaded the file. I made a second copy. I have listened to it through twice, once with headphones and once without, and the second time I noticed that the breaths between sets are not breaths. They are the same length every time. 1.4 seconds. Count them again. They are.

Between roughly 3100 and 2900 BCE, in what is now southwestern Iran, scribes produced about 1,600 clay tablets in a script we call Proto-Elamite. Most are accounts. Grain, livestock, workers. Numbers and commodity signs. Partial decipherment has been possible for the administrative corpus because accounting is structurally redundant. You can guess at a sign for "sheep" by watching it behave like a sheep over a hundred ledgers.

Sixteen of the documents are not accounts.

They are monumental. Carved or impressed on objects meant to be displayed, not filed. The sign inventory overlaps with the administrative script but the syntax does not. The repetitions that let you crack a ledger are absent. Jacob Dahl at Oxford has worked on the corpus for years. He has said publicly that the monumental texts use the script for a different purpose. He has not said what purpose.

Sixteen.

1,600 administrative tablets. 16 monuments. The ratio is exact to two significant figures, which is probably accidental. A multiple of 7 with a remainder of 2. I wrote that down and then wrote it down again.

The monuments are scattered across Susa and a few smaller sites. They are not a set. They were not produced together. They share a function we cannot name and a grammar we cannot parse. The administrative texts can be read partially because they were meant to be read by someone other than the writer. The monuments were not. Or they were meant to be read by someone who already knew what they said.

I spent the evening with the published photographs. The CDLI has good scans of most of them. One detail I had not noticed before: on at least four of the sixteen, the sign sequence runs along one edge of the object only. The other edges are blank or bear decorative impressions that are not script. The text occupies one side. It is not centered. It is lateralized.

I went back to the administrative tablets to check. The accounting texts use both faces of the tablet and both registers. They fill the available surface. The monuments do not.

Count them again. Sixteen. Four with the script confined to one side. I have not checked the remaining twelve carefully. The photographs of three of them are partial.

Dahl's 2018 paper notes in a footnote that one monumental fragment, M-346, has been broken in antiquity along a line that removes what would have been the central portion of the inscription. The break is clean. The two surviving sections do not join. The middle is gone.

I am keeping a list of breaks like this. It is getting long.

I ran the KALENDA program again last night. Same tarball, same interpreter, same 49 bytes. I wanted to look at it as a hex dump instead of the ASCII garbage I logged in 2012.

Here it is, xxd output, seven rows of seven bytes each padded to the tool's default width:

00000000: 07 1a 2c 07 1a 2c 07                     ..,..,.
00000007: 1a 2c 07 1a 2c 07 1a                     .,..,..
0000000e: 2c 07 1a 2c 07 1a 2c                     ,..,..,
00000015: 07 1a 2c 07 1a 2c 07                     ..,..,.
0000001c: 1a 2c 07 1a 2c 07 1a                     .,..,..
00000023: 2c 07 1a 2c 07 1a 2c                     ,..,..,
0000002a: 07 1a 2c 07 1a 2c 07                     ..,..,.

Three bytes cycling: 0x07, 0x1a, 0x2c. BEL, SUB, comma. Sixteen instances of 0x07, sixteen of 0x1a, seventeen of 0x2c. The cycle length does not divide 49 evenly. It runs 07-1a-2c sixteen times and then one extra 0x07 at the end. Except that is not what happens. Count the last byte again. It is 0x07. Count the middle. Byte 24, zero-indexed, is 0x07. The cycle is not 3. The cycle is 3 with a shift.

I wrote it out as a 7x7 grid, row-major:

07 1a 2c 07 1a 2c 07
1a 2c 07 1a 2c 07 1a
2c 07 1a 2c 07 1a 2c
07 1a 2c 07 1a 2c 07
1a 2c 07 1a 2c 07 1a
2c 07 1a 2c 07 1a 2c
07 1a 2c 07 1a 2c 07

The diagonals are constant. Top-left to bottom-right: all 0x07. The anti-diagonal, top-right to bottom-left: also all 0x07. Four corners 0x07. Center cell (row 3, column 3, zero-indexed): 0x07.

I read the rows left to right because that is how xxd prints them. There is no reason the output has to be read that way. It is a stream of bytes. It came out of a stack. The stack is LIFO. If I reverse the byte order I get the same grid rotated 180 degrees. Which is the same grid.

The garbage-file problem. Statistical tests would call this low entropy and move on. It is low entropy. That is not the same as random. It is not the same as meaningful either. It is a shape.

I checked the esolangs wiki page again. The KALENDA comment is still gone from the live page. The Wayback capture from 2009-03-15 still has it. "this produces correct output." No other edits from that account, ever.

The 0x07 byte is BEL. The terminal rings when you print it. I piped the output to cat and my laptop chimed seven times before I realized what was happening and killed it. There are sixteen 0x07 bytes in the output. It chimed seven times because I hit Ctrl-C. I ran it again and counted. Sixteen chimes. I had miscounted the first time.

Count them again. Sixteen.

Sixteen is also the Proto-Elamite monument count. I am writing this down because I am writing everything down. It is a multiple of eight and I do not know what to do with that.

I wrote an esolang in 2011. It is called Toadskin. Stack-based. Six of Brainfuck's eight instructions kept as they were. Two modified to operate on the stack instead of a tape. Word definitions in the Forth style, :name ... ;. Single characters otherwise. I described it at the time as Brainfuck meets Forth, a fight ensues. That description is still on the wiki.

The reference interpreter is a tarball. 14 KB. I have not touched it since 2012.

It has a bug.

The bug is in [. The loop entry. In the normal case, [ checks the top of the stack. If zero, jump past the matching ]. If nonzero, enter the body. At ], return to [ and check again.

The edge case: if the body produces exactly seven stack operations before reaching ], the interpreter does not re-check. It continues into the next iteration without consulting the stack. The loop does not terminate when the condition would have said to stop. It terminates when something else stops it. End of input. Stack underflow. Whatever comes first.

Six operations: fine. Eight: fine. Seven: the check is skipped.

I do not remember writing this. I have read the source. The conditional is structured so that the counter increments before the comparison rather than after, and the comparison uses > where it should use >=. Off by one, in a way that only matters at one value.

There is one program in the wiki's example section that triggers the case. Five lines. I did not put it there.

:f 0 1 + 1 + 1 + 1 + 1 + 1 + ;
:g [ f . ] ;
0 g
1 g
g

The comment under it reads: this produces correct output.

The contributor's username is KALENDA. The account has one edit, this one, made on 2009-03-14 and reverted by an admin the same day for lacking context. The Wayback Machine has the page from before the revert. I have the page from after. The revert removed the program. It did not remove the comment.

I ran it. The output is 49 bytes. Seven groups of seven. I expected the output to vary by machine, because the bug consults whatever happens to be on the stack when the check is skipped, and the stack at that point contains uninitialized values on some implementations. It does not vary. I ran it on my laptop, on a Raspberry Pi, on a VM with the locale set to Icelandic. Same 49 bytes.

I have the hex dump in a text file. I have not done anything with it.

The other thing I noticed, reading the source again: the loop counter variable is named kal. I do not remember naming it. I would not have named a counter that. I name counters i and j and occasionally n.

The tarball's mtime is 2012-07-02. I have a backup from 2011-11 on an external drive. The 2011 version has the same bug. The counter is named i in that one.

I have not opened the 2010 version. It is on a drive in a box I have not unpacked.

The Vinča culture lived along the Danube and its tributaries between roughly 5300 and 4500 BCE. They made pottery, figurines, spindle whorls. On many of these objects they cut marks. Geometric. Combs, chevrons, ladders, crosses inside circles, the slash, the V, the X.

There are over 200 distinct signs in the corpus. The count varies by scholar. Winn cataloged 210 in 1973. Other counts run higher depending on what gets treated as a variant.

The marks are older than Sumerian cuneiform by more than a thousand years. They are older than Egyptian hieroglyphs. There is no known descendant script. Nothing carries forward.

The current consensus is that the marks are not writing. They convey meaning. They do not encode language. What they do instead is unsettled.

I have been reading Shan Winn, Harald Haarmann, and a 2016 paper by Marler. The marks recur across hundreds of kilometers. The same signs at Tărtăria, at Vinča-Belo Brdo, at Turdaș, at Banjica. People who never met agreed on what these were for.

The detail I keep returning to. On a subset of the figurines, the marks appear only on the left side of the body. Left shoulder. Left thigh. Left side of the face. The right side is blank. This is documented in the Tărtăria material and in figurines from the Gradešnica plaque region. Not universal. But present often enough that it has been noticed and not explained.

I counted the figurines in the Winn plates where the marks are clearly lateralized. I got seven. Then I counted again. Still seven.

The Tărtăria tablets, found in 1961 by Nicolae Vlassa, are three small clay objects. One is round, pierced, with marks arranged around a cross that divides the face into quadrants. One is rectangular with a register of signs across the top. One shows an animal and a branch and a row of signs along one edge. The find context was disturbed. The dating has been argued for sixty years.

The round tablet is read by some scholars clockwise starting from the upper right. By others counterclockwise starting from the lower left. By others not at all, on the grounds that reading is the wrong verb.

Haarmann's position is that the Vinča marks are a "sign system" rather than a script. He distinguishes between systems that record speech and systems that do something else. He does not say what the something else is. He says it is not our business to assume.

I want to note one more thing and then stop. The figurines with left-side marks are, in the cases I checked, also marked on the underside of the base. The base marks are different signs from the body marks. I have not found a paper that addresses this. It may be addressed somewhere. I have not found it yet.