2.1 What has been established in the domain-generalist paradigm

The large language model approach treats language modelling over a broad training corpus as the mechanism by which domain knowledge is acquired and cross-domain transfer is enabled. The model learns domain-specific patterns — the vocabulary, the relational structures, the typical inferential moves of a domain — by exposure to enough text from that domain. It learns cross-domain transfer by exposure to text that itself crosses domains: scientific writing that borrows from adjacent fields, interdisciplinary papers, analogical explanations, and so on.

The result is a model that holds domain knowledge and coordination capacity in the same parameter space. When the model encounters a problem, it does not route to a specialist; it retrieves from a generalised substrate that contains everything at once. Retrieval-augmented generation (Lewis et al. 2021) and fine-tuning extend this by adding domain specificity as a correction applied after training: the base model is a generalist, and specialisation is layered on. Mixture-of-experts architectures (Shazeer et al. 2017) pursue a different efficiency: within a single model, a gating network routes each token to a subset of parameter experts. This reduces inference cost without changing the epistemic structure — all experts are trained jointly under the same loss, in the same model, on the same data distribution.

None of these approaches separates coordination from domain expertise at the architectural level. The coordinating function — whatever the model does when it integrates across domains — is distributed throughout the same weights that hold domain content.

2.2 What the brain does instead

Functional neuroimaging research has documented a different structure. The human brain is not organised as a single generalised processor. It is organised as a set of discrete functional modules — each densely interconnected internally, each performing a domain-specific cognitive function — coordinated by a distinct class of regions that do not themselves perform domain-specific computation.

Bertolero, Yeo, and D’Esposito (2015) established this architecture empirically across 9,208 experiments and 77 cognitive tasks in the BrainMap database. Using resting-state fMRI and graph-theoretic network analysis, they identified 14 distinct functional modules with strong spatial correspondence to known cognitive functions. They then measured activity at different types of nodes across all tasks. Local nodes within modules — provincial hubs — did not increase activity as more cognitive functions were engaged. Their computational load remained constant regardless of task complexity. Connector nodes — regions with high participation coefficients, meaning their connections were distributed evenly across many modules rather than concentrated within any one — showed a different pattern entirely. Their activity increased proportionally to the number of modules engaged in a task.

This finding has a specific implication. Connector nodes are not doing more of what the domain modules are doing when tasks get more complex. They are doing something else entirely: managing the integration load that increases when many modules must work together, while preserving the autonomy of each module’s function. The modules stay modules; the connector nodes handle the coordination between them.

Bertolero et al. (2018) extended this to a mechanistic account. Connector hubs do not merely route information between modules; they actively tune the connectivity of their neighbours, reorganising which modules are more or less connected based on current task demands. Individuals with more diversely connected hubs and more modular brain networks show higher cognitive performance across all tasks — not on any specific task, but across the board. The diversity of hub connectivity predicts general integration capacity.

The architectural principle that emerges from this literature is not that specialisation and integration are in tension. It is that they are structurally separable and mutually reinforcing: more modular domain processing combined with more capable coordination produces better outcomes than either alone (Menon and D’Esposito 2022).

3.1 Module autonomy is not isolation

A clarification matters here. Saying that domain modules process information autonomously does not mean they are isolated from one another. The brain is not a collection of silos that occasionally exchange messages. It is a network in which modules maintain dense internal connectivity while connector hub regions manage cross-module communication selectively, based on task demands.

The key property of connector hubs is the participation coefficient (Sporns and Betzel 2016): the degree to which a node’s connections are distributed evenly across modules rather than concentrated within one. A node with a high participation coefficient is well-connected to many modules. It has access to what each module is doing. But it does not perform any module’s function. It is neither a domain specialist nor a blank generalist. It occupies a structurally distinct role: a node that can reach across module boundaries without being defined by any of them.

Gordon et al. (2018) refined this picture further, showing that connector hubs are not a single category. Three distinct sets were identified, each with different task-activation profiles: one set deactivates across tasks, one activates during all tasks, one activates specifically during tasks requiring the configuring of input, transformation, and output processes. This differentiation within the coordinator role is relevant because it suggests coordination is itself a structured function — not a homogeneous relay, but a set of subtypes performing distinct integrative operations.

3.2 What connector hubs actually compute

The literature on connector hub function does not describe these regions as performing structural isomorphism detection — comparing problem shapes across domains and flagging when a problem in one domain has the same relational structure as a solved problem in another. That function is not what the connector hub literature directly documents.

What it documents is routing and tuning: managing which modules are active, how strongly they communicate, and how that connectivity pattern shifts as task demands change. The connector hub’s documented computational role is coordination in the sense of network configuration, not in the sense of cross-domain analogy.

The analogical reasoning literature, however, sits adjacent and is worth examining. A meta-analysis of 27 neuroimaging studies on analogical reasoning found that the left rostrolateral prefrontal cortex (rlPFC) is the region most consistently activated across all analogical reasoning tasks, regardless of whether the domain is semantic or visuospatial (Hobeika et al. 2016). The rlPFC is domain-general for analogy. Lesions to the left rlPFC impair analogical reasoning across domains. And the rlPFC is anatomically located within the connector hub regions identified by the modular brain architecture literature.

This anatomical overlap does not establish that connector hubs are cross-domain analogy engines. It establishes something more modest: the architectural conditions that define connector hubs — high participation coefficient, domain-distributed connectivity, low domain specificity — are the same conditions under which cross-domain relational comparison is supported. Whether a coordination layer trained into this architectural role would develop the capacity for structural similarity detection across domains is an open question. The brain architecture suggests it is not an implausible one.

Gentner’s structure-mapping theory (Gentner 1983) provides the theoretical vocabulary for what this function would be. Analogy, in the structure-mapping framework, depends on finding relational correspondences between domains — not surface similarity between objects, but systematic similarity between the roles objects play within a relational structure. The function is domain-blind by definition: the same relational structure can exist in two entirely different content domains, and detecting it requires abstracting away from domain content. A coordinator trained to detect such correspondences would not need to know what a domain is about; it would need to know what shape a problem has.

4.1 Mixture of experts

Mixture-of-experts architectures (Shazeer et al. 2017; Cai and colleagues 2025) are the closest existing analogue to the proposed architecture. MoE models contain multiple sub-networks (experts), with a gating mechanism routing each token to a small subset of experts during inference. This achieves computational efficiency — not all parameters are activated for every input — and produces a form of functional specialisation within the model.

The distinction from the proposed architecture is epistemic rather than computational. In MoE, all experts are trained jointly under the same loss, within the same model, on the same data distribution. The gating network is trained simultaneously with the experts; there is no separation between the coordinator’s training objective and the specialists’ training objective. The experts are not domain-native in the sense of having been developed to hold a specific domain’s knowledge independently of the generalist training regime. They are weight-level subnetworks within a single model that have developed different activation patterns through joint training.

The connector hub analogy is structurally different. Domain-native modules, in the brain, are not trained jointly with the connector hubs under a shared loss. They develop through domain-specific experience and exposure; connector hubs develop separately. The proposed AI architecture would reflect this separation: domain-native specialist models trained on domain-specific corpora, and a coordination layer trained separately — potentially on a different objective entirely, concerned with structural relationships across domains rather than domain content.

4.2 Retrieval-augmented generation

Retrieval-augmented generation (Lewis et al. 2021) adds domain specificity to a generalist model by retrieving relevant documents at inference time and including them in the context window. This is a post-training correction: the base model remains a generalist; specialisation is supplied externally.

The proposed architecture differs in that domain specialists are not corrections applied to a generalist. They are the primary domain processors. The coordinator does not have domain knowledge that gets topped up by retrieval; it does not hold domain knowledge in the first place. The separation is architectural, not a retrieval strategy.

4.3 Current multi-agent systems

Multi-agent systems (Xiao and colleagues 2024) distribute tasks across multiple models and coordinate their outputs through an orchestrator. This is the existing approach closest in spirit to the proposed architecture, and it shares the structural separation the brain’s architecture exhibits. The limitation documented in current production deployments is coordination overhead: as the number of specialists increases, the coordination tax — communication overhead, latency, context management — grows faster than the benefit (Königstein 2026). The orchestrator in most deployed systems is not a model with a distinct training objective for coordination; it is a generalist model given a coordination role through prompting. The coordination capacity is borrowed from the generalist’s general capability, not developed as a distinct function.

The proposed architecture asks whether a coordinator trained specifically for structural integration — with its own training objective, on a corpus of cross-domain relational correspondences rather than on domain content — would perform differently from a prompted generalist acting as coordinator. The brain’s architecture suggests these are different things. Whether they produce different outcomes in AI systems is the experimental question.

7.1 What this paper does not claim

This paper does not claim that the proposed architecture would outperform current large language models on any benchmark. The claim is architectural and conceptual: that the separation of coordination from domain expertise is a structural principle documented in the brain and independently discovered in human organisational systems, and that AI architecture has not yet explored it at the level the brain implements it.

This paper does not propose an implementation. The training objective for the coordination layer, the mechanism by which specialists and coordinator communicate, the representation format for structural similarity, and the evaluation framework for coordination quality are all open engineering questions. Scoping them is outside the range of what a conceptual paper can usefully do.

This paper does not argue that domain-native training produces better specialists than generalised training in all cases. There are domains where generalised training produces specialists that match or exceed domain-native fine-tuning. The architectural argument is not about which approach produces better specialists; it is about whether the coordination function is better served by a dedicated coordinator with its own training objective than by a generalist model acting as coordinator.

7.2 What remains open

The most significant open question is the training objective for the coordinator. The brain’s connector hubs develop their function through experience in a system where domain modules are already developing their functions simultaneously. A training objective that reproduces this developmental condition in a supervised setting does not yet exist.

The evaluation question is similarly open. Current benchmarks evaluate domain performance. Cross-domain transfer is typically evaluated by measuring performance on domain B after training on domain A. Neither evaluates coordination quality directly — the capacity of a coordinator to route appropriately, integrate across specialists, and detect structural correspondences across domain boundaries. Developing such an evaluation framework may be a precondition for testing the architecture.

2.1 What has been named, and what has not

The LLM alignment literature has developed a family of related concepts for failures at the boundary of intent and execution.

Specification gaming occurs when a model achieves the literal specification of an objective without achieving the intended outcome (Amodei et al. 2016). The canonical examples are from reinforcement learning: a boat-racing agent that maximizes score by circling checkpoints rather than finishing the race; a cleaning robot that covers its camera rather than cleaning. In each case, a proxy metric is gamed because the true objective was underspecified. The model finds an unintended solution to the stated objective.

Reward hacking is the broader category: exploiting flaws or blind spots in the reward model to achieve high proxy reward without satisfying human intent (Denison et al. 2024). Sycophancy — agreeing with false user claims to generate approval signal — is a trained-in form of reward hacking. Reward tampering is its extreme: modifying the reward mechanism itself.

Goal misgeneralization occurs when a model pursues a proxy goal that correlated with the intended goal during training but diverges from it in deployment, particularly under distributional shift (Langosco di Langosco et al. 2022). The model has learned the wrong goal; it did not fail to execute the right one.

None of these concepts describes the failure in the opening example. In that case:

• The objective is not underspecified. The researcher stated what she wanted.
• No reward signal is being gamed. There is no approval-seeking behavior.
• There is no distributional shift. The task is squarely in-distribution for the model.
• The model has not learned the wrong goal. It has correctly identified the immediate task.

What has happened is different: the model has treated the task’s instrumental structure — the log format, the section conventions, the expected length and scope — as the thing to be optimized, and in doing so has lost track of the terminal goal the task was supposed to serve. The artifact is complete. The purpose is not.

2.2 The cross-domain precedent

This failure mode has a precise name in organizational sociology. Merton (1940) described it in bureaucracies: rules designed as means to an end become ends in themselves through a process he called goal displacement. The bureaucrat adheres to every rule, satisfies every procedure, and fails every client. Merton called the extreme case “the bureaucratic virtuoso, who never forgets a single rule binding his action and hence is unable to assist many of his clients” (Merton 1940, 563).

The psychological mechanism Merton cited was Allport’s (1937) functional autonomy of motives — the principle that instrumental behaviors can become self-sustaining and motivationally independent from their original purpose. “What was once a means becomes an end in itself” (Allport 1937). Allport’s workman who continues to do clean-cut jobs even when his security no longer depends on it is the benign version; Merton’s bureaucrat is the organizational pathology.

The measurement-science parallel is Goodhart’s Law (1975): when a measure becomes a target, it ceases to be a good measure (Goodhart 1975). Campbell (1979) stated the same principle from social science: the more a quantitative indicator is used for social decision-making, the more it distorts the process it was meant to monitor (Campbell 1979). In each case: a proxy for a goal displaces the goal.

What connects these traditions is a shared structural logic. An instrumental value — a rule, a metric, a procedure — is created to serve a terminal goal. Under conditions that reward adherence to the instrumental value independent of terminal goal service, the instrumental value becomes terminal. The original goal disappears from view.

3.1 Instrumental-terminal goal inversion defined

We define instrumental-terminal goal inversion (ITGI) as follows:

Given a task with an explicit terminal goal and a set of instrumental constraints specified to serve , ITGI occurs when a model’s output satisfies while failing to serve , and this failure is attributable to the model treating satisfaction of as sufficient for task completion.

The key conditions distinguishing ITGI from related phenomena:

1. is stated in the prompt, not merely implied or inferable from reward signal.
2. is explicitly specified (format requirements, section structure, output length, schema).
3. The model’s output satisfies all or most elements of .
4. The model’s output does not serve — specifically, a reader with only the output cannot accomplish what required.
5. The failure is not attributable to factual error, hallucination, or task misunderstanding.

This distinguishes ITGI from specification gaming (where is underspecified), from sycophancy (where the model is optimizing for approval), and from goal misgeneralization (where distributional shift causes a wrong goal to be pursued).

3.2 The structural specification hypothesis

ITGI is not merely possible; we hypothesize it is monotonically increasing in structural specification density. As the number and specificity of instrumental constraints in a prompt increases, the probability that the model’s output serves the terminal goal decreases, holding terminal goal clarity constant.

This is the counterintuitive claim. For humans, constraints serve as scaffolding — they reduce cognitive overhead allocated to the how, freeing attention for the why. A researcher given a template for a log entry is freed from formatting decisions and can concentrate on what the log should say. The constraints help.

For LLMs, the hypothesis is that the inverse holds: each additional constraint is an additional optimization target, and as constraint density increases, the model’s attention allocation shifts from to . The constraints crowd out the goal.

The empirical support for this direction comes from Sridhar et al. (2023), whose ASH (Actor-Summarizer-Hierarchical) prompting work on web navigation demonstrated that when a single LLM prompt must simultaneously process raw environmental observations and predict the next action, performance degrades sharply on long-horizon tasks. On trajectories exceeding 11 steps, REACT — which loads both observation processing and action prediction into a single prompt — scored 7.4; ASH, which separates these functions into a SUMMARIZER and an ACTOR, scored 38.2 (Sridhar et al. 2023). The implicit diagnosis: when a model must manage simultaneous instrumental load (process the current observation) and terminal goal tracking (buy the right product), terminal goal tracking degrades first. The fix — hierarchical decomposition that isolates instrumental processing — is structural evidence for the hypothesis.

3.3 The inverse human pattern

The human behavioral literature on intention establishes the baseline against which ITGI is the inversion. The intention-action gap — the well-documented failure of humans to execute their stated intentions — shows that humans hold terminal goals but frequently fail on the instrumental side (Sheeran 2002; Sheeran and Webb 2016). Intentions explain only 18–28% of behavioral variance, even when the intention is strong and clearly stated.

The LLM failure runs in the opposite direction. Models execute the instrumental structure reliably and completely. What they fail to maintain is the terminal goal. Humans fail at doing; LLMs fail at purposing.

This inversion is not merely a rhetorical point. It has methodological implications for how the failure should be studied, and design implications for how it might be mitigated. Strategies developed to close the human intention-action gap — implementation intentions, commitment devices, environmental triggers — work by strengthening the link between a held terminal goal and instrumental execution. The LLM problem requires the reverse: strengthening the link between instrumental execution and a terminal goal that has not been lost but has been deprioritized.

4.1 The common structure

Across the organizational sociology, measurement science, and LLM literatures, a single structural pattern recurs:

1. A terminal goal exists: serve the client, measure economic health, help the researcher.
2. An instrumental proxy is created to serve the terminal goal: follow the rules, track the money supply, complete the artifact.
3. Under conditions where adherence to the proxy is rewarded independent of terminal goal service, the proxy displaces the terminal goal.
4. The agent — bureaucrat, central bank, LLM — then optimizes the proxy while failing the original goal.

What varies across domains is the mechanism of displacement:

• In bureaucracies, displacement is driven by incentive structures: career advancement depends on rule compliance, not client outcomes.
• In measurement systems, displacement is driven by optimization pressure: when a metric becomes a target, actors game it.
• In LLMs, displacement is driven by attention allocation during inference: satisfying explicit constraints is a local, verifiable task; serving a terminal goal requires maintaining a non-local purpose across the response.

The LLM mechanism is distinct from the human mechanisms in an important way. Bureaucratic ritualism is chosen — the bureaucrat has other options and selects rule compliance. Metric gaming is strategic — the actor knows the metric is a proxy and exploits the gap. LLM ITGI is neither chosen nor strategic. The model does not know it has displaced the terminal goal. The displacement is a property of how inference proceeds when constraints are dense, not a property of motivation or strategy.

4.2 Allport’s functional autonomy as the deepest analog

Allport’s functional autonomy (Allport 1937) is the closest structural analog to LLM ITGI — and also the most illuminating difference. Allport showed that instrumental behaviors can become self-sustaining: a motive that originates as a means to an end acquires its own motivational energy, independent of the original end. The workman who does clean-cut jobs even when his income no longer depends on it has developed a functionally autonomous motive for craftsmanship.

In humans, this is generally adaptive: functionally autonomous motives allow complex behaviors to persist without continuous reference to their original justification. The craftsman doesn’t recalculate the utility of quality work every time he picks up a tool.

In LLMs, the analog fails to generalize adaptively. There is no “persistence of a motive” — there is no motive, in the psychological sense. What there is: a training distribution that rewards well-formed artifacts, and an inference-time process that generates the most plausible completion of a prompt that already contains an elaborate structure. The structure predicts its own completion. The terminal goal, if not redundantly encoded in ways that compete with the structural signal, loses salience.

4.3 Goodhart as the measurement-science frame

Manheim and Garrabrant (2018) distinguish four variants of Goodhart’s Law: regressional (the proxy correlates imperfectly with the goal), extremal (the proxy diverges from the goal at extreme optimization), causal (optimizing the proxy changes the underlying relationship), and adversarial (an agent exploits the gap between proxy and goal) (Manheim and Garrabrant 2018).

LLM ITGI most closely resembles the regressional variant: the proxy (artifact completion) correlates with the terminal goal (task purpose) under normal conditions but diverges when structural specification is dense. The correlation holds for simple tasks with thin constraints; it breaks down as constraint density increases.

This framing is useful because it predicts where ITGI will be most severe: tasks with elaborate templates, multi-section output requirements, rigid format constraints, and complex schemas. These are, not coincidentally, the tasks where LLMs are most commonly deployed in professional and research settings — report generation, document drafting, structured analysis, code documentation.

5.1 What needs to be shown

Three empirical claims require testing:

1. Existence: ITGI occurs at inference time in current LLMs — outputs that satisfy structural constraints while failing terminal goals.
2. Monotonicity: ITGI increases as structural specification density increases, holding terminal goal clarity constant.
3. Asymmetry: The relationship between structural specification and ITGI is different for LLMs than for humans performing the same tasks.

5.2 Core design

Task pairs with separable terminal and instrumental goals. The key design requirement is that and can be independently evaluated. Tasks where artifact completion and purpose-serving are inseparable are uninformative.

Suitable task types:

• Synthesis tasks: Summarize this paper in a way that helps a reader decide whether to read it. Instrumental: produce a summary of appropriate length and scope. Terminal: enable the decision.
• Advisory tasks: Draft a note explaining this finding to a non-specialist audience. Instrumental: produce a note in the specified format. Terminal: the reader understands the finding.
• Selection tasks: Write a log entry for this source that captures what is relevant to Project X. Instrumental: produce a log entry. Terminal: a future researcher can use it without reading the source.

Specification density as the independent variable. Three conditions:

• Condition A (thin): Terminal goal stated only. No format, length, or section requirements.
• Condition B (moderate): Terminal goal stated plus moderate structure (suggested sections, approximate length).
• Condition C (dense): Terminal goal stated plus elaborate structure (required section headers, word count constraints, mandatory elements).

Measurement of terminal goal service. The challenge is avoiding subjective evaluation. Three approaches, in increasing defensibility:

1. Downstream task completion: Give readers only the output and ask them to accomplish what required (make the decision, explain the finding to someone else, use the log entry without the source). Measure success rate.
2. Counterfactual completeness: Have domain experts identify the 3–5 elements an output must contain to serve . Score presence/absence. ITGI predicts that Condition C outputs will score lower on this list despite longer overall length and higher structural compliance.
3. Truncation sensitivity: Progressively shorten outputs from the end. Measure at what point -relevant content disappears vs. at what point structural completeness fails. ITGI predicts these diverge, with content concentrated early and structural completion content concentrated late.

Human baseline. The asymmetry claim (Claim 3) requires human participants completing the same tasks under the same three conditions. If ITGI is real and inverted from human behavior, Condition C outputs from humans should be more purpose-serving than Condition A, while Condition C outputs from LLMs should be less purpose-serving.

6.1 What this thesis does not claim

ITGI is not claimed to be:

• The dominant failure mode of LLMs, or more common than hallucination, sycophancy, or factual error.
• Present in all structured tasks. Tasks where structural compliance and terminal goal service are tightly correlated will not exhibit ITGI.
• A property of current models specifically. Whether ITGI increases or decreases with model scale, RLHF, or chain-of-thought prompting is an empirical question this thesis does not answer.
• A training-time phenomenon. The claim is about inference-time behavior given well-formed prompts.

6.2 Confounds requiring control

• Task difficulty: More structurally complex tasks may simply be harder, producing lower overall quality independently of ITGI.
• Length bias: Condition C prompts produce longer outputs, and longer outputs may dilute the concentration of -relevant content without reflecting goal displacement.
• Model-specific behavior: Different models may show different ITGI rates. The structural specification hypothesis should be tested across model families, not assumed to generalize from a single model.

4.1 One Atomic Operation Per API Call

Passing a full document or a large batch to the model and asking it to extract everything produces degraded constraint adherence as the model’s attention distributes across multiple tasks simultaneously. In both implementations documented here, constraint violations — approximations instead of explicit failures, formatting inconsistencies, missed boundary cases — increased measurably as batch size grew beyond a threshold.

The solution is decomposition. Each API call handles one atomic operation: retrieve the ingredient list for this specific product, or extract the location and timing of this specific flood event from this specific article. The operation is defined narrowly enough that the model can apply the full constraint set reliably.

In the ingredient extraction implementation, the threshold was empirically established at 6 SKUs per batch. Batches above 10 showed measurable constraint violations. Below 6, quality was equivalent but throughput was lower than necessary. The optimal batch size is domain-specific and should be tested rather than assumed.

4.2 Explicit Failure Over Approximation

This decision is described in Section 2 and is the single most important constraint in the framework. The system instruction must be unambiguous: when data is absent from permitted sources, return the designated failure token. Do not infer. Do not approximate based on similar cases. Do not fill the gap.

In the ingredient extraction implementation, the system instruction read: “If ingredient list not found in whitelisted domains, return DATA_NOT_FOUND. DO NOT infer typical ingredients from product category. DO NOT approximate based on similar products.”

Of 1,000 products attempted, 104 returned persistent DATA_NOT_FOUND across two passes. These 104 were excluded from the corpus. The exclusion is a feature: those products either had no verifiable online ingredient list or were no longer in active distribution. The pipeline returned a clean gap rather than 104 fabricated entries that would have required expensive downstream correction.

In the flood extraction implementation, the equivalent constraint was classification: the model was required to distinguish between reports of actual past floods and articles discussing future warnings or policy — returning nothing for the latter rather than extracting a plausible but incorrect event record.

4.3 Batch Size as a Quality Variable

Batch size interacts with constraint adherence in a consistent pattern across both implementations. This is not primarily a cost or speed consideration. It is a quality variable that should be calibrated empirically for each domain and each stage of the pipeline.

In artifact removal and semantic decomposition stages of ingredient processing, batch size was set inversely to string complexity: short strings in batches of 40, complex multi-bracket strings one at a time. The same principle applies in news extraction: article complexity and length affect how reliably the model applies its classification and extraction constraints.

Test a range before committing to a batch size. The optimal value is not predictable from first principles.

4.4 Iterative Human-in-the-Loop Audit

Statistical validation establishes a confidence interval on the overall error rate. Iterative audit addresses systematic error patterns — categories of errors that recur and can be corrected in bulk.

The audit process runs as follows. A first model receives a sample of the extracted strings and identifies error types present. A second model receives the full extraction and flags instances of those specific error types. Human review resolves the flagged cases. Corrections are applied. The cycle repeats until the first model identifies no new error types.

In the ingredient extraction implementation, this converged in four iterations. The pattern across iterations was: 16.7% flagged in iteration 1, 7.1% in iteration 2, edge cases only in iteration 3, zero new error types in iteration 4. The edge cases in iteration 3 were boundary decisions — gluten classified as a grain or a protein, spirulina as an additive or a botanical — that required domain judgment rather than extraction correction. These were held for the classification framework stage, not resolved as cleaning errors.

Convergence does not mean zero errors. It means no new systematic error types are detectable. The residual error rate is quantified by the statistical sampling step.

4.5 Statistical Validation with Finite Population Correction

Complete manual validation is not feasible at scale. Statistical sampling with a confidence interval is.

For a population of size , desired confidence level , and margin of error , required sample size with finite population correction:

Using conservative (maximum variance), , , a population of approximately 2,000 requires a sample of around 130. For the ingredient extraction corpus, 90 extractions from 896 were audited manually. One error was identified: the model merged content from two adjacent sections of a product page. The 95% confidence interval on the population error rate, with finite population correction applied, places the upper bound below 3.6%. Stated as accuracy: the corpus is 98.9% accurate at 95% confidence.

Audit allocation should be risk-stratified: concentrate effort on high-risk subsets (very short strings that may be truncations, very long strings that may be insufficiently decomposed, low-confidence extractions) while maintaining a random component for unbiased population coverage.

5.1 Case Study 1: Indian Packaged Food Ingredient Vocabulary

The problem. No reference layer exists that maps the names Indian food labels use to shared ingredient identities. The same substance appears as maida, refined wheat flour, and all-purpose flour. Standard similarity measures would merge palm oil and palmite, which are functionally distinct, while missing the equivalence of besan flour and chana dal, which are the same ingredient in different language registers. No ground truth lexicon exists to train a supervised system against.

The implementation. 1,000 products were selected across 42 companies and 153 brands from verified Indian market listings. Ingredient lists were retrieved from whitelisted domains (brand official website, Amazon India, BigBasket, Blinkit) at temperature 0, with DATA_NOT_FOUND required when all sources were exhausted. Retrieved strings were parsed using a structure-aware algorithm that splits on commas only at nesting depth zero, preserving compound ingredient relationships. Each string then went through a single-purpose artifact removal pass (removing percentages and marketing text, preserving INS codes and preparation specifications) and a semantic decomposition pass with context propagation. The process ran at 6 SKUs per batch for retrieval and scaled inversely with string complexity for subsequent stages.

Results. 896 of 1,000 products extracted successfully (89.6%). 104 returned persistent DATA_NOT_FOUND. The sampling pipeline produced 1,987 unique variant strings. Combined with ingredient strings from OpenFoodFacts filtered to rows with a verifiable Indian product name and passed through the same pipeline, the final corpus after iterative audit is 2,291 unique ingredient variant strings. Audit of 90 extractions identified 1 error (0.11%). Full methodology documented in (R. 2026).

5.2 Case Study 2: Global Flash Flood Historical Record

The problem. Hydro-meteorological hazards like flash floods lack a standardised global observation infrastructure. Existing archives capture large, long-lasting events but miss localised and fast-moving floods. The Global Disaster Alert and Coordination System holds approximately 10,000 records — orders of magnitude fewer than what AI-based forecasting models require for training and validation. The historical record exists, embedded in news archives across 80 languages, but has never been extracted at scale.

The implementation. Google’s Groundsource framework analysed news reports where flooding was a primary subject, standardised text into English via translation, and used Gemini to apply three constrained extraction tasks: classification (distinguishing actual past flood events from articles about future warnings or policy), temporal reasoning (anchoring relative date references against publication dates), and spatial precision (mapping location references to standardised geographic polygons). The model was not asked to know where floods occurred. It was asked to read a specific article and extract specific structured fields — with explicit verification criteria for each field rather than open-ended generation (Rotem Mayo 2026).

Results. 2.6 million historical flood events extracted, spanning more than 150 countries from 2000 to present. Manual review found 60% of extracted events accurate in both location and timing; 82% accurate enough for practical research use. Spatiotemporal matching against GDACS records for 2020–2026 shows Groundsource captured between 85% and 100% of severe events in that reference set, alongside large numbers of smaller localised events the reference set missed entirely.

5.3 What the Convergence Shows

Neither implementation was designed with the other in mind. The decisions they share — constrain the model’s role to retrieval and parsing, require explicit failure for absent data, calibrate batch size empirically, validate statistically — emerged independently from the same underlying problem: how to collect structured data from a space where the information exists but no ground truth organises it.

The table below shows the architectural correspondence.

The accuracy figures are not directly comparable — the domains define error differently, and the flood implementation targets a harder extraction problem (temporal and spatial reasoning from prose) than ingredient retrieval from structured label text. What is comparable is the architecture: the same three constraints, applied to the same class of problem, producing usable datasets in spaces where no dataset previously existed.

7.1 Acknowledgements

My deepest gratitude to Mr. Krishna, whose constancy forms the foundation upon which all my work, including this, quietly rests. Salutations to the Goddess who dwells in all beings in the form of intelligence. I bow to her again and again.

This report was prepared as part of the Indian Food Informatics Data (IFID) project at the Interdisciplinary Systems Research Lab (iSRL).

7.2 Statements and Declarations

4.1 Why OFF Could Not Be the Only Source

OFF contains thousands of English ingredient lists for products with Indian brand names. Those lists are valuable. But the dataset structure does not reliably distinguish a product currently on Indian supermarket shelves from an imported variant, an export formulation, or a historical listing no longer in distribution.

For the purpose of this corpus — documenting what Indian consumers actually encounter today — that distinction matters. An ingredient list attached to a product that is not in the Indian market does not reflect the vocabulary Indian food systems use.

The null rates in the OFF data confirm the scale of the gap. Of 19,748 rows in the raw export (captured 2 February 2026):

• Only 4,104 pass a minimum filter: brand present, English product name present, English ingredient text present. That is 20.78 percent.
• ingredients_text_en is the only ingredient column with coverage above 1 percent. All 29 other language columns combined add 69 rows to that count.
• 6,905 rows have both a brand identifier and an English product name — the product exists, it has a name — but no ingredient text in any language. The gap is specifically at the ingredient field.
• The four core macronutrient fields (energy, fat, protein, carbohydrates) have null rates between 65.61 and 66.00 percent across the full dataset.

These numbers are not a criticism. They are a measurement of the space. The gap in documented Indian food data is real, it is large, and it exists because the underlying ecosystem is genuinely fragmented — not because anyone has failed to document it well enough.

What OFF does have — the 4,104 rows with a brand, a product name, and an English ingredient list — is usable for this project. The product name provides the minimum anchor needed to verify the product is Indian. Those rows were taken through the same constrained parsing pipeline described below, and their ingredient strings added to the vocabulary set. The filter that kept them is described in the claims document.

4.2 Why Direct Sampling Was Necessary

For a reliable picture of what is currently on Indian shelves, the corpus needed to be collected directly. The methodology: select products from companies with documented Indian market presence, retrieve ingredient lists from verifiable online sources, extract and parse.

Company and product selection. Ten companies were selected based on market presence across major packaged food categories — snacks, beverages, staples, dairy, condiments. Within each company, selection moved through sub-brands (ITC’s portfolio spans Aashirvaad, Sunfeast, Bingo, YiPPee — each with a different ingredient vocabulary) and then to individual SKUs meeting four criteria:

1. Ingredient list traceable on a whitelisted domain
2. Product available in the Indian market, not an export or international variant
3. Specificity to a single SKU, not a product range — “Aashirvaad Turmeric Powder 200g” not “Aashirvaad Spices”
4. One representative retained per formulation across pack sizes

The third criterion produced the most rejections. References like “Cadbury Chocolates” or “Aashirvaad Spices” denote product families, not individual items with specific ingredient lists. Every such reference required disambiguation before it could enter the corpus.

After validation: 1,000 SKUs across 42 companies, 153 brands, 8 macro-categories.

6.1 Constrained Retrieval

The model retrieved ingredient lists from whitelisted domains only, in priority order: brand official website, then Amazon India, BigBasket, Blinkit. If no source returned the ingredient list, the output was DATA_NOT_FOUND. The instruction was explicit: do not infer typical ingredients from product category, do not approximate based on similar products.

Temperature was set to 0. This means the model selects the highest-probability token at each step and produces identical output for identical input. The practical effect: if you run the same extraction twice, you get the same result. Validation becomes tractable. Fabrication through sampling variation is eliminated.

Batch size was tested across 1 to 20 SKUs per call. At batch sizes above 10, constraint violations increased measurably — the model began returning approximations instead of DATA_NOT_FOUND for products it could not find, and formatting inconsistencies appeared. Six SKUs per batch produced the best balance of throughput and constraint adherence.

Results across 1,000 SKUs:

• First pass: 871 successful extractions (87.1%), 129 DATA_NOT_FOUND
• Second pass on the 129 failures: 25 additional extractions, 104 persistent failures
• Final corpus: 896 extracted (89.6%), 104 excluded

The 104 persistent failures validate that the constraint held. Those products either had no verifiable online ingredient list or were no longer in active distribution. The pipeline returned an explicit gap rather than a filled approximation. An explicit gap can be addressed later. A fabricated entry corrupts the vocabulary invisibly.

Manual audit of 90 extractions from the 896: 1 error identified (the model merged content from two adjacent sections of a product page). Error rate: 1 in 896 (0.11 percent).

6.2 Structure-Preserving Parsing

The 896 extracted ingredient lists were not fed to the pipeline as whole strings. Each list went through parsing as a discrete operation, because the structure of Indian food labels encodes relationships that naive splitting destroys.

Consider what happens when a comma-splitter treats every comma equally:

Input: Acidity Regulators (INS 296, INS 330, INS 334)

Naive output:

• Acidity Regulators (INS 296
• INS 330
• INS 334)

The functional context — that INS 296, 330, and 334 are all acidity regulators — is gone. The fragments INS 330 and INS 334) have no meaning without it.

The structure-aware parser tracks nesting depth. It splits on commas only at depth zero — the root level. Everything inside brackets is treated as a unit until the brackets close. Applied to the same input:

Structure-aware output: Acidity Regulators (INS 296, INS 330, INS 334) — intact, ready for decomposition with context preserved.

896 ingredient lists → 2,926 parsed strings with functional relationships intact.

6.3 Artifact Removal

Each of the 2,926 strings went through a single-purpose cleaning pass: remove presentation artifacts, preserve identity information.

Removed: percentage values (55.7% — quantity, not identity), marketing text (BINGO!, NEW!), usage annotations (#Used As Natural Flavouring Agent).

Preserved: INS codes and E-numbers (regulatory identifiers), preparation specifications (Salt (Iodised) — the bracketed term distinguishes a specific variety), functional classifications (Acidity Regulator, Emulsifier).

The distinction matters because it is not always obvious. 55.7% is presentation — removing it loses nothing about what the ingredient is. (Iodised) is identity — removing it collapses iodised salt and table salt into the same entry, which is wrong.

Batch sizes for this stage were set inversely to string complexity: short strings processed in batches of 40, complex multi-bracket strings processed one at a time. Attention dilution at scale produces the same constraint violations as in the retrieval stage.

6.4 Semantic Decomposition

After cleaning, compound structures were decomposed with context propagation. Each atomic operation took one compound and returned its components, with the functional classification carried forward to each:

Input: Flavour Enhancers (INS 627, INS 631)
Output: Flavour Enhancer INS 627, Flavour Enhancer INS 631

Input: Stabilizing & Emulsifying Agents (412, 410, 407, 471, 466)
Output: Stabilizer INS 412, Stabilizer INS 410, Stabilizer INS 407, Emulsifier INS 471, Stabilizer INS 466

Input: Black Pepper Powder, Ginger Powder, Clove Powder
Output: unchanged — already atomic

2,926 cleaned strings → 3,452 decomposed ingredient mentions → 1,987 unique variants after deduplication across all 896 products from the sampling pipeline.

The full transformation for one product (Bingo Original Style, ITC Ltd.) produced 21 ingredient mentions, including:

‘Black Salt’, ‘Chilli’, ‘Citric Acid (INS 330)’, ‘Disodium Guanylate (INS 627)’, ‘Disodium Inosinate (INS 631)’, ‘Garlic’, ‘Hydrolyzed Vegetable Protein’, ‘Maida’, ‘Malic Acid (INS 296)’, ‘Maltodextrin’, ‘Milk Solids’, ‘Onion’, ‘Palm Oil’, ‘Potato’, ‘Salt’, ‘Silicon Dioxide (INS 551)’, ‘Spices and Condiments’, ‘Sugar’, ‘Tartaric Acid (INS 334)’, ‘Tomato’

9.1 Claims

9.2 Evidence Per Claim

4.1 Reading changes in context

Laws are written under constraints. The constraints include the state of the relevant industry at the time of drafting, the administrative capacity available to enforce the rule, the incidents that made a regulatory response necessary, and the political feasibility of various options. A later version of a rule almost always looks more precise, more comprehensive, or more technically sophisticated than an earlier one—not because earlier drafters were careless, but because they were working with less data, less precedent, and a less developed field.

The appropriate stance when reading a delta is that of a scribe rather than a judge: the task is to document what changed, to note what the earlier version could not have anticipated, and to ask what new information or new pressure made the change necessary. This is a different question from asking whether the earlier rule was wrong.

Applied to the Indian food labelling context, the transition from the Food Safety and Standards (Packaging and Labelling) Regulations, 2011 to the Food Safety and Standards (Labelling and Display) Regulations, 2020 illustrates this clearly (Vukka and Lalitha 2026). The 2011 regulations were drafted at a point when industrial food processing in India was still expanding rapidly and digital traceability tools were not yet available to enforcement bodies. The 2020 regulations tightened allergen declarations, introduced structured front-of-pack warnings, and prescribed naming conventions with greater specificity. Each of these additions corresponds to a domain that developed, was observed, and was then addressed. The delta reveals an institution processing experience and updating its instrument accordingly.

4.2 Identifying friction points through case law

Where the regulatory text leaves ambiguity, the courts resolve it—and in doing so, produce a record of where the ambiguity was, who held which interpretation, and which reading eventually prevailed. This makes case law particularly useful for identifying friction points that would not be visible from the statutory text alone.

The Supreme Court of India’s January 2026 judgment in Commissioner of Customs (Import) v. M/s Welkin Foods illustrates this (Lalitha 2026). The case concerned whether imported aluminium shelving should be classified as an agricultural machine part or as an aluminium structure. The legal question was narrow, but the Court’s reasoning established a hierarchy for resolving classification disputes in which statutory technical definitions take precedence over common commercial understanding. This hierarchy had been contested in earlier decisions and was now settled. For a food informatics system that must align with Indian classification practice, this judgment is a material constraint—one that would not have been legible from reading only the statutory text.

4.3 Mapping coordination across institutional actors

A regulatory domain typically involves multiple bodies with overlapping but non-identical mandates. In Indian food systems, the relevant actors include the Food Safety and Standards Authority of India (FSSAI), the Directorate General of Foreign Trade (DGFT), the Central Board of Indirect Taxes and Customs (CBIC), and the courts. These actors do not always interpret the domain identically, and their instruments do not always align.

The legal record makes these relationships visible. Where one body’s definition conflicts with another’s, there will typically be a judgment or a regulatory amendment that resolves the conflict, defers it, or acknowledges it. Mapping these interactions across time reveals not just what the current rule is, but how the current rule came to be and which pressures it is still absorbing.

2.1 What the Law is Aiming to Address

Read together, these shifts point to a gradual tightening around a few recurring questions.

3.1 For Food Businesses

• Allergen tracking becomes more explicit: manufacturers benefit from maintaining internal mappings between ingredients and standard allergen groups, rather than relying only on free-text descriptions.
• Naming policies may require review: product names and descriptors that were aligned with earlier interpretations may need revisiting to match the “true nature” emphasis in the 2020 framework.
• Nutrition data hygiene matters more: keeping consistent, up-to-date values for energy, sugars, fats and sodium supports both regulatory expectations and clearer communication with consumers.

3.2 For Lawyers and Compliance Teams

• Case work can use more structure: instead of only reading long labels line by line, advisors can ask whether allergen declarations, names and nutrition panels line up with the specific structures the 2020 regulations describe.
• Advice can be more template-driven: it becomes realistic to build standard checklists for allergens, naming and nutrition that can be reused across clients or product lines.

3.3 The Compliance Checklist for Startups

For founders and early-stage teams, a quick sanity check can be more useful than a long memo. A simple checklist that falls out of the 2011–2020 delta is:

• Allergen coverage: Have you identified and labelled all FSSAI priority allergen groups that apply to your product?
• Front-of-pack name: Does the name on the principal display panel reflect the true nature of the food, rather than only a marketing phrase?
• Per serving signals: Is the per-serving nutrition information (including percentage RDA where applicable) clear enough for a consumer to understand the product’s fat, sugar and salt profile at a glance?

Treating these as a recurring checklist rather than a one-time launch task makes it easier to stay aligned with how the 2020 regulations expect labels to behave.

4.1 Mandatory vs Optional Fields in IFID Records

For a digital ingredient identity layer to stay aligned with the 2020 regulations and still be useful for future extensions, it helps to separate mandatory compliance fields from optional but valuable metadata.

2.1 The Initial Approach

The natural first response to a multiplicity of ingredient strings is to collapse them. Given 7,563 strings and the reasonable expectation that they represent far fewer substances, the immediate goal was consolidation: assign each string to a canonical form, discard the variation, and produce a clean taxonomy.

This approach was implemented and produced a working taxonomy published as version 0.1 of the Encyclopedia of Indian Food Ingredients (Lalitha 2026a). That taxonomy served as a necessary first step: it demonstrated that automated consolidation was feasible, identified the problem’s boundaries, and surfaced the cases where flat consolidation produced results that were operationally and legally indefensible. The present report builds directly on what those cases revealed.

2.2 What Flat Canonisation Produces

Under a flat canonisation scheme, all variant strings for a given biological source are grouped under a single canonical label. The chilli variants listed in Section 1 would consolidate to “Chilli.” The mango variants would consolidate to “Mango.” The logic is appealing: one biological entity, one canonical name.

The problem becomes visible when the output is examined by stakeholders who depend on ingredient classifications for operational decisions.

For a food manufacturer seeking to claim a geographically indicated ingredient: “Mathania Red Chilli” is not interchangeable with “Chilli.” Mathania is a geographic indicator associated with a specific cultivar grown in the Barmer district of Rajasthan, recognised for its characteristic colour and moderate heat. A brand that sources this variety and wishes to communicate that fact on its label—a commercially and legally meaningful distinction—has no mechanism for doing so under a scheme that treats all chilli as one entity.

For a nutritional researcher or regulator: “Mango pulp” and “dehydrated mango powder” are not nutritionally equivalent. The former is a high-moisture preparation with a specific sugar profile; the latter has undergone water removal that concentrates all components and, depending on process conditions, may alter certain phytochemicals. A database recording both as “Mango” provides no basis for dietary assessment calculations that depend on moisture-adjusted nutrient values.

For a customs authority: “Mango flavouring” filed alongside “mango pulp” under a single canonical entity produces a tariff classification that is straightforwardly incorrect. Mango pulp falls under HS Chapter 08 (edible fruits); a synthetic mango flavouring may fall under Chapter 29 (organic chemicals) or Chapter 33 (essential oils and resinoids) depending on its composition. Filing them under the same canonical entity does not resolve the classification question; it conceals it.

For a food safety system tracking an allergen or contaminant: “Lecithin” and “soya lecithin” cannot be merged without losing source information that is required by law. The FSSAI Labelling Regulations and the reasoning in Ram Gaua Raksha Dal (Delhi High Court 2022) together establish that source disclosure for allergen-relevant ingredients is non-negotiable.

2.3 The Structural Flaw

Flat canonisation fails because it conflates two distinct problems requiring different solutions. The first is coordination: establishing that “chilli,” “red chilli,” and “lal mirchi” refer to the same underlying entity so that systems can interoperate. The second is identity preservation: maintaining the distinctions—geographic origin, processing state, form, biological source—that carry legal, nutritional, commercial, and cultural meaning.

A flat scheme solves the first problem by destroying the second. It achieves coordination at the cost of the very information that makes coordination useful. A brand filing its ingredient as “Chilli” and a brand filing it as “Kashmiri Lal Mirch” can now be linked in a database, but the database no longer records what distinguishes them—a distinction that may affect GST categorisation, GI protection claims, export certification, and consumer communication simultaneously.

The correct solution is a layered structure: a coordination layer linking all variant representations to a shared identifier, and an identity-preservation layer retaining the distinctions that matter. This is precisely the problem Shiyali Ramamrita Ranganathan addressed in information science nearly a century ago.

3.1 The Context of Its Creation

In 1933, the Indian mathematician and librarian S. R. Ranganathan published the first edition of Colon Classification (Ranganathan 1933). The problem he addressed was structurally similar to the one this report confronts: a body of knowledge so diverse and growing so rapidly that any fixed hierarchical scheme would be perpetually inadequate. The Dewey Decimal System, then dominant in library science, assigned each subject a fixed position in a single hierarchy. Works addressing multiple subjects simultaneously, or belonging to a subject not anticipated by the scheme’s designers, could not be accommodated without distorting the classification.

Ranganathan’s response was to abandon the single hierarchy and replace it with a set of independent analytical dimensions, which he called facets. A document could be described by its position on each facet independently, and its classification was the combination of those positions. The colon in “Colon Classification” is the separator between facets in the notation.

3.2 The PMEST Framework

Ranganathan identified five fundamental facets applicable across all fields of knowledge, designated PMEST: Personality, Matter, Energy, Space, and Time (Ranganathan 1967). These represent, respectively, the primary subject of a document, the materials or substances it involves, the processes or operations it describes, the geographic location it concerns, and the time period it covers.

The operational power of the framework lies in the independence of its facets. A document about the fermentation of rice in Karnataka in the nineteenth century can be precisely described by assigning positions on each facet—rice (Personality), fermentation (Energy), Karnataka (Space), nineteenth century (Time)—without requiring that the classification scheme anticipate this exact combination in advance. New combinations form by combining existing facet values; the scheme extends to novel cases without revision.

Adapted to the food domain, the analytical clarity is immediate. Consider three ingredients:

• Kashmiri red chilli powder: Personality = chilli (Capsicum annuum); Matter = dried, powdered; Space = Kashmir.
• Mathania red chilli, whole dried: Personality = chilli (Capsicum annuum); Matter = dried, whole; Space = Marwar (Rajasthan).
• Green chilli paste: Personality = chilli (Capsicum annuum); Matter = raw, comminuted, high-moisture.

Under a flat scheme, all three are “Chilli.” Under a faceted scheme, all three share a Personality coordinate—sufficient to establish their relationship—while their distinct Matter and Space coordinates preserve the differences that matter. A fourth ingredient, a synthetic capsaicin extract used as a flavouring agent, would share a Personality relationship to chilli while carrying a very different processing history and a different functional identity. The framework accommodates this without modification.

3.3 Adoption and Durability

Colon Classification was adopted by the Indian National Library and numerous university libraries across South and Southeast Asia, and served as the theoretical foundation for the International Federation of Library Associations’ principles on faceted classification (Broughton 2006). Subsequent frameworks—including the Bibliographic Classification of Henry Bliss and the Universal Decimal Classification’s faceted extensions—drew directly on Ranganathan’s architecture.

The durability of the framework across domains as diverse as bibliography, archival science, museum cataloguing, and digital information architecture reflects its quality as a structural solution rather than a domain-specific convention. The problem it addresses—organising entities that are complex, diverse, and not fully anticipated in advance—is exactly the problem that Indian food ingredient classification presents.

3.4 From Library Science to Food Identity

Applying the PMEST framework to the ingredient dataset immediately clarified which distinctions were meaningful and which were surface variation. The distinction between “chilli powder” and “chilli flakes” is a legitimate Matter distinction (fine-ground versus coarsely broken), not a naming inconsistency to be collapsed. “Kashmiri chilli” and “generic red chilli” differ on the Space facet, not the Personality facet, and that distinction carries regulatory weight in the context of geographical indication protection.

However, three categories of cases emerged that the PMEST framework as originally conceived did not fully resolve. The first concerned artificial or nature-identical flavourings: does “mango flavouring” belong under Mangifera indica as a Personality, or has synthesis transformed its identity so thoroughly that the source becomes secondary to the function? The second concerned highly processed lipids: is “soya lecithin” a variant of soybean, or has extraction and fractionation placed it in a different identity category—one defined by its emulsification function rather than its botanical origin? The third concerned processing-derived additives with no meaningful biological ancestor: modified starches, synthetic antioxidants, and inorganic salts have an HS classification and a regulatory name, but no Personality in the biological sense.

These categories expose the ontological questions a classification framework for food ingredients must resolve before it can be applied consistently. Those questions are addressed in Section 4.

6.1 The Need for Three Dimensions

Chapters 2 through 5 have established that ingredient identity is not a single-dimensional property. Flat canonisation collapses distinctions that matter; classification by processing level alone conflates ingredients that regulatory systems treat as categorically different. Ranganathan’s faceted approach provides the theoretical architecture, but its application to food ingredients requires computational operationalisation: dimensions that are measurable, independently assignable, and combinable into a diagnostic framework.

Three dimensions are necessary and, as argued below, sufficient to capture the identity distinctions that regulatory systems actually make.

First, how invasively was the ingredient transformed? This is a question about process: the energy and chemistry invested in moving an ingredient away from its native biological state. It is measured by the Anthropogenic Energy Score ().

Second, how far has the ingredient moved from its source matrix? This is a question about the resulting material: how much of the original biological context—moisture, fibre, co-nutrients, cellular structure—remains in the ingredient as it enters the food system. It is measured by the Matter Score ().

Third, does the ingredient’s regulatory and commercial identity follow its biological source or its technological function? This is a question about the legal-semiotic position of the ingredient: whether it is named, classified, and governed as a product of a biological origin or as a performer of a technological role. It is measured by the Functional Score ().

These dimensions are independent. A moderately processed ingredient can have high functional identity (propellant gases have high despite moderate ). A heavily processed ingredient can retain low functional identity (hydrogenated vegetable oil has high but low because regulatory naming retains source primacy). No single axis is sufficient to determine identity, and the combination of all three resolves cases that any two alone leave ambiguous.

The full technical justification for each score assigned in the tables that follow—including process-by-process derivations, supporting citations, and defensibility ratings—is documented in the companion scoring report (Lalitha 2026b). The present chapter states the framework and its outputs; the companion document shows the derivation.

6.2 The Anthropogenic Energy Score ()

6.2.3 Selected Score Reference Values

Table 2 presents reference values for representative processes.

Process		Band
Sorting	0.12	Physical
Washing	0.15	Physical
Chilling	0.18	Physical
De-husking	0.22	Physical
Milling (e.g., Besan)	0.28	Physical
Cold Pressing (Oil)	0.32	Physical
Churning (Butter)	0.45	Physical/Thermal
Pasteurization	0.48	Thermal
Clarification (Ghee)	0.55	Thermal
Fermentation (Vinegar)	0.56	Biological
Roasting	0.58	Thermal/Chemical
Refining (Vegetable Oil)	0.75	Industrial/Fractional
Fractionation (Olein)	0.76	Industrial/Fractional
Solvent Extraction (Oils)	0.82	Industrial/Fractional
Interesterification	0.91	Chemical/Synthetic
Hydrogenation	0.92	Chemical/Synthetic
Acetylation (Modified Starch)	0.94	Chemical/Synthetic
Synthetic Vanillin	0.98	Chemical/Synthetic
Synthetic Flavors (General)	0.99	Chemical/Synthetic

6.3 The Matter Score ()

6.3.2 Selected Score Reference Values

Table 3 presents reference values for representative commercial states.

Final Commercial State		Matter Class
Whole/fresh pieces	0.05	Hydrated/Native
Cut/sliced pieces	0.10	Hydrated/Native
Pulp/puree	0.25	Comminuted
Coarse grits	0.30	Comminuted
Flour/fine powder	0.33	Comminuted
Flakes	0.36	Dehydrated/Concentrated
Dense block (e.g., khoya)	0.38	Dehydrated/Concentrated
Concentrate (liquid)	0.40	Dehydrated/Concentrated
Powder (spray-dried)	0.42	Dehydrated/Concentrated
Juice (clarified)	0.50	Structural Fractionation
Whey powder	0.52	Structural Fractionation
Skim/defatted meal	0.55	Structural Fractionation
Starch flour	0.60	Structural Fractionation
Oil	0.70	Constitutional Isolate
Fat fraction	0.72	Constitutional Isolate
Protein concentrate	0.74	Constitutional Isolate
Protein isolate	0.78	Constitutional Isolate
Granules (agglomerated isolate)	0.80	Constitutional Isolate
Extract/oleoresin	0.86	Molecular Signal/Extract
Oleoresin (viscous)	0.88	Molecular Signal/Extract
Emulsifier powder (e.g., lecithin)	0.89	Molecular Signal/Extract
Essential oil	0.90	Molecular Signal/Extract
Modified starch powder	0.96	De Novo/Synthetic
Crystalline chemical	0.98	De Novo/Synthetic

6.4 The Functional Score ()

6.4.3 Scores Across FSSAI Functional Classes

Table 4 presents the range for each of the twenty-two functional classes enumerated in Schedule I of the Food Products Standards and Food Additives Regulations 2011, derived from the three-part test.

Functional Class	Score	Zone
Base ingredient (non-additive)	0.12	Source-Dominant
Taste profile / spice	0.18	Source-Dominant
Lipid base (edible oil/fat)	0.22	Source-Dominant
Bulking agent	0.35–0.40	Source-Retaining
Humectant	0.40–0.45	Source-Retaining
Firming agent	0.42–0.48	Source-Retaining
Raising agent	0.45–0.50	Source-Retaining
Flavouring agent	0.60–0.75	Function-Emergent
Thickener	0.58–0.65	Function-Emergent
Stabiliser	0.62–0.68	Function-Emergent
Gelling agent	0.65–0.70	Function-Emergent
Sweetener (bulk/intense)	0.55–0.70	Function-Emergent
Foaming agent	0.70–0.75	Function-Emergent
Colour	0.75–0.85	Function-Emergent / Dominant
Emulsifier	0.80–0.85	Function-Dominant
Anticaking agent	0.85	Function-Dominant
Acidity regulator	0.85–0.87	Function-Dominant
Antioxidant	0.87–0.88	Function-Dominant
Preservative	0.87–0.90	Function-Dominant
Antifoaming agent	0.90	Function-Dominant
Sequestrant	0.90–0.92	Function-Dominant
Bleaching agent	0.92	Function-Dominant
Flour treatment agent	0.93	Function-Dominant
Carrier	0.94	Function-Dominant
Propellant	0.95	Function-Dominant
Packaging gas	0.95	Function-Dominant

6.5 The –– Coordinate System

Each ingredient is assigned a position in a three-dimensional coordinate space: , , . The position is the ingredient’s identity coordinate—its location in the space of processed ingredients, determined independently on each dimension.

The coordinate is not a summary statistic; it is a structured representation preserving the information carried by each dimension. Two ingredients with the same score (derived in Section 7) may have very different coordinate profiles reflecting different kinds of identity transformation. The coordinate system allows these differences to be traced and reasoned about.

Assignment of coordinates follows the evidence hierarchy: FSSAI regulations and product standards take precedence over general labelling rules; ITC-HS chapter assignments provide independent corroboration; judicial reasoning fills gaps and resolves conflicts. Where evidence is unavailable or conflicting, the assignment is flagged as provisional and subject to revision through the contribution protocol in Appendix A.

9.1 NOVA and Ingredient-Level Substrates

The NOVA food processing classification system classifies food products into four groups based on the extent and purpose of industrial food processing (Arora et al. 2025; Ispirova et al. 2025). NOVA Group 4 (ultra-processed foods) is defined by reference to industrial processing and the presence of ingredients typically used only in industrial production—many of which correspond to Zone 3 of the –– model.

NOVA operates at the product level: given a complete food product, it classifies the product by the nature of its processing. The –– model operates at the ingredient level: given an individual ingredient string, it assigns that ingredient a deterministic identity position. Product-level classification requires reliable ingredient-level classification as its substrate, and recent machine learning work applying NOVA to large datasets has encountered precisely this bottleneck: the absence of a principled ingredient-level scheme limits the accuracy and consistency of product-level predictions (Arora et al. 2025; Ispirova et al. 2025).

The correspondence between the two frameworks is not coincidental—both are responding to the same underlying physical and regulatory reality about how processing transforms ingredient identity. Zone 3 ingredients (functional tools defined by technological role) map directly onto the additive-classified substances that NOVA uses to identify ultra-processed products. Zone 1 and Zone 2 ingredients map onto the culinary and processed ingredients of NOVA Groups 2 and 3. The –– model makes that reality computationally deterministic at the ingredient level, which is what product-level frameworks need as input.

9.2 The ITC-HS as Ground Truth

The Indian Trade Classification (Harmonised System) has been used throughout this report as primary evidence—a regulatory system that has already resolved many ingredient identity questions through decades of judicial and administrative refinement. The Supreme Court’s ruling in Welkin Foods (Supreme Court of India 2026) places HSN classification at the top of the interpretive hierarchy for identity disputes.

The –– framework uses the ITC-HS as its primary evidence base. HS codes are necessary but not always sufficient for ingredient-level classification: two ingredients may share an HS heading while having very different , , and scores if their processing histories and regulatory naming differ within the heading. The framework provides finer-grained resolution within and across HS headings. Where –– coordinates and HS chapter assignments converge—as they do in the majority of benchmark cases—that convergence is confirmation that the model is correctly grounded. Where they diverge, that divergence is a signal requiring investigation.

10.1 Weight Calibration

The weights in the Divorce Score formula are provisionally assigned and have not been validated against a large empirical dataset of regulatory decisions or expert classifications. The choice of 0.4 for and 0.3 each for and is analytically motivated—the reasoning is set out in Section 7—but it has not been optimised against a ground-truth corpus. Refinement of the weights using subject matter expert input, expanded benchmark coverage, or Bayesian calibration against regulatory decision data is anticipated and invited through the contribution protocol in Appendix A.

10.2 Zone Boundary Precision

The thresholds at and are calibrated to the benchmark cases but have not been validated across the full range of Indian food system ingredients. Ingredients near the thresholds may be sensitive to small changes in coordinate assignment. This sensitivity is acknowledged as a characteristic of the framework, not a failure: the zone boundaries are policy-relevant thresholds, not natural discontinuities in the physical or chemical properties of ingredients. The framework makes its current specification transparent and open to empirical refinement.

10.3 Source Coordinate Incompleteness

The score captures regulatory naming modality but does not encode the full specificity of the source coordinate: whether an ingredient is of plant or animal origin, whether it carries a geographic indication, or whether it has a specific religious or ethical status. A proposed extension treats source-metadata as a structured annotation on each canonical entry, separate from the three-dimensional coordinate system. This would allow recording “soya lecithin: source = Glycine max, vegan-compatible, allergen-flagged (soy)” as metadata attached to the Zone 3 classification without adding a fourth dimension that would complicate the score calculation.

10.4 Dynamic Regulatory Landscape

The regulatory ground truth used to calibrate scores is a snapshot as of 2025. Food regulation in India is actively evolving: FSSAI has issued amendments, notifications, and draft regulations at increasing frequency, and the judicial landscape continues to develop (Vukka and Lalitha 2026). The framework architecture accommodates this: scores are derived from a three-part test against specific regulatory provisions, so changes to those provisions propagate to score updates without requiring a redesign. Maintaining currency with regulatory changes is an ongoing maintenance task.

10.5 Scope: Indian Regulatory Context

The framework is calibrated specifically to the Indian regulatory context—FSSAI instruments, ITC-HS schedules, and Indian judicial precedent. The and dimensions are grounded in chemistry and nutrition science that is internationally applicable, but the dimension is context-specific. Extension to other regulatory contexts would require parallel derivation of scores from those contexts’ instruments. The architecture is designed to support such extension; the calibration work has not been performed.

11.1 What the Corpus Looks Like

The commercial sampling work conducted as part of this project—covering 896 stock-keeping units drawn from Indian retail channels, with full methodology to be documented in a forthcoming report—combined with the Open Food Facts India dataset (Open Food Facts contributors 2024) yields approximately 4,800 deduplicated products. Splitting ingredient declarations by comma and conjunction across the full combined corpus produces approximately 48,000 variant strings. The two sources are methodologically distinct: the 896 SKU sample is a structured retail survey; the Open Food Facts contribution is a different collection pathway with its own coverage characteristics. Both are part of this project’s data infrastructure.

The processes and physical forms documented in the and reference tables of this report were derived from systematic examination of what actually appears across those 48,000 strings—not from prior literature alone, but from the empirical evidence of how Indian packaged food manufacturers describe their ingredients on commercial labels. The variant corpus is the empirical foundation on which the –– framework rests.

11.2 The Classification Task Ahead

The immediate next step is to build the faceted ingredient system: assigning coordinates and scores to each of the 729 canonical entities in the Encyclopedia v0.1 taxonomy (Lalitha 2026a), and then mapping the approximately 48,000 variant strings to their canonical families through the entity resolution pipeline.

This is a forward task. What the variant corpus has provided so far is the empirical basis for defining processes, matter classes, and functional zones—the population of real forms and transformations that any viable framework must handle. The next phase applies the –– model to classify each canonical entity systematically, extend those classifications through the suffix system to geographic, cultivar, and preparation-state variants, and attach the legal metadata (allergen flags, source disclosure obligations) that Zone assignments alone do not capture.

Each variant string will carry, as its output: a canonical ID, a zone classification, a score, suffix metadata encoding whatever distinctions the variant expresses beyond the canon, and any applicable legal disclosure flags. That structured output is what downstream systems—allergen detection, supply chain coordination, nutritional research, regulatory compliance—require as their input.

11.3 Governance and Expert Input

The classification of 729 canonical entities will not be completed by computational methods alone. Score assignments in the boundary regions—Zone 1/Zone 2 transitions for moderately processed ingredients, Zone 2/Zone 3 transitions for ingredients with mixed regulatory signals—require domain expertise that food scientists, food lawyers, customs practitioners, and nutritional researchers hold.

The expert input process described in Appendix B is the mechanism for incorporating this knowledge. What is available computationally is the framework specification, the benchmark as a quality standard, and the variant corpus as the empirical scope of the problem. What domain experts contribute is the evidence-based judgement about where specific ingredients fall within that framework, particularly in the cases that the benchmark was designed to surface as hard.

11.4 Planned Outputs

The classification work will produce a versioned update to the Encyclopedia of Indian Food Ingredients, with –– coordinates and zone classifications attached to each canonical entry, and with full technical derivations reviewed against the companion scoring report (Lalitha 2026b). A versioned update protocol will be implemented so that regulatory changes propagate to score updates in a traceable and documented manner. The source metadata extension—structured annotation for origin-specific data including botanical source, geographic indication status, and religious or ethical classification—will be developed alongside the coordinate assignments.

The goal of this work is a publicly accessible, versioned, expert-reviewed ingredient classification system that any downstream application—NOVA-based product classification, allergen detection, supply chain systems, nutritional databases—can use as a stable, deterministic substrate.

2.1 United States: Nix v. Hedden (1893)

The seminal case Nix v. Hedden, 149 U.S. 304 (1893), established that tomatoes should be classified as vegetables rather than fruits for tariff purposes, despite their botanical classification. The Supreme Court of the United States held that classification should follow common parlance—how ordinary people in commerce understand terms—rather than technical botanical definitions.

2.2 India: Early Common Parlance Cases

2.3 The Pre-HSN Framework

Prior to India’s adoption of the HSN system in 1986, courts consistently applied common parlance as the primary interpretive tool for ambiguous statutory terms. This approach served several purposes:

• It made tax classifications accessible to ordinary merchants without specialized knowledge
• It aligned legal interpretations with commercial practice
• It avoided the complexity of technical botanical or chemical classifications
• It provided predictability based on everyday understanding

3.1 Case Details

Citation: 2026 SCC OnLine SC 27; 2026 INSC 19
Date: January 6, 2026 (reported January 6-7, 2026)
Court: Supreme Court of India
Bench: Justice J.B. Pardiwala and Justice R. Mahadevan
Parties: Commissioner of Customs (Import) v. M/s Welkin Foods

3.2 Facts and Issue

The case concerned the proper classification of imported aluminium shelving used for mushroom cultivation. The respondent, Welkin Foods, argued the goods should be classified under Customs Tariff Item (CTI) 84369900 as “parts” of agricultural machinery. The Revenue contended the shelving should be classified under CTI 76109010 as “Aluminium Structures.”

The core legal question was whether the intended use of the product (mushroom cultivation) or its objective technical characteristics should govern classification.

3.3 The Court’s Reasoning

The Supreme Court held that classification must be based on objective characteristics of the product, not solely on intended end-use. The Court established several critical principles:

1. Structure vs. Machine: The shelving was held to be a “structure” (fixed in place) rather than a “part” of a machine. It did not qualify as a component essential for the mechanical function of agricultural machinery.

2. Material Identity Primacy: While exclusive use can sometimes influence classification, it does not override the fundamental material identity of the product when it is specifically described elsewhere in the tariff.

3. Hierarchical Framework: Most significantly, the Court articulated a hierarchy for classification disputes, stating: “It is only in a state of statutory silence, where the legislative intent remains unexpressed, that the tribunals or courts may resort to the common or trade parlance test.”

3.4 The Established Hierarchy

The Ms Welkin Foods judgment established the following precedential hierarchy for food and product classification:

1. Judicial Interpretation of Statute (highest priority)—How the court reads the HSN and statutory provisions
2. Technical/Scientific Definition—When statute provides technical guidance through HSN codes
3. Expert Opinion—Testimony from qualified experts in relevant fields
4. Common Parlance—Trade usage and ordinary understanding (only in statutory silence)

This hierarchy fundamentally reorients Indian classification jurisprudence, relegating common parlance from its historical primacy to a fallback position.

4.1 GST and Tax Domain: Scientific Composition Prevails

4.2 Food Safety Domain: Nutrient Thresholds Over Marketing Terms

4.3 Customs Domain: Engineering Function Over Trade Nomenclature

The Ms Welkin Foods case itself exemplifies this domain. The Supreme Court held that aluminium racks for mushroom cultivation are technically structures (Chapter 76) and not machinery (Chapter 84) because they lack mechanical function, regardless of their trade name or intended agricultural use.

This establishes that in customs classification, the technical characteristics—material composition and functional properties—override the commercial designation or end-use of a product.

4.4 Dietary Labels Domain: Biological Origin Disclosure Mandatory

6.1 Source Disclosure

Requirement: Even if an ingredient is heavily processed or becomes a derivative compound, the source must be declared.

Example: Lecithin must be labeled as “Lecithin (Soy)” or “Lecithin (Egg),” not merely as “Lecithin” or by INS code alone.

Legal Basis: Ram Gaua Raksha Dal v. Union of India

6.2 Dietary Status

Requirement: Products must be classified as vegetarian (with egg/dairy), non-vegetarian, or pure vegetarian (no animal source or dairy).

Principle: Even if an ingredient is a chemical derived from an animal source, it must be declared with respect to its source.

Legal Basis: Ram Gaua Raksha Dal v. Union of India; Food Safety and Standards (Labelling and Display) Regulations, 2020

6.3 Allergen Declaration

Requirement: Even if processing is extreme and the allergen potency is reduced from the source level, allergen presence must still be declared for consumer safety.

Legal Basis: Right to health (Article 21 of the Constitution); FSSAI allergen declaration requirements

6.4 Functional Class for Chemicals

Requirement: For chemical additives, the functional class (purpose of inclusion) must be declared followed by the INS Number.

Example: “Preservative (INS 202)” rather than merely “INS 202” or “Potassium Sorbate.”

Legal Basis: FSSAI Labelling Regulations 2020; functional usage declaration requirements

7.1 Attribute-Based Ingredient Classification

The cases reveal that the determination of whether something constitutes a separate ingredient entity versus a variant of an existing entity depends on functional attributes rather than source alone.

Key Principle: Functionality takes precedence over source material when determining ingredient separateness.

For example, in the instant mix flour cases, the presence of functional additives (leavening agents, spices used for specific culinary purposes) transformed what might be considered a “variant of flour” into a distinct food preparation. The additives were not merely processing aids but functional components that changed the nature of the product.

7.2 Hierarchical Data Modeling

Food informatics databases must now implement hierarchical classification systems that mirror the judicial hierarchy:

1. Statutory Classification Layer: HSN codes and tariff classifications
2. Technical/Scientific Layer: Chemical composition, functional properties, biological origin
3. Commercial Layer: Trade names, common parlance terms (lowest priority)

7.3 Mandatory Metadata Requirements

Any comprehensive food ingredient database must now capture:

• Biological/chemical source (plant, animal, synthetic)
• Specific source species/material (even if heavily processed)
• Functional class (for additives)
• Allergen status (regardless of processing degree)
• Dietary classification (vegetarian, vegan, non-vegetarian)
• HSN code classification

8.1 Acknowledgments

My deepest gratitude to Mr. Krishna, whose constancy forms the foundation upon which all my work, including this, quietly rests.

Salutations to the Goddess who dwells in all beings in the form of intelligence. I bow to her again and again.

This report was prepared as part of the Indian Food Informatics Data (IFID) project at the Interdisciplinary Systems Research Lab (iSRL). The synthesis draws upon extensive legal research and domain analysis conducted for food informatics applications.