Indus-CoT: Chain-of-Inference Framework
- Indus-CoT is a cumulative framework that converts artifact images into structured grapheme sequences to analyze the undeciphered Indus script.
- It employs n-gram Markov chains, network analysis, and segmentation trees to reveal directional, positional, and recurrent structural features.
- The framework integrates image processing, corpus curation, and archaeological context to probe administrative practices, craft production, and measurement systems.
“Indus-CoT” (Editor’s term) denotes a chain-of-inference framework for research on the undeciphered Indus sign system in which artifact images are converted into grapheme sequences, those sequences are tested for directionality, positional classes, local dependencies, and recurrent subunits, and the resulting structural evidence is interpreted against archaeological context, especially administration, craft production, and measurement. The label is not a standard term in the literature, but the methodological architecture is explicit across work on automated corpus preparation, -gram and network analysis, genre-sensitive administrative modeling, and comparative scorecards for non-linguistic baselines (Palaniappan et al., 2017, 0901.3017, Sinha et al., 2010, Rao, 2018, Nair, 20 Apr 2026).
1. Conceptual scope and evidentiary base
Indus-CoT is best understood as a cumulative research program rather than a decipherment claim. Its point of departure is the severe constraint structure of the corpus: the Indus script remains undeciphered; there are no bilingual inscriptions; the underlying language is unknown; inscriptions are brief; and many texts are damaged or illegible. Under these conditions, the literature repeatedly shifts attention from semantic certainty to intermediate levels of inference: sign order, positional regularities, recurrent genres, artifact context, and production procedures (0901.3017, Rao, 2018).
A central feature of this program is that corpus definition itself is methodological. Different studies work with different cleaned or deduplicated subsets, and each subset supports a different type of inference. Mahadevan’s electronic concordance (M77), used for -gram analysis, records 417 unique signs in 3573 lines belonging to 2906 texts; its cleaned Extended Basic Unique Data Set (EBUDS) retains 1548 texts. Wells’s WUCS corpus, used for network analysis, contains 1821 unique complete single-line sequences with 593 signs. The ICIT/Yajnadevam digitization used for the 2026 synthetic-baseline scorecard analyzes 1916 deduplicated inscriptions, 584 unique sign types, and 11,110 total sign tokens (0901.3017, Sinha et al., 2010, Nair, 20 Apr 2026).
| Corpus/frame | Size reported | Principal use |
|---|---|---|
| M77 / EBUDS | 417 signs; 2906 texts; 1548 cleaned texts | -gram modeling and restoration |
| WUCS | 1821 sequences; 593 signs | Directed network and segmentation trees |
| ICIT/Yajnadevam | 1916 inscriptions; 584 signs; 11,110 tokens | Multi-metric non-linguistic scorecard |
This evidentiary base implies a distinctive epistemic order. Instead of beginning with phonetic values, Indus-CoT begins with representational hygiene, distributional structure, and object context. This suggests a research logic in which semantic proposals are downstream of image processing, corpus curation, and structural modeling, not prior to them.
2. Structural inference from sign sequences
The structural core of Indus-CoT is the claim that Indus inscriptions are not random strings. Yadav and collaborators model sign sequences with -gram Markov chains and show strong positional asymmetry, local dependence, and directionality. In EBUDS, approximately 82 text-beginner signs account for 80% of all text-beginner usage, while only 23 text-ender signs account for 80% of text-ender usage. Held-out perplexity drops sharply from unigram to bigram and only marginally thereafter: $68.82$ at , $26.69$ at , $26.09$ at , and 0 at 1 and 2. The paper therefore treats the bigram model as effectively optimal for the present corpus (0901.3017).
The basic probabilistic form is the boundary-aware bigram factorization
3
supplemented by Witten-Bell smoothing and Katz backoff to avoid zero-probability transitions in a sparse corpus. This supports restoration and constrained prediction: across five folds, mean sensitivity for restoring deleted signs under a 90% cumulative-probability criterion is 4 with standard deviation 5 (0901.3017).
Network analysis extends the same point from probabilistic modeling to graph structure. Using WUCS, the directed weighted sign network has 2719 observed ordered pairs out of 6 possible directed pairs, yielding connectivity 7. Reciprocity is low, 8, compared with 9 in randomized controls that preserve within-sequence co-occurrence and sequence length. Positional specialization is also strong: 21 signs appear only as solo inscriptions; 128 signs are beginners only; 43 are enders only; and 127 are always internal. Pair significance is quantified by
0
with 377 pairs satisfying 1 and 31 pairs satisfying 2 (Sinha et al., 2010).
A further step toward Indus-CoT is the segmentation-tree procedure. Adjacent pairs in a long inscription are merged in descending order of pair significance, yielding recursive trees that expose recurrent subunits. In the 13-sign inscription M-0355, the subsequences 3, 4, 5, 6, and 7 are identified as units that also occur elsewhere in the corpus. This does not prove natural-language syntax, but it does support a grammar-like compositional layer. A plausible implication is that Indus-CoT operates most effectively when it reasons over boundary classes, significant local templates, and recurrent subsequences rather than isolated signs (Sinha et al., 2010).
3. Genre-sensitive administrative reading
The strongest explicitly semantic layer in Indus-CoT is genre-sensitive rather than fully decipherment-based. Rao’s analysis of miniature tablets and stamp seals argues that a substantial subset of Indus inscriptions is best read as administrative notation tied to rationing, labor management, and exchange in a barter-based economy. The key archaeological trigger is the 1997 Harappa find published by Meadow and Kenoyer: 22 three-sided miniature tablets, with 16 forming a coherent cluster, each side bearing a short inscription, generally of two signs, found together in a clump dumped over a wall or outside a perimeter wall. Standardization, clumped disposal, and overlap with seal inscriptions are used to motivate an administrative function rather than a decorative or purely ownership-mark function (Rao, 2018).
The formal contribution is a positional grammar for “Patterned Texts.” Following Wells and Bonta, Rao partitions these texts into Prefix, Medial, Core, and Terminal classes and proposes the structural mapping
8
with the functional hypothesis
9
This is not presented as decipherment, but as a role-based structural model for a subset of inscriptions. In this model, Medial positions often contain numerals and fish signs; the sign 0 is treated, following Wells, as a possible volumetric measure of about 40 liters; and anthropomorphic compounds are cautiously interpreted as labor categories, especially the sign of a person carrying two bundles on a pole over the shoulders, which Rao treats as a strong clue for porter-type labor (Rao, 2018).
The same paper reframes the function of seals. Instead of reducing seals to owner-name devices, it proposes that many stamp seals were productive administrative instruments used to generate ration tokens, wage tokens, miniature tablets, sealings on goods, and exchange tokens. Evidence cited for this includes inscriptional overlap between tablets and seals, a two-sided Mohenjo-daro seal whose sides parallel tablet inscriptions, pendant-like tokens from Kanmer with identical “unicorn” impressions and holes, pierced bosses on stamp seals that may have allowed stringing, deliberately broken or discarded seals, and the absence of seals in burials in the expected personal-ownership sense (Rao, 2018).
This administrative layer yields a distinctive Indus-CoT inferential chain: observe standardization; note numeral-like signs; identify a possible metrological sign; compare with proto-Elamite and proto-Cuneiform ration texts; add iconographic support from anthropomorphic labor signs; situate the system within Harappan transport and exchange; and only then infer rationing, labor allocation, or exchange semantics. The resulting model is narrower than decipherment but more concrete than purely formal syntax.
4. Material and geometric priors in the Indus world
Indus-CoT is not limited to inscriptions. A broader archaeological substrate comes from work on measurement, design, and constructive geometry in the Indus Civilization. “In Square Circle: Geometric Knowledge of the Indus Civilization” argues that South Asian mathematical thinking, especially geometry, should not begin exclusively with the \emph{Sulbasutras} of roughly 800–500 BCE, but that significant geometric reasoning was already present in the Mature Harappan Civilization of ca. 2500–1900 BCE. The argument is cumulative: planned city layouts, streets aligned to cardinal directions, practical mensuration, standardized systems of length and weight, an ivory scale from Lothal with 27 uniformly spaced lines over 1, a compass-like device inferred by Mackay from Mohenjo-daro evidence, and shell ring objects with four slits from Lothal that may have served for angle measurement (Sinha et al., 2011).
The paper’s distinctive move is methodological. It treats recurring decorative motifs on seals, pottery, and ornaments as traces of geometric operations that can be reverse-engineered. The most important example is the Mohenjo-daro seal M-1261, whose space-filling fan-like design is reconstructed from a rectangle of dimensions 2 containing two non-overlapping circles of radius 3. Because the tile is obtained by cutting and repositioning regions outside the circles, area is preserved: 4 and the four congruent fan-shaped constituents therefore each have area
5
The claim is not that a formal area theorem survives in text, but that curved-edge tiling, quarter-turn rotation, translational periodicity, modular proportion, and area-preserving rearrangement were operationally available to Indus artisans (Sinha et al., 2011).
The same constructive logic is applied to intersecting-circle lattices, imbricate or fish-scale patterns, and star-like or quincross motifs derived from circle overlap and square framing. The repeated use of simple relations—6 rectangles, offsets by 7, squares of side 8, fourfold rotational symmetry—supports the paper’s claim of a “deep understanding of the properties of circular shapes.” Continuity with later Vedic geometry is explicitly not demonstrated, but the existence of a sophisticated constructive geometry centered on circles is taken as evidence that visual and material reasoning in the Indus world was already highly developed (Sinha et al., 2011).
For Indus-CoT, this matters because it supplies non-textual priors about workshop procedure, measurement culture, and rule-based design. A plausible implication is that epigraphic reasoning in the Indus domain should be situated within the same artisanal environment that produced exact metrology, modular construction, and repeatable transformations.
5. Perceptual grounding and automatic corpus preparation
A full Indus-CoT architecture requires a perceptual front end, since most computational work presupposes already prepared sign strings. “Deep Learning the Indus Script” addresses this bottleneck by presenting an OCR-style pipeline that takes as input images of archaeological artifacts and returns as output a string of graphemes suitable for inclusion in a standard corpus. The pipeline proceeds through seal extraction, Selective Search region proposals, heuristic region grouping, CNN-based classification of proposals into text, no-text, or both, text-box formulation and trimming, symbol segmentation, and final sign classification (Palaniappan et al., 2017).
The first learned component is a transfer-learned GoogLeNet region classifier with three output classes: text, no-text, and both. It operates on 9 crops, uses ImageNet-pretrained BVLC GoogLeNet, freezes inception layers $68.82$0, doubles learning rates in $68.82$1, and reaches 89.30% validation accuracy on a manually labeled Text-NoText dataset of 2091 proposals. The second learned component, SymbolNet, is a binary Jar/NoJar classifier operating on $68.82$2 grayscale inputs, with two convolutional layers, a dropout layer, two fully connected layers of 500 and 2 outputs, and a SoftMax classifier; it reaches 92.07% validation accuracy on the Jar-NoJar dataset (Palaniappan et al., 2017).
At the pipeline level, results are partial but concrete. On 50 held-out seal images, 43 yielded complete text regions, corresponding to 86% “perfect case accuracy” for text-region extraction. Symbol segmentation was perfect in 34 of 50 cases, or 68%, and rose to 94% when “fair” outputs were included. The authors are explicit that the final stage is not full transcription across the 417-sign Mahadevan inventory; it only detects the most frequent sign, the “jar” sign, Mahadevan sign 342. The paper is therefore best categorized as a front-end perceptual module and corpus-preparation support system, not as decipherment or sequence reasoning (Palaniappan et al., 2017).
Within an Indus-CoT framework, this front end has a clear role. It grounds downstream structural inference in artifact images, separates text from iconography, and creates ordered symbol crops that can later be scored by $68.82$3-gram, network, or genre-sensitive models. The paper also states, with equal clarity, what remains missing: full sign-inventory recognition, reading-order inference, uncertainty-aware decoding, sequence models, syntax or semantics, and artifact-metadata integration.
6. Comparative scorecards, maximalist decipherments, and unresolved controversies
Indus-CoT remains controversial because structural regularity does not by itself settle the linguistic status of the sign system. The 2026 scorecard paper responds directly to that problem by comparing the Indus corpus against synthetic heraldic and administrative non-linguistic baselines on four metrics derived from the Farmer-Sproat-Witzel critique: text brevity, repeated formulaic phrases, hapax rate, and positional rigidity. On 1916 deduplicated inscriptions, the paper reports a Zipf slope of $68.82$4 and conditional entropy of 3.23 bits, replicating earlier evidence for non-random sequential structure; but its main claim is comparative. Across the four scorecard metrics, Indus occupies an intermediate position relative to the two baseline families and matches neither cleanly. No tested real-world non-linguistic corpus reproduces the full statistical profile either (Nair, 20 Apr 2026).
That result is deliberately limited. It raises the burden on simple non-linguistic explanations, but it does not prove that the system encodes spoken language. This is an important correction to a common misconception: “structured sign system showing features of a formal language” is not equivalent to “demonstrated natural language.” The earlier $68.82$5-gram paper makes the same point explicitly, and the scorecard paper confirms that single metrics such as conditional entropy or positional rigidity are insufficient on their own (0901.3017, Nair, 20 Apr 2026).
At the opposite end of the interpretive spectrum lies the maximalist rebus program of “Indus script corpora, archaeo-metallurgy and Meluhha (Mleccha).” That paper treats the script as a full writing system of “Indian hieroglyphs,” read by rebus against a hypothesized Meluhha/Mleccha lingua franca of the Indian sprachbund, with a dominant semantic field of archaeo-metallurgy, workshop production, and trade administration. Its inferential pipeline is explicit: archaeological artifact context $68.82$6 glyph identification $68.82$7 pictured object name $68.82$8 lexeme from the Indian sprachbund $68.82$9 rebus transfer 0 domain assignment 1 transactional reading. Examples include elephant 2 ibh 3 ib 4 iron, swastika 5 sathiya 6 jasta 7 zinc, and one-horned young bull 8 kodiyum/kodiya 9 kod $26.69$0 workshop or smithy (Kalyanaraman, 2012).
The paper is methodologically important because it offers a fully articulated decipherment workflow, but it is also the clearest case where caution is required. The same source notes the breadth of the lexical search space, the use of many later language stages, the heavy semantic funnel toward metallurgy, and the relative lack of structural epigraphy. A cautious reading therefore distinguishes between the value of its inferential architecture and the much stronger claim that the proposed rebus readings are established (Kalyanaraman, 2012).
A further comparative line comes from the 2025 thesis on visual similarity between the Indus script and Tibetan-Yi Corridor writing systems. Using a hybrid CNN-Transformer ensemble, it reports that Tibetan-Yi Corridor scripts exhibit approximately six-fold higher visual similarity to the Indus script, 0.635, than to Proto-Cuneiform, 0.102, or Proto-Elamite, 0.078; in an alternate summary scale, the Indus script maps closer to Tibetan-Yi Corridor scripts with a mean cosine similarity of 0.930 than to Proto-Cuneiform, 0.887, or Proto-Elamite, 0.855. The thesis explicitly states that this is not decipherment, not proof of phonetic connection, and not a claim that Dongba is a direct descendant of the Indus script. Its strongest defensible claim is visual-morphological affinity and a possible corridor-mediated cultural transmission hypothesis (Reddy, 27 Mar 2025).
Taken together, these debates define the current limits of Indus-CoT. The framework is strongest when it remains modular: image grounding, structural modeling, genre-sensitive interpretation, and archaeological contextualization. It is weakest when one module is allowed to substitute for the others—when local syntax is treated as decipherment, when visual similarity is treated as genealogy, or when a rebus lexicon is treated as if corpus-wide structural constraints had already been solved. The literature therefore supports Indus-CoT as a disciplined architecture for inference under radical uncertainty, not as a settled reading of the Indus script.