Findings

What holds up, how strongly, and what has been ruled out

Eleven findings about how the script works, ranked by the strength of their evidence, then every route to a reading that has been tested and closed. Each links to its code and full results in the repository.

Square seal: signs along the top, a striped animal below
Pl. 4 · Mohenjo-daro · 32 × 34 mmSeal. Marshall, Mohenjo-daro and the Indus Civilization (1931), seal no. 91.

AStrength

How to read the strength labels

Strong
Held up on texts the method had not seen, survived an independent attack, and survives correction for every test the project ran.
Moderate
Held up on unseen texts, but rests on small numbers or weakens after correction for all tests.
Weak
The broad pattern is real; the specific claim is barely supported.
Exploratory
Found while looking at all the data. Not yet tested on unseen texts.
Descriptive
A measurement, not a test of a hypothesis.

“q” is the chance of a false positive after correcting for all the tests in a family or in the whole project (Benjamini-Hochberg). “Held-out” means texts kept apart while the idea was built. None of these findings gives a sound or a meaning to any sign.

BRanked

Eleven findings, strongest first

Open any finding for its numbers, caveats and code.

  1. 01Sign order follows rules, not just positionsStrong

    Knowing the signs of an unseen text, the real order can be picked out from all orderings of the same signs far more often than chance. It still works with the commonest signs removed.

    40% of 321 held-out runs vs 5.7% by chance; q ≈ 8e-72

    Code and results: structure redteam revalidation

  2. 02About 400 signs: word-signs mixed with sound-signsDescriptive

    The sign count and how often each sign is used look like Proto-Elamite or Sumerian, not like Linear B or an alphabet.

    ≈400 signs (≈565 with unseen ones; 95% CI 508–658); 120 signs cover 90% of uses (Linear B: 63, Latin: 14)

    Code and results: structure

  3. 03Corner ticks move the fish sign to the end of the text Moderate

    A fish with four small ticks closes a text far more often than a plain fish: a mark inside a sign that changes where the sign goes.

    final in 40 of 73 vs 26% expected (p = 1e-7); held-out 11 of 19 vs 7% for plain fish (p = 0.004; q = 0.03 within its family, 0.54 across all tests)

    Code and results: novel_methods redteam

  4. 04Round Gulf-type seals break normal sign order, wherever they were foundModerate

    Round seals of the Gulf type, found abroad or at home, do not follow the usual sign order. This fits Indus signs used to write another language, but plainer explanations are not ruled out.

    order excess +1.64 bits/sign on 8 home round seals (Holm p = 3e-4, rare-sign-matched null); foreign vs home round +0.11 (p = 0.66)

    Corrected on 17 Nov 2026: a label error had hidden the home round seals. Only 3 of the 8 are confirmed round in a printed excavation report.

    Code and results: round_label_audit redteam_round_format redteam_round_order

  5. 05Six followed by fish is a fixed expression Strong

    The six-stroke sign followed by the fish behaves like one set phrase. The proposed ‘seven stars’ reading of seven plus fish has almost no support.

    held-out 7 of 10 vs 3% base rate; q ≈ 6e-8

    Code and results: language redteam

  6. 06Each kind of object has its own vocabularyStrong

    Seals, tablets, copper tablets, rods and pottery each favour different signs, enough to tell them apart on unseen texts.

    held-out AUC: copper 0.87, tablet 0.76, pottery 0.73, seal 0.72, sealing 0.67 (chance ceiling 0.55–0.61)

    Code and results: context redteam

  7. 07Texts written left to right are true reversalsStrong

    The few texts that run left to right are written backwards on purpose, not misread.

    31 of 34 copy-free lines; q = 6e-6

    Code and results: structure redteam

  8. 08Copper tablets come in a few dozen fixed typesExploratory

    Copper tablets repeat about 18 to 64 fixed text-and-picture types. A one-sign face can stand in for the picture, but the sign does not depict it.

    same legend → same sign face 8 of 8 (p = 1e-4, all data); blind matching 5 of 24 vs 3.3 by chance

    Code and results: context logograms

  9. 09Texts can be parsed into a small set of rolesWeak

    An 11-role reading parses most unseen texts, far more than shuffled ones. That positional roles exist is solid; that these 11 are the right ones is not.

    parses 76% of unseen texts vs 44% for shuffles; specific roles p = 0.013, q = 0.085

    Code and results: functional_reading grammar

  10. 10The core of a seal text is one or two word-signsModerate

    The part of a seal text between its fixed opening and closing signs is short and made of whole words, not spelled out sound by sound.

    34% of cores are one sign; recurring pairs keep one order in 91% of uses (held-out 115 of 138, p = 3e-16)

    Code and results: typology core_semantics

  11. 11The fish signs form one word classModerate

    Six fish signs swap for each other, combine with each other and share partners. The idea that fish signs are god names is not supported.

    held-out p = 0.0002–0.0004

    Code and results: core_semantics

CClosed

79 routes tested and closed

38 are ruled out at a stated power. The others cannot be decided with the data that exists today; each says what new evidence would reopen it. Source: the atlas of closed routes.

  • Ruled out the test ran and failed, with enough power to see a real effect
  • Can’t test yet the data cannot decide
  • WeakenedNot supported an earlier result shrank or failed its pass mark
Phonetic routes10 routes · 5 ruled out
  • 1.1Published phonetic keys
    Ruled out
    Verdict: Refuted. Test power: Bench z 11.5 on Linear B.Would reopen with: Fixed one-value key beating noise.language literature
  • 1.2Ventris grid
    Ruled out
    Verdict: Refuted. Test power: Synthetic ARI 0.86.Would reopen with: Long texts with inflection.phonetic_grid
  • 1.3Phonetic complements
    Can’t test yet
    Verdict: Untestable. Test power: 40%.Would reopen with: Larger corpus.internal_phonetics
  • 1.4Homophones
    Can’t test yet
    Verdict: Untestable. Test power: 8/80.Would reopen with: Multi-impression names.internal_phonetics
  • 1.5Ligatures
    Ruled out
    Verdict: Refuted (these signs). Test power: 0 events.Would reopen with: Spelled-out duplicate.internal_phonetics
  • 1.6Determinatives
    Can’t test yet
    Verdict: Untestable. Test power: 20%.Would reopen with: Long texts.internal_phonetics
  • 1.7Substitution X↔YZ
    Ruled out
    Verdict: Refuted ≥20%. Test power: P ≤ 0.08.Would reopen with: Multi-impression names.substitution
  • 1.8Fit to candidate language
    Can’t test yet
    Verdict: Untestable. Test power: Planted z 0.3-2.8.Would reopen with: ~10 anchors.decipherability
  • 1.9Cognate decipherment
    Can’t test yet
    Verdict: Untestable. Test power: Gate failed.Would reopen with: Word division + lexicon.cognate_decipher
  • 1.10Gulf phonetic names
    Ruled out
    Verdict: Refuted. Test power: Low, n = 12.Would reopen with: More Gulf seals.typology
Language identification6 routes · 3 ruled out
  • 2.1Family classifier
    Can’t test yet
    Verdict: Untestable. Test power: LOSO 0.621.Would reopen with: Bilingual.typology language
  • 2.2Family from names
    Can’t test yet
    Verdict: Untestable. Test power: 0/7.Would reopen with: Valid naming statistic.naming_sim
  • 2.3Head direction
    Ruled out
    Verdict: Refuted. Test power: CI [0.36, 0.54].Would reopen with: Larger held-out.core_semantics
  • 2.4External grounds
    Weakened
    Verdict: Weakened (Vedic, Sumerian, Munda). Test power: n/a.Would reopen with: aDNA, loanwords.literature
  • 2.57+fish
    Ruled out
    Verdict: Refuted. Test power: 0/13 held-out.Would reopen with: Nothing in sight.language redteam
  • 2.63+fish
    Ruled out
    Verdict: Refuted. Test power: 2/22 held-out.Would reopen with: Larger corpus.language
Language vs code4 routes · 1 ruled out
  • 3.1Language vs code
    Can’t test yet
    Verdict: Untestable. Test power: LOCO 0.61.Would reopen with: Long texts.language_vs_code
  • 3.2Short-label calibration
    Can’t test yet
    Verdict: Untestable. Test power: 0.69 vs 0.80 gate.Would reopen with: Emblem corpus.short_labels
  • 3.3Entropy statistics
    Ruled out
    Verdict: Refuted. Test power: Fails controls.Would reopen with: Nothing in sight.literature novel_methods
  • 3.4F5 discriminator
    Can’t test yet
    Verdict: Untestable. Test power: n/a.Would reopen with: Long texts.functional_reading
Naming and genre9 routes · 5 ruled out
  • 4.1Two-inventory names
    Ruled out
    Verdict: Refuted. Test power: 0.60-0.85 vs 0.075.Would reopen with: Nothing in sight.pair_numeral_long
  • 4.2Naming outlier
    Weakened
    Verdict: Weakened. Test power: Construct-dependent.Would reopen with: New core definition.naming_sim core_genre
  • 4.3Genre match
    Can’t test yet
    Verdict: Untestable. Test power: 0/6.Would reopen with: More genres.core_genre
  • 4.4Closer = servant of
    Ruled out
    Verdict: Refuted. Test power: C1 53-85%.Would reopen with: Bilingual.genre_alignment
  • 4.5CONN = son of
    Ruled out
    Verdict: Refuted. Test power: 0/48.Would reopen with: Nothing in sight.genre_alignment closer_ownership
  • 4.6Closer = ownership
    Can’t test yet
    Verdict: Untestable. Test power: Gate 0.25 vs 0.75.Would reopen with: Discriminating references.closer_ownership
  • 4.7Fish = deity
    Can’t test yet
    Verdict: Untestable. Test power: AUC 0.32-0.54.Would reopen with: Valid statistic.core_semantics
  • 4.8Tool cluster
    Ruled out
    Verdict: Refuted. Test power: p = 0.66.Would reopen with: Nothing in sight.core_semantics
  • 4.9Fixed compounds
    Ruled out
    Verdict: Refuted. Test power: Missed by 0.003.Would reopen with: Nothing in sight.typology
Administrative and relational readings8 routes · 6 ruled out
  • 5.1Administration (14)
    Ruled out
    Verdict: Refuted. Test power: Large effects only.Would reopen with: CISI 3.1 lots.administration
  • 5.2Rank ladder
    Ruled out
    Verdict: Refuted. Test power: MDE rho 0.12.Would reopen with: More measured seals.rank_ladder redteam_rank
  • 5.3Tablet archives
    Ruled out
    Verdict: Refuted (T1, T3). Test power: 0.96-0.98.Would reopen with: CISI 3.1 lots.tablet_archives
  • 5.4Toponyms
    Ruled out
    Verdict: Refuted. Test power: 0.83 at OR 3.Would reopen with: New off-home sealings.toponyms
  • 5.5Roofed signs
    Ruled out
    Verdict: Refuted at OR 3. Test power: 0.96-0.99.Would reopen with: Citadel finds.roof_context
  • 5.6Copper calendar
    Not supported
    Verdict: Inconclusive. Test power: Ring 0.38.Would reopen with: More copper tablets.copper_series
  • 5.7Context C1-C5
    Ruled out
    Verdict: Refuted / untestable. Test power: Mixed.Would reopen with: Foreign finds.context
  • 5.8Titles on door sealings
    Can’t test yet
    Verdict: Untestable. Test power: 0.004-0.012 at 3×.Would reopen with: 40-80 door sealings.sealing_function
Numerals and metrology9 routes · 6 ruled out
  • 6.1Weight series
    Ruled out
    Verdict: Refuted. Test power: n/a.Would reopen with: Inscribed weights.context
  • 6.2Tall = 10
    Ruled out
    Verdict: Refuted. Test power: n/a.Would reopen with: Nothing in sight.literature
  • 6.3Numeral class
    Ruled out
    Verdict: Refuted. Test power: PPC 9/9 fail.Would reopen with: Class with 12 spike.numeral_class
  • 6.4Numeral base
    Ruled out
    Verdict: Refuted. Test power: 1.00.Would reopen with: Nothing in sight.numeral_grammar
  • 6.5M328 measures
    Ruled out
    Verdict: Refuted. Test power: ≥0.98 at rho 0.3.Would reopen with: Hi-res photos.m328_measure
  • 6.6Counted items
    Ruled out
    Verdict: Refuted. Test power: n/a.Would reopen with: Nothing in sight.redteam_pnl
  • 6.7Long texts as lists
    Can’t test yet
    Verdict: Untestable. Test power: n/a.Would reopen with: More long texts.redteam_pnl pair_numeral_long
  • 6.8Magnitude
    Can’t test yet
    Verdict: Untestable. Test power: 0.15.Would reopen with: ~40 objects.magnitude
  • 6.9Number-names
    Weakened
    Verdict: Weakened. Test power: 14/24 without M59.Would reopen with: New texts.pair_numeral_long redteam_pnl
Iconography and depiction7 routes · 6 ruled out
  • 7.1Animal in text
    Ruled out
    Verdict: Refuted. Test power: V ≈ 0.3.Would reopen with: Nothing in sight.iconography
  • 7.2Logograms
    Ruled out
    Verdict: Refuted. Test power: 17 tests.Would reopen with: Nothing in sight.logograms
  • 7.3Self-labels
    Ruled out
    Verdict: Refuted. Test power: 1/78.Would reopen with: More bangles.self_label
  • 7.4Modifier paradigm
    Ruled out
    Verdict: Refuted. Test power: ≥0.8 at +0.5.Would reopen with: Nothing in sight.modifiers
  • 7.5Referent coherence
    Ruled out
    Verdict: Refuted. Test power: n/a.Would reopen with: Nothing in sight.core_semantics
  • 7.6Direction from text
    Can’t test yet
    Verdict: Untestable. Test power: n/a.Would reopen with: Nothing in sight.structure
  • 7.7Allograph perplexity
    Ruled out
    Verdict: Refuted. Test power: Fails control.Would reopen with: Nothing in sight.structure
Cross-script look-alikes4 routes · 3 ruled out
  • 8.1Linear Elamite
    Ruled out
    Verdict: Refuted. Test power: ~0.7.Would reopen with: More LE values.elamite_bridge
  • 8.2Proto-cuneiform
    Ruled out
    Verdict: Refuted (strong). Test power: 0.92-0.94.Would reopen with: Nothing in sight.protocuneiform
  • 8.3Tamil Nadu graffiti
    Ruled out
    Verdict: Refuted. Test power: n/a.Would reopen with: Nothing in sight.literature
  • 8.4Potters’ marks
    Can’t test yet
    Verdict: Untestable. Test power: 0.22-0.33.Would reopen with: Dated Early marks.potters_marks
External anchors3 routes · 0 ruled out
  • 9.1Meluhhan names
    Can’t test yet
    Verdict: Untestable. Test power: n = 2-5.Would reopen with: New texts.meluhha
  • 9.2Mesopotamian sealings
    Can’t test yet
    Verdict: Untestable. Test power: n/a.Would reopen with: Photos of 2 tablets.sealing_hunt
  • 9.3Bilingual
    Can’t test yet
    Verdict: Untestable. Test power: n/a.Would reopen with: A find.elamite_bridge literature
Chronology2 routes · 2 ruled out
  • 10.1Sign turnover
    Ruled out
    Verdict: Refuted (strong). Test power: 1.0 strong.Would reopen with: Dated strata.half_life redteam_chrono context
  • 10.2Dholavira style
    Ruled out
    Verdict: Refuted. Test power: n/a.Would reopen with: Nothing in sight.redteam
Cycles 7-1017 routes · 1 ruled out
  • 10b.1M328 on offering scenes
    Can’t test yet
    Verdict: Untestable. Test power: Power 0.Would reopen with: ≥ 600 dpi scene tablets.scene_text
  • 10b.2Bangle oval
    Can’t test yet
    Verdict: Untestable. Test power: n = 0 (19 needed).Would reopen with: HARP bangle list.bangle_oval
  • 10b.3Firing stage
    Can’t test yet
    Verdict: Underpowered. Test power: 0.06.Would reopen with: HARP sherd DB.firing_stage
  • 10b.4Seal texts as identifiers
    Can’t test yet
    Verdict: Undecidable; formula layer reversed. Test power: 0.03-0.07 (repulsion).Would reopen with: Lot-level groups.identifiers redteam_identifiers
  • 10b.5City-specific sign stock
    Not supported
    Verdict: Inconclusive; large stock excluded. Test power: 0.92 whole / 0.12 slot.Would reopen with: 3× dated Harappa seals.site_signal
  • 10b.6Name + title formula
    Not supported
    Verdict: Not supported. Test power: 0.94 at f = 0.5.Would reopen with: Register-split nulls.slot_novelty
  • 10b.7Omission order
    Weakened
    Verdict: Reversed. Test power: Null rho median 0.52.Would reopen with: Independent measurements.space_and_sealing reg_size redteam_size
  • 10b.8Closer marks used seals
    Not supported
    Verdict: Not supported. Test power: 0.51-0.75 at OR 2.Would reopen with: Lothal per-impression IDs.space_and_sealing
  • 10b.9Repeats as re-cut/office seals
    Weakened
    Verdict: Withdrawn / reversed. Test power: n = 15 families.Would reopen with: Dated formula seals.local_repeats redteam_local_repeats
  • 10b.10Line breaks
    Can’t test yet
    Verdict: Untestable. Test power: 0.77 / 0.51.Would reopen with: 150-300 true wraps.line_breaks
  • 10b.11Detached signs as a class
    Not supported
    Verdict: Not supported; covariates inconclusive. Test power: ≤ 0.32.Would reopen with: Face position from photos.detached_signs
  • 10b.12Direction class / slot time / animal
    Can’t test yet
    Verdict: Untestable / inconclusive / refuted (closer type). Test power: 0.00-1.0.Would reopen with: Photo direction coding.direction_time_animal
  • 10b.13Grammar as predictor
    Not supported
    Verdict: Not validated. Test power: Bigram wins by 0.3 bits.Would reopen with: Annotated positive control.grammar
  • 10b.14Household/office signature
    Ruled out
    Verdict: Refuted (two cities). Test power: 0.90-1.0 at half strength.Would reopen with: HARP lots.reg_findspot reg_harappa
  • 10b.15Ticked fish by block
    Weakened
    Verdict: Weakened (lead). Test power: 0.44 / 0 replication.Would reopen with: More located seals.reg_findspot redteam_fish_block fish_marshall
  • 10b.16Depth chronology
    Not supported
    Verdict: Inconclusive. Test power: 0.92-0.94 broad.Would reopen with: HARP 3B vs 3C.reg_chrono
  • 10b.17Carver hands; craft assemblage
    Can’t test yet
    Verdict: Untestable / inconclusive. Test power: 92-100% / 0-4.5%.Would reopen with: ≥ 30 px/mm; craft coding.

DReversed

36 of our own results that the red team broke

These looked positive and did not survive an independent attack. They show where the traps are.

Show all 36
  1. v1 split leakage. 45% of held-out texts had a near copy in dev. Order prediction fell from 63% to 40.2% (vs 5.7% chance) on the clean split.
  2. Dholavira regional. 0.71 vs 1.19 became “not regional” once the comparison sites were added.
  3. Left-to-right direction. 134/150 became 31/34 on copy-free lines. Jar 33 vs 3 became 5 vs 0 (p = 0.03, q = 0.18).
  4. 7+fish. p 0.0025, then 0.017, then 0.07 on seals.
  5. M99 seal-specific. q = 1e-6 became 1.08×, p = 0.07.
  6. Rod register. 9 vs 0.5 became 6 vs 0.34, Mohenjo-daro only.
  7. Copper picture. 272 pairs became 18 legends. Sign level 0.485 became 0.030.
  8. Template test. 67% vs 25% was circular.
  9. Functional-reading lexicon null. q = 0.085.
  10. 3+fish. Failed held-out (2/22).
  11. Counted items. Code bug in `ll_spike`.
  12. Long texts. Rigged null.
  13. Number-names. M59 gives 15/16 hits.
  14. Rank signs. M336 is a W706 carving variant. M249 weakened. Zebu +42% became +11% (split luck).
  15. CONN = dumu. Built into the grammar; 0/48 grammar-free.
  16. naming_sim outlier. Construct-dependent.
  17. half_life P3. Family-driven.
  18. 1+12 numeral. Came from 3 families.
  19. Allograph perplexity. z −3.6 became p = 0.67.
  20. Typology spec A. IA 30/30 and NONLING 30/30 from the same machinery.
  21. Malayalam reads Sanskrit. 36% through loanwords.
  22. M99|M402 homophone. Did not replicate.
  23. M12 ligature. Below the bigram null.
  24. paragorth AAB repeats. p = 0.023; did not replicate in Wells.
  25. shaunmortal F23. Retracted.
  26. tamil-full-80. Also “passed” on planted Sanskrit and English.
  27. Cross-site short formula layer (identifiers S3). 145 objects on repeated 1-3 sign texts vs 151 expected under a bigram null; 0.89 vs 0.18 locality contrast was a length artefact (equal length: 0.51 vs 0.53).
  28. Seal cores more novel than personal names (identifiers). Ur III cores built the same way are more novel (0.96-0.995 vs 0.91-0.92); survives only against single names.
  29. No identifier repulsion as evidence against identifiers. Power 0.03-0.07 among contemporaries; real Ur III identifiers cluster.
  30. Repeated 4+ sign texts are re-cut seals of one holder, local like Ur III (red-team reading). Withdrawn by local_repeats: 43% same-site pairs vs 45% random; the 0.18 came from small reference sites.
  31. Numeral formula as a new find; “other 20 objects exceed chance” (local_repeats). The formula is the known M162 count frame (Wells W390 lumps five signs); the residual is 20 vs 18.6 expected under a fair null (p = 0.42).
  32. “Texts grow, never shrink” (local_repeats). p = 9e-9 shrank to 6 vs 1.8 expected (p = 0.005 trigram; 0.07 bigram); the singleton comparison measured commonness.
  33. Graded omission order on small seals (space_and_sealing T4, reg_size). Closer slopes equal a keep-one-closer random-deletion null (−0.0212 vs −0.0212); only “about one closer per text” stands.
  34. reg_size “new” seals. 41 of 138 were T4 seals counted twice (register-ICIT link gap).
  35. Ticked fish cluster by block (reg_findspot F3). Weakened: q = 0.114 over all 163 tests; not replicated elsewhere at Mohenjo-daro (p = 0.31).
  36. Connective M99 dropped first on small seals (T4). Did not replicate on register seals (+0.002, p = 0.41); not significant against the keep-closers null.
Impression of a one-line seal text
Pl. 5 · Mohenjo-daro · 36 × 17 mmSeal impression. Mackay, Further Excavations at Mohenjo-daro (1938), no. 461.

EOutputs

What can be published along the way

None of this needs a decipherment. Clean negatives and tools are worth publishing on their own.

  • A fair test bench for Indus sound keys Ready to write

    Paper and code. Five published readings, including the 185-sign key from 2026, score no better than their own values shuffled. The bench and its planted checks can be released for any new claim.

  • Round Gulf-type seals break Indus sign order, at home and abroad Needs one check

    Short paper. Passed two independent attacks. Before publishing, the shape of 5 of the 8 home seals needs checking, and a specialist should review the readings.

  • Structure of the script on unseen texts Ready to write

    Paper. Sign order predicted on held-out texts, the ticked fish that moves to the closing slot, and six plus fish as a fixed expression.

  • An atlas of closed routes Ready to write

    Open dataset. Over 80 tested routes, each with the size of effect its test could have caught, so nobody has to repeat them.

  • A small agent team with a shared task board Later

    Tool and write-up. The task board, the claim and red-team rules and the compute gate, as a reusable pattern for research with AI agents.