Data Readiness Levels

• Data Readiness Levels are frameworks that define and measure the progression of data from raw existence to AI-ready status through accessibility, fidelity, and contextual fitness.
• DRLs employ structured bands and multi-dimensional scoring to assess aspects such as data availability, representational accuracy, and task-specific readiness.
• They guide project planning and risk assessment by quantifying hidden curation efforts and aligning data processing with application-specific requirements.

Data Readiness Levels (DRLs) are frameworks for expressing how far a dataset has progressed from mere existence to effective use in analysis, machine learning, or decision-making. The term was introduced to provide a common language around the hidden work of locating, loading, validating, and contextualizing data, and later work recast DRLs as a qualitative banding scheme, a real-valued function tied to a question, a multi-dimensional AI-readiness score, and a stage-based maturity matrix for scientific AI (Lawrence, 2017, Guan et al., 2017, Brewer et al., 30 Jul 2025).

1. Origins and rationale

The original motivation for DRLs was organizational rather than algorithmic. Lawrence’s position paper begins from the observation that application of models to data is fraught, and that data-generating collaborators often only have a very basic understanding of the complications of collating, processing and curating data. The cited obstacles include poor data collection practices, missing values, inconvenient storage mechanisms, intellectual property, security and privacy. In that formulation, the “universal formula of machine learning” is rendered as

$$\text{data} + \text{model} \longrightarrow \text{prediction},$$

but the central claim is that the left-hand side is often under-specified: project overruns can occur due to failure to account for the amount of time required to curate and collate, and a common language is needed for assessing the readiness of a particular data set (Lawrence, 2017).

This project-management and communication role was carried into later applied NLP work. In that literature, the state of data is identified as an important reason for failure in applied Natural Language Processing projects, and the gap between academic research in NLP and its application to problems outside academia is argued to be rooted in poor mutual understanding between academic researchers and their non-academic peers. The proposed use of DRLs in that context is to improve communication regarding the accessibility, validity, and utility of data between researchers and external stakeholders (Olsson et al., 2021).

2. The original three-band framework

In the original DRL framework, readiness is divided into three bands with potential sub-levels: Band C, Band B, and Band A. Band C concerns Accessibility and asks, in effect, “Does the data exist, and can I load it?” At the lowest end, C4 denotes “hearsay data”: one has only second-hand claims that the data exist, existence is unverified, format is unknown, legal or privacy constraints are unknown, and storage location is unclear. At the upper end of the band, C1 means that the data have been located, permissions and ethics approvals granted, the data are machine-readable, and storage-topology issues have been resolved.

Band B concerns Faithfulness & Representation and asks whether the loaded data actually reflect what they claim. B1 is reached when missing-value conventions have been identified and, where possible, correctly encoded; noise characteristics or sensor error models have been characterized; data-entry or spreadsheet-induced errors have been found and removed; units and scales are consistent; and simple Exploratory Data Analysis has been performed so that an analyst understands key biases or quirks in the data.

Band A concerns Contextual Fitness and asks whether the data can answer a precise question. A1 is defined by the presence of a clear task or question, completion of any remaining annotations or supplemental data collection needed for that task, attention to statistical-design considerations such as sampling scheme and potential confounders, and a state in which the data are ready to be plugged directly into modeling or decision-making pipelines for that task. A crucial feature of this formulation is task-specificity: a data set may be A1 for one question and only B1, or lower, for another if new variables or different biases come into play (Lawrence, 2017).

3. Quantitative and multi-dimensional reformulations

A more formal line of work defines DRL as a scalar-valued function of a dataset and a question. In that formulation,

$$DRL(D\mid q)=f\bigl(N(D),\,B(D),\,R(D\mid q),\,C(D)\bigr)\in[0,1],$$

where $N(D)$ measures Noisiness, $B(D)$ measures Believability, $R(D\mid q)$ measures Relevance to a question $q$, and $C(D)$ measures internal Coherence. The paper proposes explicit information-theoretic metrics for Relevance and Coherence: average cosine similarity in an LDA topic space for $R(D\mid q)$, and Document Disparity based on Jensen–Rényi divergence for coherence-related assessment. It also lists desiderata such as easy computability, stability or continuity, scalability, minimal supervision, and discriminative power (Guan et al., 2017).

A distinct line of work treats readiness as a multi-dimensional AI-evaluation problem. In the 2024 AIDRIN prototype, readiness is reported as a continuous score in nine orthogonal dimensions: Completeness, Outliers, Duplicates, Privacy, Feature correlations, Feature relevancy, Class imbalance, Fairness, and FAIR principle compliance, with discrete DRL bands left for future standardization. AIDRIN 2.0 provides one concrete thresholded construction by defining

$DR(D)=[Q(D),B(D),F(D),P(D)],$

with $Q$ for Data Quality readiness, $$DRL(D\mid q)=f\bigl(N(D),\,B(D),\,R(D\mid q),\,C(D)\bigr)\in[0,1],$$0 for Bias readiness, $$DRL(D\mid q)=f\bigl(N(D),\,B(D),\,R(D\mid q),\,C(D)\bigr)\in[0,1],$$1 for Fairness readiness, and $$DRL(D\mid q)=f\bigl(N(D),\,B(D),\,R(D\mid q),\,C(D)\bigr)\in[0,1],$$2 for Privacy readiness, and an overall score

$$DRL(D\mid q)=f\bigl(N(D),\,B(D),\,R(D\mid q),\,C(D)\bigr)\in[0,1],$$3

subject to $$DRL(D\mid q)=f\bigl(N(D),\,B(D),\,R(D\mid q),\,C(D)\bigr)\in[0,1],$$4. In that version, Level 0 is “Unusable” for $$DRL(D\mid q)=f\bigl(N(D),\,B(D),\,R(D\mid q),\,C(D)\bigr)\in[0,1],$$5, Level 1 is “Low readiness” for $$DRL(D\mid q)=f\bigl(N(D),\,B(D),\,R(D\mid q),\,C(D)\bigr)\in[0,1],$$6, Level 2 is “Medium readiness” for $$DRL(D\mid q)=f\bigl(N(D),\,B(D),\,R(D\mid q),\,C(D)\bigr)\in[0,1],$$7, and Level 3 is “High readiness” for $$DRL(D\mid q)=f\bigl(N(D),\,B(D),\,R(D\mid q),\,C(D)\bigr)\in[0,1],$$8 (Hiniduma et al., 2024, Hiniduma et al., 22 May 2025).

The resulting literature does not converge on one single mathematical object. It includes qualitative labels, scalar functions, weighted dimension vectors, and thresholded categorical levels.

Formulation	Core representation	Output
Lawrence DRLs	Bands C, B, A with sub-levels	Qualitative label
Question-specific DRL	$$DRL(D\mid q)=f\bigl(N(D),\,B(D),\,R(D\mid q),\,C(D)\bigr)\in[0,1],$$9 from $N(D)$0	Scalar in $N(D)$1
AIDRIN 2.0	$N(D)$2, weighted $N(D)$3	Levels 0–3
Scientific AI DRLs	DRL$N(D)$4 to DRL$N(D)$5 plus processing stages	Matrix or stage state

4. Stage-based readiness for AI and scientific computing

For scientific AI, DRLs have been reformulated as a five-level progression from raw data to fully trainable assets. One framework defines DRL$N(D)$6 as Raw, DRL$N(D)$7 as Cleaned, DRL$N(D)$8 as Labeled, DRL$N(D)$9 as Feature-engineered, and DRL$B(D)$0 as Fully AI-ready. In parallel, it defines five processing stages—ingest, preprocess, transform, structure, and shard—and places datasets in a $B(D)$1 maturity matrix $B(D)$2, where $B(D)$3 describes the state of a dataset at readiness level $B(D)$4 and processing stage $B(D)$5, with cells $B(D)$6 designated “N/A” (Brewer et al., 30 Jul 2025).

REDI turns that conceptual matrix into an automated five-stage pipeline: ingest, preprocess, transform, structure, and output. Ingest moves data from raw to usable when all files can be read without error and schema inferred; preprocess moves data to a transformed, schema-consistent state when domain-specific checks pass; transform produces fully defined and normalized features without leakage; structure makes data “model-plug-and-play,” including proper shape and split-balance checks; and output serializes AI-ready artifacts with metadata validation. REDI defines a stage-wise maturity vector

$B(D)$7

and the entire pipeline maturity is given as $B(D)$8. The same work also reports strong-scaling speedup and efficiency,

$B(D)$9

with preliminary climate results showing efficiency $R(D\mid q)$0 up to 100 nodes, and identifies file I/O as the dominant pipeline cost, with format selection a first-order optimization lever (Wilkinson et al., 2 Jul 2026).

A related but non-identical formulation appears in space life sciences. There the proposed readiness stack is three-tier rather than five-level: FAIR Data, AI-Ready Data, and Space-Ready Data. FAIR data aim at human reuse; AI-Ready data add structured, machine-actionable formats, provenance, automated metadata curation, readiness metrics, and high-throughput programmatic access; Space-Ready data add spaceflight-specific contextualization, small-$R(D\mid q)$1 robustness guarantees, mission-context metadata, federated-learning readiness, governance-aware access controls, and certification by a neutral international body. Each tier is described as strictly contingent on satisfying the prior tier’s requirements (Costes et al., 27 Jun 2026).

5. Assessment practices, workflows, and case studies

In the original DRL literature, the framework was explicitly tied to planning, effort estimation, and risk assessment. The PMLR proceedings case study reports that converting 3,200 PDF proceedings into a fully parsed pandas-ready data frame took roughly six working days even with no privacy constraints, thereby illustrating hidden software-engineering effort in Band C work. The Uganda malaria monitoring case study shows how missing-value biases introduced during Health Management Information System aggregation produced a false negative in spatial correlations, and how obtaining disaggregated data and re-modeling led to recovery of expected spatial signals and successful deployment of an early-warning system (Lawrence, 2017).

AIDRIN 2.0 turns readiness assessment into an explicit pre-training workflow. The stated sequence is: ingest raw dataset $R(D\mid q)$2; compute dimension-specific sub-metrics; normalize; compute each dimension score; compute overall $R(D\mid q)$3; assign DRL; and, if readiness is below threshold, flag the dataset for remediation such as data cleaning, bias mitigation, fairness intervention, or privacy enhancement. One decision rule is to aim for DRL$R(D\mid q)$4 before large-scale model training, with per-dimension minima $R(D\mid q)$5 and $R(D\mid q)$6 for regulated scenarios. In the FLamby Heart Disease cross-silo PPFL case study, before exclusion of one client the global model accuracy was $R(D\mid q)$7 and overall $R(D\mid q)$8 (Level 1) due to $R(D\mid q)$9 and $q$0; after removing Client 3, global accuracy became $q$1 and overall $q$2 (Level 2) (Hiniduma et al., 22 May 2025).

Visual analytics has also been presented as a method for raising readiness. One extension retains the C–B–A structure, adds an explicit Band B question on distribution shifts during data collection time, and splits Band A into seven task-solution aspects $q$3 through $q$4, including model assumptions, feature–target information, feature importance, task performance, tuning difficulty, robustness under distribution shift, and trust for deployment. The same work maps readiness questions to concrete chart types for continuous, categorical, text, and time-varying data, and recommends a workflow of selecting one readiness question and one variable, generating the mapped visualization, inspecting for anomalies, discussing findings with stakeholders, and iterating when Band C or B corrections occur. Its reported use cases include train-traffic disturbance duration prediction, public-transport arrival and departure prediction, and coalition structure and task assignment, with concrete outcomes such as removal of pre-2016 data after discovering a reporting change, and use of feature-importance and hyperparameter-loss visualizations to support model selection and trust (Tiger et al., 2024).

At leadership-computing scale, REDI demonstrates closed-loop readiness assessment and validation across climate, proteomics, materials science, and nuclear fusion. The reported examples include ClimaX with 5.4 TB and 4,766 samples, OpenFold with 4.3 TB and 250k samples, HydraGNN with 10 TB and 482M samples, and XGC1 with 106 TB and 7,089 runs. Validation is reported against domain-expert references, with examples such as climate Pearson $q$5 and MAE $q$6, and exact or near-exact agreement for the other domain cases (Wilkinson et al., 2 Jul 2026).

6. Standardization, misconceptions, and governance

A common misconception is that “Data Readiness Levels” refers to a single universally accepted ladder. The published literature instead contains several non-identical constructions. Lawrence’s position paper deliberately stops short of introducing a quantitative readiness metric; the 2024 AIDRIN prototype reports continuous scores and leaves exact thresholds for future standardization; the 2025 AIDRIN 2.0 paper defines one particular thresholding scheme; and the 360-degree survey explicitly states that its taxonomy does not refer to a set of hierarchical readiness levels but to a structured catalogue of data-quality dimensions and associated metrics (Lawrence, 2017, Hiniduma et al., 2024, Hiniduma et al., 22 May 2025, Hiniduma et al., 2024).

A second persistent theme is that readiness increasingly includes governance and scientific usability, not only cleanliness. SciHorizon-DataEVA organizes AI-readiness into four complementary dimensions—Governance Trustworthiness, Data Quality, AI Compatibility, and Scientific Adaptability—and aggregates normalized atomic element scores into per-dimension scores and a global DRL by a weighted combination

$q$7

with equal weights by default. Its operational contribution is a hierarchical multi-agent workflow over a directed cyclic graph, in which dataset-aware evaluation specification, adaptive tool orchestration, and review-driven verification and self-correction are combined into a scalable evaluation pipeline for heterogeneous scientific data (Liu et al., 29 Apr 2026).

In the space-data literature, governance becomes a readiness condition in its own right. The proposed neutral international coordinating body is assigned responsibilities that include maintaining global metadata schemas, authorizing and auditing AIDRIN scoring criteria and thresholds, certifying MCP tools and agents, governing federated-learning frameworks, maintaining the central data registry and digital-seal authority, and convening cross-jurisdictional working groups. This suggests that, in some domains, the highest readiness levels are inseparable from provenance, certification, access control, and durable institutional coordination rather than from preprocessing alone (Costes et al., 27 Jun 2026).

Across these formulations, DRLs function less as a single metric than as an evolving family of schemes for making data preparation explicit. The unifying idea is stable: readiness is not exhausted by possession of data, but depends on accessibility, representational faithfulness, task fit, and—depending on domain—provenance, fairness, privacy, machine-actionability, scalability, and governance.