the serial against the code
the speaker in the story is invented. the tape is not, the numbers are not, and this
page says day by day which is which.
the short version. a serial runs on @ventry089 about six agents saving a
dying language from its last speaker. that man does not exist. the desk does: it runs on
1.2 hours of Torwali and 27 minutes of Piedmontese, both recorded by
Wikitongues from real speakers, and every
number below opens the second of audio it came from.
where the story goes past what a program can do, this table says so instead of quietly agreeing.
three of the days need a human being in a room, and no amount of code closes that.
TAPE1.2 h
WORD CANDIDATES3,218
RULES SIGNED19
RULES REFUSED41
FORMS NOBODY SAID1,105
RANDOM CHAINS THAT LAND0
DIE WITH HIM2,041
day by day
DAY 1
six agents pull sounds, words and rules off raw tape
desk.py run on 72 minutes:
1,164 utterances cut,
33,877 phones heard,
3,218 word candidates,
60 rules written and
19 signed
see itRAN ON REAL TAPE
DAY 2
contradictions are kept as variants instead of being resolved
9 fights stay open in the run - pairs one phone apart where both forms keep coming back. Nothing picks a winner, because nobody alive can
see it
RAN ON REAL TAPE
DAY 3
rules turn out to be him doing somebody else's voice
every rule carries
in_shifted_voice. On this tape
6 of the 19 signed rules live where his pitch or tempo is 2σ off his own median, and
60 register shifts are logged with positions
see itRAN ON REAL TAPE
DAY 4
the desk speaks the language back to him
sentence assembly is not written. Cloning a voice off 45 hours of one speaker is a solved problem; a native speaker's reaction to it is not, and that is the half that matters
NEEDS A LIVING SPEAKER
DAY 5
the words come back 40 ms slower and he takes them anyway
pitch and tempo per utterance are already measured. Comparing one word across two recording dates is a short script and it is not written yet
NOT WRITTEN YET
DAY 6
the machine passes the desk's own test for a speaker
two of the three checks are mechanical. The third one - takes a correction - is a person changing their mind in a room. Nothing here fakes it, and nothing here will
NEEDS A LIVING SPEAKER
DAY 7
the bot says a word that is on no tape anywhere, and it is real
agents/derive.py runs the signed rules backwards.
1,110 forms this grammar allows ·
5 turned out to be on the tape after all ·
1,105 are on no second of it. Control:
0 of 1,110 random chains of the same lengths land anywhere
see itRAN ON REAL TAPE
DAY 8
the queue is sorted by which words are safe to lose
derive.queue() -
2,041 of 3,218 word candidates were heard twice or less and no signed rule rebuilds them. That is the order to work through with a speaker, not the frequency list
see itRAN ON REAL TAPE
DAY 8
his voice has fewer minutes left than the queue has words
fade is measured in 10-minute windows inside each tape
see itHALF OF IT RUNS
the one number that decides whether any of this counts
day 7 is the load-bearing one: a machine producing words nobody ever recorded.
that is easy to fake - stack phones together and call the pile a language. so the desk runs a
control on every pass: 1,110 random chains of exactly the same lengths, phones
drawn by how common they are on this very tape.
0 of them land anywhere in 1,164 utterances. the derived
forms hit 5 times. that gap is the difference between grammar and noise, and
the line prints on every run - including the runs where it would kill the result.
on the Piedmontese tape it does kill it: the gate signed one rule out of 18, an
analogy needs two, so DERIVE returns nothing at all and says why. 27 minutes is not enough
tape. that is the desk working, not the language being empty.
what the desk will never do
judge whether a rule is a joke. hear that a word belonged to somebody's mother.
decide that a pattern is dead. those need the speaker, and the gate exists exactly because the
desk cannot do them. the serial is allowed to imagine that room. this repo is not.