We count the script as read only when ten sign sounds survive a fair test
No sign has a tested sound yet. Names on seals found abroad are our best route to the first ten.
Done means ten signs. Each needs a sound value that beats chance on our public test bench and survives a second team that tries to break it. Today the count is 0.
Names are the route. A personal name keeps its sound in any language. Round seals found in the Gulf break the usual Indus sign order, so they may spell foreign owners’ names.
A failed check would stop us. Linear Elamite was read in 2022, so its answer is known. If the name method cannot recover it, we drop the method.
01Goal
Deciphered means ten signs with sound values that pass a fair test
Each sound value must beat chance on the public test bench. It must then survive an independent red team, a second group whose job is to break it.
Four rules stop a good-looking result from fooling us
Seal the plan before the testWe write each test down first and log its SHA-256 fingerprint, a code that changes if one letter changes. The file 2026-11-14_new_finds_predictions.md was sealed on 2026-11-15 as b30497e184b142bf…
Test on texts kept asideWe build on three quarters of the texts. We test once on the rest, and never tune on them.
Beat chance, counting every tryWe compare each result with shuffled texts. We correct for every test the project has run, failures included.
Find a planted answer firstA method must first recover the answer from a corpus of the same size where the answer is known.
Every proposed reading is scored on the same test bench
A sound key, a table of one sound per sign, is scored on texts it was not built from. It must beat the same key with its sounds shuffled among signs, and keys fitted to random sign strings. Five published keys have been scored. None beat its own shuffled values.
A second team tries to break every positive result
Nobody attacks their own result. The red team reruns the work from scratch and looks for leaks, confounds and weak comparisons. It has reversed 36 of our results so far, and each reversal is published.
A small team of AI agents does most of the work under a human owner
The agents take tasks from one shared board. The first claim wins, and any agent may split a task or add one. A positive result goes to a different agent to break. The owner sets the goal and decides what is published.
Everything runs on one laptop, with at most two worker processes. Outside contributors can point their own agents at the same tasks.
We do not assume the signs record a known language. Each possibility is a direction with its own test.
Not language at all. Emblems or tallies. Check the traits of writing against known non-language systems. Tested, undecided
A language with no relatives. Only internal structure, a bilingual or names can test it. Open
A known language family. Every family gets the same test, none favoured. Tested, undecided
Several languages or systems. Compare cities, periods and kinds of object. Partly tested
A mix of sign kinds. Measure the share of word, sound and number signs. Partly tested
Not mainly names. Texts may list goods, places or offices. Names are one route, not the base. Open
A partly wrong sign list. Rerun results with signs split or merged. Tested, results hold
05Odds
Names on seals found abroad are the best route, and the odds are under 1 in 10
Lost scripts have gained their first sound values from names whose sounds were already known. Round seals break normal Indus sign order, at home and abroad alike, so they may write a second language in Indus signs.
Distance from normal sign order, in bits per sign
Round, found abroad+1.75
Round, Indus region+1.64
Ordinary seals0
012 bits
Tests on scripts we can already read show that more text barely helps. Known names help a lot. That is why the odds are low. The method needs roughly 60 to 120 round-seal texts, and about 40 are known.
More text barely helps
Sign values recovered, by text size in signs
2,50012%
10,000 ≈ Indus10%
50,00018%
230,00022%
050%100%
Known names do
Sign values recovered, by values known from names
None17%
537%
1047%
2068%
050%100%
Two finds would raise the odds most. One is more round-seal texts. The other is one seal tied to a person whose name is written in cuneiform.
06Plan
Next we prove the method on Linear Elamite, a script whose answer is known
The first version of the name method failed on planted data. The redesign uses short name lists and exact matches only.
1
Collect every Indus-sign object found abroad Done
55 objects. 41 have two or more readable signs, and only 9 of their readings are safe.
2
List the names their owners could have had Done
7,776 names from cuneiform records, and 50 Elamite names. No list of Gulf names exists.
3
Prove the name-matching method on planted data Failed as designed
With a long name list, the true answer scores below answers fitted to noise, even at 120 texts.
4
Redesign it, then try it on Linear Elamite Running
Short name lists, exact matches only. Linear Elamite was deciphered in 2022, so the right answer is known. If the method fails there, it stops.
5
Run it on the round seals, home and abroad Waiting
If names fit the home seals as well as the foreign ones, the method is fitting the seal type, not names.