Executive Summary

As AI agents search enterprise documents and query data, data quality work reaches further into daily operations. Organizations need agents to understand where information comes from, when it applies and how it may be used. AI-ready data is a general term for data prepared for AI use; AIRD (AI-Ready Data), the Korean draft framework and demonstration program reviewed here, translates these needs into documented data descriptions, quality measures and purpose-specific readiness requirements.

The public recall case illustrates data prepared for agent search and aggregation. Its source passed a quality assessment, while the derived RAG pack completed linking and checks but has its purpose-fitness verdict on hold. Exact matches in the published recall comparison rose from 23/42 to 42/42, within a test whose answer key is linked to the rules used to build the pack.

Pebblous sees this development as an opportunity for two cooperating products. Data Greenhouse diagnoses and improves data against multiple requirements. Agent Greenhouse is the next R&D step: improving capabilities that fall short in operational records and tests, then testing and approving new agent versions. Better-prepared data supports agents, and improved agents help cultivate the next data version.

Key Numbers

The first two numbers are scores from the published comparison, and the last two describe the measurement scope and the size of the data. The scores split by task; the measured items split by data type.

Source: AIRD evidence, test conditions and limitations (measured 2026-10-21) · AIRD indicator list and type applicability table · description of the recall retrieval pack.

23/42 → 42/42

Exact match on the recall questions

7 questions repeated, with an answer key tied to the rules that built the pack

28/28 → 28/28

Hydrogen and fuel questions do not move

14 questions. What preparation bought differed from task to task

3–13

Mandatory indicators per data type

3 for images, 13 for structured tables and time series. All 16 are not uniformly required

843 chunks

Size of the retrieval pack

The published source CSV has 844 rows. This is not a count of products or training demonstrations

1

This Dataset Was Checked for Search, Not for Training

The case starts from ordinary public data. The Korean Agency for Technology and Standards publishes a domestic recall report on the national open-data portal, filed under dataset number 15040695. Row by row it records products that broke a safety threshold, products whose labeling did not match, and descriptions tied to accidents. Open the published CSV and there are 844 rows, three fields used in the work itself, and one identifier that points at the row. The public pack is prepared for retrieval and aggregation.

What changed is the side that repackaged it. The derived pack splits the same content into 843 retrieval chunks and tags each chunk with the substance involved and how many times over the limit it ran. It also keeps the identifier of the source row each chunk came from. Seven evaluation questions ship with it, and the one asking for a manufacturer is built so that the answer is withheld for lack of evidence. This is not a textbook for a model to memorize. It is input for a program that has been handed a question and has to look things up and count them.

Between 844 source rows and 843 chunks there is a gap of one. Saying what caused it takes the row-by-row conversion rule, and the published description on its own does not settle it as "one empty row was dropped." When the number travels, it is worth carrying the note that 843 counts chunks, not products and not training demonstrations.

How the source data becomes retrieval input Korean recall CSV (KATS) 844 rows 3 working fields + 1 row ID Cut into chunks + tag Gap of 1 — different counting units Retrieval chunks 843 Substance + exceedance tags + source row ID ✗ Not training material ✓ Input for search and aggregation Source: recall source data (844 rows) and the published recall retrieval pack description (843 chunks)
▲ An original Pebblous diagram, drawn from the published figures

This one case unsettles a common assumption. "Preparing data for AI" usually reads as tidying up training material. What was prepared here is working data for search and aggregation, and its quality and purpose-specific readiness were assessed separately. Where an agent digs through in-house documents to build an answer, data that never entered training decides what the model says.

2

Connecting Quality Measures With Readiness for AI Use

AIRD names data that meets the conditions for being used by AI. The framework published here splits that readiness in two. One half is Quality-Ready: the data itself is measured against set dimensions and indicators and has to land at or above a given tier. The other is Purpose-Ready: whatever the data is meant for carries its own conditions, and those have to be met separately. The published guide counts six purposes, namely machine learning, deep learning, retrieval-augmented generation, fine-tuning, knowledge graphs and statistical analysis. The same document puts one more step in front of those two, asking whether the data can be found at all, and describes the result as three gates: Discoverable (G1), Quality-Ready (G2) and Purpose-Ready (G3).

Each gate is assessed once the required evidence and criteria are available. An empty later assessment should be read alongside the readiness state and the reason it was not assessed. An unregistered purpose profile, for example, does not mean that the source failed its earlier quality assessment.

Status is the part to read carefully. What the published guide points to as the governing text is TTAK draft v0.95 together with vocabulary table v0.10.5, and enactment is expected in December 2026. It is neither a finished international standard nor a certification carrying legal force. The packs distributed by the demonstration site carry their own version numbers for vocabulary, schema and operator thresholds, so folding the draft version and an individual implementation into one name blurs the record.

The other side of the pairing holds a different status. ISO/IEC 5259-2:2024 was published in November 2024, and its formal name is data quality for analytics and machine learning, part 2, data quality measures. It carries measures over from ISO/IEC 25012 and 25024 and widens them into a machine learning context, so properties that cause trouble before and after training, such as the diversity of a dataset or its auditability, sit in the measure list.

Shrinking the 5259 series to a single measurement document is not accurate either. Alongside terminology and measures, the series holds data quality management requirements, processes and governance documents. The split that says ISO inspects files while AIRD looks at use is convenient to explain, but it erases the ground the two frameworks share.

Readers who want the quality characteristics and measures of 5259-2 item by item will find them in our cheat sheet, and how the series divides into documents and what schedule it has moved on is written up in the standardization roadmap. This article does not repeat that material and looks only at where AIRD touches it.

3

What Each Framework Asks For

A practical way to tell the two frameworks apart is to write their questions side by side. What gets evaluated, what evidence is demanded, and who sets the pass mark all differ. The table below puts those questions in five rows and records, on each row, how Pebblous intends to answer them.

Question ISO/IEC 5259-2 This AIRD framework How Data Greenhouse responds
What is evaluated? Quality characteristics and measures chosen to fit the purpose Readiness requirements per data type, plus quality and per-purpose readiness Connect thresholds and measurement tools to the intended use
What material is required? The data plus the basis for each measure: answer keys, rules, thresholds The data plus machine-readable description, versions, diagnostics and operating records Check the submitted material and the missing evidence at intake
Can a calculation be reused? The definition, unit and denominator of an item govern The definition and applicability conditions of that indicator have to be checked against it Reuse only calculations whose definitions match, and keep the provenance
How is a pass decided? Evaluated against the intended use and the quality requirements By operating guidance, threshold values and the purpose profile Separate measurement from verdict, and record the threshold version and who signed off
Does it guarantee real AI results? Quality measurement alone does not guarantee success on every task A readiness verdict and the effect on actual work are confirmed separately Compare the results and the cost of search, aggregation and training as well

▲ A Pebblous comparison and design note. It is not an accredited equivalence table between the two specifications.

The easiest trap when reading the table is the names. Completeness, consistency, accuracy and a fair number of other words appear in both frameworks. Where the formula and the thing being measured differ under one name, though, the values are not the same number. The moment a completeness score computed on one side is copied into a report on the other, the number survives and the evidence disappears.

4

Applying Seven Quality Dimensions to Each Data Type

The published guide names seven quality dimensions, which are completeness, consistency, accuracy, timeliness, validity, uniqueness and machine readability, with 16 indicators under them. All 16 do not apply uniformly. Structured tables and time series carry 13 mandatory indicators, prompt-and-response pairs carry 8, documents carry 5, and images carry 3. The minimum measurable indicator set (MMI) contains four candidates; only those applicable to the data type are measured. What an image dataset is asked for cannot be the same as what robot multimodal time series are asked for, where physical time, coordinate frames and the meaning of an action have to be verified together.

Mandatory indicators by data type (out of 16 total)

Structured tables & time series
13
Prompt-response pairs
8
Documents
5
Images
3
▲ An original Pebblous diagram · mandatory indicator counts by data type, out of 16 total | Source: AIRD indicator list and type applicability table

The scoring is not one block either. Diagnostic maturity, which records how much has been measured, the quality tier, and the per-purpose readiness level each answer a different question. The quality tier uses the lowest dimension score with the operating guideline's tier ranges and required-dimension pass conditions. The scale runs from Tier0 to Tier4, five steps, and clearing the quality gate takes every mandatory indicator for that type measured and a tier of 1 or higher. A tier carries its meaning only when the pass mark travels with it. The per-purpose readiness level, by contrast, can be judged only once a profile for that purpose is registered, and the published vocabulary document states that the procedure is still to come.

This is where the everyday damage happens. Write an item that has not been measured yet, an item that does not apply to that type, and an item with no threshold value yet all down as zero, and sound data looks like it failed. The three differ in kind and have to be recorded one by one, and the example threshold values in the guide do not belong in an official tier calculation.

The misreading also runs the other way, in the claim that AIRD has no thresholds for deciding a pass at all. One published verdict document refutes it. The quality verdict for the recall source states that it applied operator threshold profile TP-KR-2026-1.0, and records the fingerprint of that profile and the time of the verdict alongside it. So rather than asking in a single sentence whether thresholds exist, check the dimensions of the standard, the thresholds of the operator, and the availability of a purpose profile one at a time.

5

Source Quality and Retrieval Readiness Need Separate Checks

The verdict record for the recall case splits into two lines. The source dataset cleared the quality gate on 11 October 2026, and the verdict document states that operator thresholds were applied and the result confirmed at Tier3 (Gold). The pass mark sits at Tier1, which puts it two steps up a five-step scale. The retrieval pack produced from the same data, by contrast, is marked as on hold (판정 보류). Its published description says linking and checks are complete but the purpose-fitness verdict is deferred, and its purpose-fitness (G3) row reads 해당 없음 (N/A) with the note "to be judged in the next release." Despite the N/A label, this means the verdict has not been made yet, not that the criterion does not apply.

Where the verdicts diverge — source and derived pack Discoverable Quality-Ready Purpose-Ready Source recall data Public CSV, 844 rows Pass Pass Tier3 (Gold) · 2026-10-11 No record Derived retrieval pack 843 chunks Published Does not inherit the source tier On hold judged in the next release Source: the quality verdict for the source data (2026-10-11) and the published description of the recall retrieval pack
▲ An original Pebblous diagram, drawn from the states recorded in the published verdict documents

That divergence settles one rule. The quality tier the source earned does not become the tier of a product processed out of it. Errors or omissions introduced while chunks were cut and tags attached were never part of the measurement of the source. The same reasoning is why quality, readiness for a given use, and results in actual work cannot share one line.

Numbers for that last one are published too. In the comparison measured on 21 October 2026, exact matches on the recall questions rose from 23/42 under the original condition to 42/42 under the pack condition, which is 54.8% to 100%. In the same test the hydrogen and fuel questions came out 28/28 under both conditions. What preparation buys differs from task to task, and lifting one score out on its own turns into overstatement.

The 100% carries a caveat the operator attached to it. Because the recall answer key was tallied with the rules that built the pack, the pack condition is structurally favored on exact match. The same page notes that the question count is small and no statistical test was run, and that among the 10 per-criterion tables broken out in the linking experiment, AIRD came out behind on 2. Publishing the limits alongside the result is worth crediting. The problem is the copy that takes the number and leaves the limits behind.

6

Connecting Industry Demand With Two Cooperating Greenhouses

Requirements that make quality, provenance and usage conditions readable by AI are a welcome industry development for Pebblous. More explicit requirements clarify the work customers need in diagnosis, improvement, evidence and re-evaluation after change. The opportunity extends across training datasets and the operational information that agents consult.

The role Pebblous gives Data Greenhouse is not to pick one standard but to take in several. Written out in two layers, taking them in looks like this.

6.1The Governance Layer Holds Versions, Provenance and Approvals

Requirements, tolerances, data versions, provenance and usage conditions differ by customer and purpose. The governance layer connects them with approvals and assessment history. An integrated report can present ISO-based measurements and AIRD readiness assessments together while keeping each result traceable to its definition, version, evidence and approval responsibility. Calculations with matching definitions can be reused; each assessment remains tied to the criteria applied.

6.2Neuro-Symbolic Diagnosis Separates Detection From Judgment

What a neural network does well is gather things that mean something similar and pick out skew in a distribution, gaps in it, and cases that look wrong. Whether such a candidate breaks a working threshold is judged on explicit rules and domain knowledge. Rule checks that can be settled by computation are run by reproducible tools. Not treating the judgment of a large language model or an embedding similarity as a conformance verdict on its own is the premise of this design. A verdict is a place where evidence is demanded, and a score whose evidence cannot be retraced does not become a verdict.

What to do when a new standard arrives also comes out as a list in this structure. Register the applicable items, the measurement tools, the evidence required and the conditions for a verdict, then diagnose existing data versions again. Being able to build and hand over a data version that satisfies both 5259 and AIRD is the direction the product is headed.

Where this stands today is worth saying plainly. The measurement and improvement technology Pebblous has built up is ISO 5259-based, and on the AIRD side the adapter and the evaluation implementation are at the stage of a verification plan. That does not mean automatic evaluation of every item is finished or that a certification has been obtained. Conformance in any individual case is something to state against the specification and profile applied, the data type, and the verification result.

The widening scope is the other thing this case points out. Beyond training material, the documents used in retrieval-augmented generation, the tables staff query at work, the input that goes into tools and the material used to evaluate agents all face the same question. What is this data ready to be used for, and where is the evidence for that judgment written down?

6.3 New Requirements Also Define the Capabilities Agents Need

An agent needs to judge whether the information it finds is sufficient for the task. It should check provenance and the relevant date, request additional evidence or abstain when the record is insufficient. Distinguishing source quality from the purpose readiness of a derived RAG pack is one example of this capability.

Agents working in Data Greenhouse can apply newly registered rules and tools. When operational records and tests reveal gaps in knowledge, skills or decision procedures, Agent Greenhouse is being developed to improve those capabilities, independently test the resulting agent version and return an approved version to work. This connects applying a requirement with improving the agent itself.

Better-prepared data supports agent decisions, and improved agents can manage the quality and evidence of the next data version more effectively. The business goal is to reduce repeated review, rework and approval effort as customers introduce new data and requirements. Building on the data quality technology available today, Pebblous plans to test the contribution of this cycle through controlled comparisons.

7

A Proposed Test That Runs Both on One Dataset

There is a limit to settling the relationship between two standards in prose. Hang both on the same data and where they overlap and where they leave gaps shows up as numbers. Below are the six steps such a comparison would take.

  1. Fix first the versions and hashes of the source and derived data, along with the terms of use and the intended purpose.
  2. Map the measures of 5259 against the AIRD indicators by definition, unit, denominator and required evidence.
  3. Measure the items that can be computed, and record separately what does not apply, what has not been measured, and what has no threshold.
  4. Issue a quality verdict as far as the operating guidance reaches, and hold the purpose verdict if no purpose profile exists.
  5. Compare the original, a version built from the published rules, the existing AIRD pack and an improved version under the same questions, model, tools and cost conditions. Review the answer key independently of the improvement rules.
  6. Do not look at accuracy alone. Count wrong answers, correct abstentions, traceability to the source, review time and the cost of updating.

The caveat in step five matters most. Build the answer key with the same hand that wrote the improvement rules and you reproduce exactly the structural advantage the published test disclosed about itself. The same goes for correct abstention in step six. Stopping instead of inventing an answer when the evidence runs short is a capability the evaluation questions in the recall pack already demand.

Those six steps are a proposal. What this article checked goes as far as a static review of the published files, their structure and the verdict documents. It is not the result of evaluating every item in either standard, of running an agent again, or of a performance improvement experiment.

R

References

The comparison uses official standard descriptions, AIRD guidance and the data and assessment files published by the demonstration project. All references below are public.

Standards

AIRD published guides

Recall case evidence

Pebblous blog