Many Voices, One Label: How Pluralistic AI Flattens the World

TL;DR for operators

An AI project can interview communities, collect thousands of preference judgments, preserve several user perspectives, and still impose one rigid interpretation of the world.

That is the central warning in Rashid Mushkani’s AI Pluralism and the Worlds It Misses.¹ The paper names the failure ontological flattening: the process by which contested concepts such as safety, accessibility, inclusion, comfort, or belonging become fixed labels, measurable proxies, aggregation rules, or benchmark targets that are subsequently treated as neutral.

The problem is not simply that the model may produce an inaccurate answer. The deeper problem is that the system decides what kinds of answers are possible.

For operators, the paper changes the governance question from:

Did we collect diverse feedback?

to:

Who had authority to define the categories, decide what counted as evidence, choose how disagreement was aggregated, and require revision after deployment?

The paper supports this argument through a conceptual synthesis, aggregate themes from 11 expert interviews, and three companion cases involving public-space image generation, streetscape inclusivity prediction, and urban vision-language evaluation. These cases illustrate mechanisms; they do not validate the proposed governance framework or establish that it improves organizational outcomes.

The practical output is Pluralistic Lifecycle Governance, or PLG, a five-dimension audit scaffold covering:

1. ontological openness;
2. epistemic inclusion;
3. procedural authority;
4. evaluation pluralism; and
5. lifecycle accountability.

Organizations can use these dimensions to review datasets, benchmarks, procurement decisions, model documentation, deployment restrictions, appeal procedures, and decommissioning rules. They should not turn PLG into another reassuring composite score. The paper rather pointedly declines to provide one.

The workshop can be inclusive while the system remains closed

Consider a familiar participatory AI project.

A team recruits stakeholders, organizes workshops, records local concerns, and collects a respectable volume of annotations. The resulting model offers several outputs rather than one universal answer. The project report contains photographs of sticky notes. Everyone has apparently participated.

Then the engineering pipeline begins.

Hundreds of concepts are consolidated into six criteria. Ambiguous judgments become neutral labels. Neutral labels become inconvenient for a pairwise training objective. Group-specific assessments become averages. Lived experience becomes something visible in an image. A disputed concept becomes a benchmark column.

By deployment, the system may faithfully optimize the categories it was given. What has disappeared is the ability to dispute those categories themselves.

Mushkani’s paper argues that much of the AI pluralism debate begins too late. It asks how models should represent multiple preferences, values, populations, or acceptable outputs after the representational machinery has already decided what entities exist, which distinctions matter, and what counts as admissible evidence.

Pluralism at the output layer may therefore coexist with monism everywhere underneath it.

A planning assistant can provide recommendations from several stakeholder perspectives while assuming that “accessibility” is a property visible in street imagery. A generative model can produce diverse images of inclusive public spaces while training on a majority-voted preference signal. A benchmark can report performance across 30 social dimensions even where human annotators do not agree that some dimensions have stable visual answers.

The model is pluralistic. The ontology is not. Corporate governance does occasionally produce sentences that sound like philosophical jokes, but this one has operational consequences.

Ontological flattening happens before ordinary model error

The paper defines an ontology in practical rather than metaphysical terms. It is the set of representational commitments embedded in a system: what exists, which entities and relationships matter, what can be measured, and what qualifies as evidence.

Every AI system needs such commitments. A model cannot operate directly on the full ambiguity of social life. It requires categories, variables, labels, features, objectives, and evaluation criteria.

The paper is therefore not arguing that abstraction is inherently illegitimate. That would leave organizations with the admirably pluralistic option of building nothing.

Its concern is a particular kind of abstraction:

Ontological flattening occurs when situated, contested, embodied, or historically specific meanings are converted into restricted technical categories or proxies that are treated as neutral and made difficult to contest.

Mushkani identifies four conditions that distinguish flattening from ordinary compression:

1. The source concept is contested or context-dependent. Terms such as safety, belonging, inclusion, and comfort do not have one stable meaning across communities and situations.

2. The system fixes the concept into a restricted representation. The concept becomes a label, proxy, score, metric, aggregation rule, or benchmark target.

3. Disagreement or context is erased. Uncertainty becomes noise, abstention becomes missing data, minority interpretations disappear into an average, or local meanings are replaced by global categories.

4. Affected actors lack meaningful standing to revise the representation. People may appeal an individual output while remaining unable to challenge the schema that produced it.

This last condition is what makes the concept more useful than another general warning that “categories are political.” The paper’s distinctive contribution is to connect representational compression to lifecycle authority.

A label space is not merely a technical artifact. It is a decision about whose distinctions the institution will recognize.

A proxy is not merely an imperfect measurement. It is a decision about which evidence the institution is prepared to process.

An aggregation rule is not merely a statistical convenience. It is a decision about what happens to disagreement.

And a benchmark target is not merely an evaluation device. It is a decision about which version of the world model developers will be rewarded for reproducing.

A system can be pluralistic in output and monistic in procedure

The paper’s most useful distinction is between outcome pluralism and procedural pluralism.

Outcome pluralism asks whether the system can accommodate more than one answer.

Procedural pluralism asks who decided what the question meant.

The distinction corrects a plausible but incomplete view of participatory AI: that diversity of input naturally produces diversity of authority. It does not.

People can contribute data without controlling its interpretation. They can be consulted after the problem definition has been fixed. They can suggest categories that are later consolidated by researchers. They can annotate examples without deciding how disagreement will be aggregated. They can review outputs without having the right to pause publication, prohibit a use case, or require a schema revision.

Participation supplies voice. Governance determines whether anyone must listen.

The paper separates epistemic inclusion from procedural authority for precisely this reason. An organization may admit several forms of knowledge into a project—technical expertise, local experience, institutional records and community testimony—while retaining all binding decisions within the development team.

That project may be inclusive as a research process and closed as a governing institution.

The lifecycle is where pluralism is gradually compressed

Ontological flattening is best understood as a sequence rather than a single design mistake.

Situated experience
        ↓
Problem definition
        ↓
Category and schema design
        ↓
Labels, proxies and annotation rules
        ↓
Aggregation and training objectives
        ↓
Metrics and benchmark targets
        ↓
Scores, maps or generated outputs
        ↓
Institutional action
        ↓
Monitoring, revision or retirement

At each transition, the system becomes easier to compute and harder to contest.

A broad set of community concepts may become a manageable taxonomy. A taxonomy may become an annotation form. An annotation form may become a majority label. A majority label may become a loss function. A loss function may produce a model score. The score may appear in a procurement report or planning map stripped of the disagreements that shaped it.

No individual step necessarily looks unreasonable. Indeed, most will be defended as necessary simplification.

The lifecycle view asks whether those simplifications remain:

• bounded to the task;
• documented with their rationale;
• accompanied by rejected alternatives;
• explicit about known uncertainty;
• open to appeal and revision; and
• prevented from masquerading as complete descriptions of social reality.

Flattening is often less a dramatic failure than an accumulation of tidy decisions. Enterprise systems are unusually talented at making political choices look like dropdown menus.

Three cases expose the same mechanism at different stages

The paper does not present a new controlled experiment. Its empirical synthesis combines a purposive map of relevant literature, aggregate construct-refinement themes from 11 confidential expert interviews, and three companion cases selected from urban and public-space AI.

The cases occupy different points in the lifecycle:

Their purpose is not to prove that all pluralistic AI systems flatten social meaning. Nor do they demonstrate that PLG prevents harm. They show how the proposed failure mode can appear across several technical forms even when a project has made serious efforts to involve participants.

That evidentiary distinction matters. Conceptual papers are regularly promoted into empirical laws by readers who find caveats inefficient.

Case A: community criteria still pass through a ranking machine

The first case concerns a participatory dataset and alignment experiment for generating images of inclusive public spaces.

The underlying process was substantial. It ran for two years, involved 30 community organizations, and began with 634 concepts. Workshops consolidated those concepts into six locally defined criteria:

• Accessibility;
• Safety;
• Comfort;
• Invitingness;
• Inclusivity; and
• Diversity.

The final dataset contained 37,710 pairwise comparisons across 13,462 images. A model was then fine-tuned using Direct Preference Optimization, which learns from preferred and dispreferred output pairs.

The companion study reportedly found that the aligned model was favored over the baseline more often in held-out evaluation. The paper does not dispute this result. It uses the case to show that a successful alignment outcome can coexist with representational compression.

Several transformations occur between community deliberation and model optimization.

First, 634 concepts become six criteria. Consolidation is unavoidable at some level, but it changes what the system can represent. The governance question is who controlled that consolidation, which alternatives were rejected, and whether the schema remained revisable.

Second, multi-criteria judgments become pairwise preferences. A public-space image can perform differently on accessibility, comfort and diversity. Reducing the evaluation to “A is preferable to B” conceals the trade-off structure unless criterion-level judgments remain available.

Third, majority voting converts heterogeneous assessments into one target. This makes training tractable while deciding which disagreements the model will never see.

The preservation of neutral comparisons is therefore analytically important. Neutrality may reflect several different conditions:

• the images are genuinely equivalent;
• the evidence is insufficient;
• participants disagree;
• advantages on different criteria offset each other; or
• the question cannot be responsibly answered from the image.

These are not interchangeable states. Treating all of them as an inconvenient midpoint would discard information about the limits of the task.

The case supports a precise conclusion: better model preference performance does not establish procedural pluralism. It establishes that a model learned the training signal it was given. Models remain stubbornly literal about governance decisions disguised as data.

Case B: inclusivity becomes what the camera can see

The second case concerns prediction of streetscape inclusivity.

Its participatory material included 28 interview participants and 12 focus-group raters. Participants evaluated 20 streets at three points per street, producing 60 rated locations. Street-level image frames expanded the training material, and the resulting model was applied to roughly 45,000 street images.

The source study reports internal validation and test performance. Mushkani declines to treat the magnitude of that performance as evidence of broad generalization, because frames from the same streets, nearby viewpoints, or shared rating contexts may not be independent unless the split design groups observations appropriately.

That is an important methodological boundary, but it is not the main ontological issue.

The deeper problem is the transition from inclusivity as lived experience to inclusivity as visually inferable streetscape form.

Participants associated inclusivity with accessibility, activity, acceptance, safety and belonging. Some of these concepts leave visual traces. Sidewalk width, crossings, lighting, seating, vegetation and signs of activity can be photographed.

Others cannot be reliably recovered from pixels.

A street may appear accessible while remaining unusable because of temporary obstructions, poor maintenance or hostile enforcement. It may look active while particular groups experience surveillance or exclusion. It may display the aesthetic cues associated with safety while being experienced as unsafe by people whose histories the image does not contain.

The model can therefore become increasingly accurate at predicting a visually mediated rating while the organizational user gradually forgets that the target was mediated.

The danger becomes more serious when predictions are converted into maps. A heatmap gives a situated and uncertain construct the visual authority of infrastructure data. Once published, it may influence investment, policing, property narratives or neighborhood reputation.

The paper’s proposed remedies are correspondingly institutional, not merely statistical:

• limit map resolution where granular outputs could stigmatize places;
• show uncertainty rather than only point estimates;
• combine visual analysis with field audits and testimony;
• prohibit punitive or decontextualized uses;
• give community partners authority over publication and revision.

The ROI argument for these controls is not that they improve a generic ethics score. It is that they reduce the chance that a narrow proxy is operationalized as a complete decision variable.

Case C: low human agreement changes what model accuracy means

The third case evaluates whether vision-language models classify urban scenes consistently with local human annotations.

The benchmark contains 100 urban scenes, annotations from 12 participants representing seven organizations, 230 completed forms, and 30 evaluation dimensions. Seven vision-language models were compared using a deterministic zero-shot prompt contract.

Single-choice items were evaluated with accuracy. Multi-label items used Jaccard similarity. The reported macro-level scores ranged from 0.16 to 0.31.

That range looks conveniently rankable. The paper explains why it is not conveniently interpretable.

The dimensions have different numbers of options and different label structures, so there is no single uniform chance baseline against which every macro score can be compared. A score of 0.25 has a different meaning on a four-option single-choice task than on a multi-label task with several plausible combinations.

More importantly, some human annotation dimensions had negative Krippendorff’s alpha values. Negative reliability indicates disagreement worse than expected under the statistic’s chance model. In such cases, the benchmark does not offer a stable human target that a model can straightforwardly “match.”

This is not merely a noisy-label problem awaiting a larger dataset. The disagreement may reveal that the proposed dimension is ambiguous, insufficiently defined, highly contextual, or not visually decidable.

The study also found that dimensions closer to visible attributes were easier than appraisal-oriented dimensions, and that model performance co-varied with human reliability. The implication is almost embarrassingly sensible: models do better where people agree on what an image shows and worse where the benchmark asks them to infer contested social meaning.

Different uses of “not applicable” by humans and models further show that abstention is part of the construct, not an evaluation afterthought. A model that always chooses a substantive label may appear decisive while failing to recognize that the evidence does not support classification.

For benchmark owners, this case suggests several practices:

• publish human reliability by dimension;
• preserve abstention and “not applicable” semantics;
• disclose prompt and aggregation contracts;
• avoid ranking models on dimensions without stable targets;
• report sensitivity to alternative aggregation rules; and
• distinguish visible-description tasks from social-appraisal tasks.

A leaderboard cannot resolve an ontology. It can, however, format one attractively.

The cases motivate the framework; they do not validate it

The paper is unusually explicit about the evidentiary status of its material.

The literature map is purposive, not exhaustive. The 11 interviews refine constructs but are not used to estimate how common particular views are. Individual quotations, frequencies and transcript-level claims are not reported.

The companion cases are source-reported rather than independently reanalyzed. They were selected because they make the relevant mechanisms observable, not through an independent sampling frame designed to test competing explanations.

The paper therefore supports three kinds of conclusions with different levels of confidence.

This is not a weakness to be repaired through confident paraphrase. It is the proper reading of the contribution.

The paper offers a sharper diagnostic vocabulary and a preliminary governance scaffold. It does not offer a tested compliance standard, a maturity model, or evidence that organizations adopting PLG experience fewer harms.

Pluralistic Lifecycle Governance asks for evidence of power

PLG converts the paper’s conceptual argument into five audit dimensions.

1. Ontological openness

Question: Can affected actors contest the system’s problem definition, categories, labels, proxies and metrics?

Useful evidence includes:

• a rationale for the schema;
• records of contested categories;
• alternatives considered;
• a defined revision path; and
• documented responses to objections.

A data dictionary alone is insufficient. It tells auditors what the categories are, not why they deserve authority or who may alter them.

For businesses, ontological openness is relevant during problem definition, vendor selection and change management. Procurement teams should ask suppliers not only to describe model inputs and outputs but also to disclose how the target construct was defined and how customers or affected groups can challenge it.

2. Epistemic inclusion

Question: Which forms of knowledge count as evidence?

Technical measurements, expert analysis, operational records and lived experience may each reveal different aspects of a system’s target. Inclusion requires more than inviting diverse people into a process whose evidentiary rules have already been fixed.

The paper points to artifacts such as:

• recruitment rationales;
• compensation arrangements;
• accessibility support;
• multilingual materials where necessary; and
• inclusion of lived, technical and institutional knowledge.

This dimension also exposes a common failure in multimodal AI: using what the model can process as a substitute for what the organization needs to know.

Camera imagery may be scalable. That does not make it sufficient evidence of belonging.

3. Procedural authority

Question: Did participants receive decision rights or merely opportunities to speak?

PLG treats several forms of participation as materially different:

• consultation;
• co-design;
• delegated decision authority;
• escalation rights;
• pause or veto rights;
• appeal rights; and
• binding review requirements.

A workshop demonstrates interaction. It does not demonstrate authority.

The paper identifies procedural authority as especially likely to be overstated because project documentation often records participation without stating what consequences participation could produce.

For governance reviews, the relevant artifact is not the attendance list. It is the decision protocol:

• Which decisions could participants change?
• Who was obligated to respond?
• Could a release be paused?
• Could a use be prohibited?
• Who adjudicated unresolved objections?
• What happened when the organization rejected participant advice?

Without such obligations, participation may improve information while leaving control untouched.

4. Evaluation pluralism

Question: Does evaluation preserve disagreement, neutrality, abstention and low reliability?

Most model evaluation pipelines reward compression. A single ground-truth label is easier to score. One aggregate metric is easier to compare. A majority decision is easier to place in a benchmark.

PLG instead asks for:

• label distributions;
• neutral and abstention rates;
• inter-rater reliability;
• sensitivity to aggregation choices;
• subgroup analysis; and
• minority reports where disclosure is safe.

The operational point is not to abandon aggregate metrics. It is to prevent aggregate metrics from erasing the conditions under which they are meaningful.

A model’s accuracy against a high-consensus label means something different from accuracy against a category humans interpret inconsistently. Combining both into one number produces a clean leaderboard and a dirty inference.

5. Lifecycle accountability

Question: Who maintains pluralistic commitments after publication, procurement or deployment?

Participation often ends when the dataset is released or the model enters production. Yet deployment is when categories begin influencing institutional decisions.

Relevant evidence includes:

• an audit cadence;
• incident and recourse logs;
• revision triggers;
• version history;
• prohibited-use rules;
• named maintenance responsibilities; and
• decommissioning criteria.

This dimension moves pluralism from project methodology into operating governance.

A model may have been responsibly developed and later used for a purpose its participants never considered. A benchmark may remain in circulation after its categories become outdated. A public map may produce stigmatizing interpretations even when its original research use was bounded.

Lifecycle accountability asks who can intervene when the institutional meaning of an artifact changes.

A practical PLG review should resist numerical theatre

The paper suggests classifying each PLG dimension as:

• documented;
• partially documented;
• absent; or
• unknown.

“Documented” requires direct evidence of both process and authority. “Partially documented” means that some participation or reporting exists but rights, evidence requirements or revision arrangements remain incomplete. “Unknown” means the available evidence is insufficient.

Crucially, the framework advises recording auditor disagreement rather than averaging it away.

This is consistent with the paper’s thesis. A numerical PLG score would risk reproducing the very flattening the framework is designed to expose. Five contested governance dimensions would become one enterprise-friendly number, perhaps adorned with a green circle.

That does not mean PLG cannot support structured review. A practical audit record could use the following format:

This creates traceability without implying that governance quality has been scientifically measured to two decimal places.

What businesses should change first

PLG is broad, but its most immediate business value lies in a small number of lifecycle controls.

Treat alignment and annotation data as institutional records

Organizations typically document where data came from, who participated and which privacy controls were applied. The paper implies that they should also record:

• who proposed each category;
• who consolidated or rejected categories;
• how neutral and abstaining responses were interpreted;
• how disagreements were aggregated;
• which alternatives were considered;
• who approved the final schema; and
• who may revise it later.

This turns a dataset from a container of labels into a record of institutional decisions.

The change matters for generative alignment, ranking systems and preference learning. Training data does not merely encode what participants preferred. It encodes the process through which their judgments became optimizable.

Add schema governance to model and dataset documentation

Model cards and datasheets can reveal label definitions, collection methods and known limitations. The paper argues that disclosure is not the same as procedural openness.

Documentation should therefore distinguish:

• description: what the schema contains;
• rationale: why those categories and proxies were selected;
• authority: who approved them;
• contestability: who can challenge them;
• revision: what triggers change; and
• recourse: who must respond to objections.

A supplier saying “our categories are transparently documented” has answered only the first question. Transparency is useful. Authority is the awkward part.

Preserve disagreement before optimizing it away

Where a domain contains real trade-offs, organizations should retain more structure than one majority label.

Depending on the task, that may mean:

• criterion-specific preference outputs;
• distributions rather than winner labels;
• explicit neutral or indeterminate classes;
• reliability estimates;
• subgroup-specific summaries;
• alternative aggregation analyses; or
• multi-objective optimization rather than one composite target.

Not every disagreement should be preserved indefinitely in production. Decisions still need to be made. The paper’s replacement for universal consensus is provisional consensus: a bounded decision supported by public reasons, recorded dissent and a revision path.

The institution may choose one policy. It should not rewrite the record to imply that only one interpretation ever existed.

Restrict downstream uses of socially loaded proxies

When models infer contested social concepts from images, language traces or behavioral data, use restrictions may matter more than another round of predictive tuning.

Controls can include:

• lower-resolution publication;
• uncertainty displays;
• mandatory contextual evidence;
• prohibitions on punitive decisions;
• human review by affected-domain specialists;
• community approval for public release; and
• automatic review when outputs conflict with field evidence.

This is particularly important for maps, scores and rankings because their visual and numerical form encourages re-use outside the original scope.

Put revision and retirement terms into procurement

A vendor assessment should not end with accuracy, security and compliance documentation. For systems operating on contested social constructs, buyers should also ask:

• Can the label schema be revised?
• Can affected groups submit category-level objections?
• Who evaluates those objections?
• Are there response deadlines?
• What changes trigger revalidation?
• Which uses are contractually prohibited?
• Who can pause deployment?
• Under what conditions is the model or dataset retired?

These requirements convert pluralism from a design aspiration into a contractual operating condition.

The business case is risk control, not automatic consensus

The paper does not calculate financial returns from pluralistic governance. Any ROI interpretation is therefore a Cognaptus inference rather than a direct result.

The most plausible business value is reduced exposure to a specific class of failure: an organization operationalizes a narrow construct as though it were an uncontested fact.

That failure can surface as:

• a product that performs well against the wrong target;
• a public-facing score that stakeholders reject as illegitimate;
• a map or classification that stigmatizes groups or locations;
• a procurement dispute over undocumented assumptions;
• a benchmark that rewards unreliable inferences;
• expensive redesign after deployment; or
• inability to identify who authorized a consequential schema.

PLG may help organizations detect these problems earlier by forcing the relevant decisions into an auditable record.

It may also increase near-term costs. Meaningful participation requires compensation, translation, accessibility support, facilitation, review time and continuing maintenance. Revision rights create operational obligations. Multi-dimensional evaluation is less convenient than a single KPI.

The proper comparison is therefore not “governance versus free participation.” It is the cost of maintaining contestability versus the cost of discovering too late that a technically competent system has institutionalized the wrong world.

The paper does not tell us which side wins financially. It tells us which costs conventional model evaluation tends to omit.

The framework’s boundaries are substantial

Several limitations materially constrain practical interpretation.

First, the empirical domain is narrow. All three companion cases concern urban or public-space applications and rely heavily on visual data. These are useful settings for exposing the gap between visible form and lived social meaning, but other domains may have different mechanisms.

Healthcare may involve disputes over diagnostic categories and patient evidence. Education may involve contested definitions of achievement. Employment systems may operationalize suitability or potential. The PLG dimensions plausibly transfer, but the paper does not empirically demonstrate that transfer.

Second, the cases were selected for thematic relevance and artifact availability. This creates a risk of confirmatory selection. They illustrate the proposed mechanism rather than test how often it appears or compare it against rival explanations.

Third, the interview evidence is available only in aggregate. The 11 interviews refined distinctions such as epistemic inclusion versus procedural authority, but the paper does not provide transcript-level evidence, frequency counts or representative claims.

Fourth, the paper reuses results from companion manuscripts. It does not independently reproduce the model training, annotation analysis or benchmark calculations.

Fifth, PLG itself remains unvalidated. The paper provides:

• no demonstrated inter-rater reliability;
• no calibrated scoring rule;
• no prospective external application;
• no evidence that PLG changes procurement decisions;
• no evidence that it improves recourse;
• no evidence that it reduces harm; and
• no evidence that it produces better deployment outcomes.

These are not minor academic formalities. Before PLG becomes a standard audit instrument, it needs independent raters, recorded disagreements, adjudication procedures and longitudinal outcome tracking.

Finally, pluralism cannot mean accepting every proposed ontology. Some interpretations deny equal standing, conflict with protected rights, enable targeted harm or turn participatory processes into harassment.

The paper therefore argues for contestable boundary-setting rather than limitless inclusion. Institutions should state which claims are excluded, identify the rights or safety rationale, document who made the decision, preserve dissent where safe, and provide an appeal path.

Pluralism still requires decisions. It merely refuses to let those decisions impersonate natural law.

The real test is whether the categories can answer back

The paper’s central insight is simple enough to overlook.

An AI system does not become genuinely pluralistic because it produces many outputs, serves many users or collects many preferences. Those achievements concern variation inside a representational system.

The harder question is whether the representational system itself remains open to challenge.

Who defined inclusion?

Why was safety represented through those features?

What happened to neutral judgments?

Which disagreements disappeared into the average?

Who may revise the benchmark?

Can a community pause publication?

What event forces the organization to reconsider the model?

Ontological flattening gives a name to what happens when these questions have no operational answer. Pluralistic Lifecycle Governance provides a preliminary way to begin recording one.

For businesses, the lesson is not that every model needs an endless constitutional convention. It is that systems making claims about contested social concepts need more than diverse samples and flexible outputs. They need traceable category decisions, preserved uncertainty, bounded uses, response obligations and revision rights.

Many voices entering the pipeline are not enough.

Somebody must still be allowed to change the pipeline.

Cognaptus: Automate the Present, Incubate the Future.