LIMALEE-mah
Data Quality & Packaging
Lima is Latin for the file a craftsman used to finish a piece. Horace called careful revision “the labor of the file.” Ours is measured: every change is logged and tested against your specification.
Data quality you can inspect, not an undefined “AI-ready” label.
Billed as a fixed or milestone fee for defined work. No commission, and buying preparation does not appoint us as your broker.
Technical acceptance is not a legal clearance, buyer approval, or a promise of sale.
Is this for you?
- Data owners preparing an asset for licensing, through us, another intermediary, or a direct sale
- Teams that hold a buyer’s specification and need the dataset to meet it, with a report to show
- Companies that want to know what preparation an asset needs before deciding to license at all
The problem
“AI-ready” rarely means anything testable.
Buyers evaluate what they can measure. A dataset with unclear volume units, undocumented fields, silent duplicates, and unreviewed personal content is hard to evaluate and slow to license.
The fix is not a label. It is a specification: which fields must be complete, which duplicates are acceptable, how sensitive content is handled, and how each delivered record traces back to an approved source.
How we work
Agree the tests first. Then do the work.
We profile an approved, bounded slice first and report the baseline. Then we agree the transformations and the acceptance tests in a statement of work, and price against sources, record types, complexity, and manual review.
Every transformation is logged and every result reports its method and denominator. You keep the prepared output and the quality report whether you license through us, another intermediary, directly, or not at all.
Typical scopes
Where preparation fits.
Typical scopes, not past client work or results.
- Readiness
Readiness assessment
A diagnostic report and remediation plan, including a recommendation to go no further where warranted.
- Documents
Manual and procedure libraries
Consistent structure, language tags, deduplicated versions, and a data dictionary.
- Operations
Workflow and outcome packaging
Context, actions, timestamps, decisions, and outcomes linked, with no invented outcomes.
- Code
Repository preparation
Dependency and license review, secret scanning, and issue-to-fix linkage.
- Evaluation
Held-out evaluation sets
Separated evaluation assets and contamination checks within the defined scope.
- Annotation
Task-specific labeling
Instructions, adjudication, and sampled review, with agreement metrics reported.
- Privacy
Sensitive-data handling
Approved detection and transformation rules and a residual-risk review.
- Delivery
Release packaging
Versioned files, manifests with checksums, a dataset card, and a validation report.
What you receive
Inventory and profiling
A source register, volume definitions, field inventory, baseline quality report, and an exception map.
Normalization
Agreed schemas, encodings, units, timestamps, and identifiers, with a documented transformation log.
Deduplication and linkage
Exact and near-duplicate handling, event linkage, and documented logic for retained and rejected records.
Sensitive-data handling
Approved detection and transformation rules, restricted-source handling, and a residual-risk review.
Annotation and enrichment
Task-specific labels and instructions, adjudication, and sampled quality review. Synthetic additions are labeled synthetic.
Release packaging
Versioned files, a dataset card, a data dictionary, manifests with checksums, and a validation report.
How it works
Describe
Tell us what the dataset contains, how it was created, and what you want improved. No records or samples.
Specify
We agree the sources, the transformations, and the acceptance tests in a statement of work.
Prepare
Work runs under authorized access and bounded scope, with every transformation logged.
Test
The output is tested against the agreed criteria. Failures trigger the specified correction process.
Release
You receive the prepared output, the quality report, and the documentation, whatever you decide next.
Quality acceptance
Define “ready” as tested requirements.
Proposed measures, set per project. None is a universal standard.
| Measure | How it is reported | Acceptance approach |
|---|---|---|
| Schema validity | Records passing the agreed parser and schema, with version and denominator | A project-specific threshold; failures reported |
| Field completeness | Valid populated required fields, separating “not applicable” from missing | Fields and exceptions agreed before work |
| Uniqueness | Exact duplicates and near-duplicate clusters, with method and thresholds | Permitted duplication agreed by use case |
| Annotation quality | Sampling design, agreement or error metric, and adjudication | A documented QA plan; no claim of 100% accuracy |
| Linkage integrity | Valid references and chronological consistency; broken joins listed | Set against the task structure |
| Privacy testing | Tests performed, sampled findings, and method limitations | A separate release gate, not a promise of zero risk |
| Traceability | Delivered records linked to approved sources and transformations | Complete release documentation for the scoped output |
What you receive
A report you can hand to a buyer.
Each measure is reported with its method, sample, and denominator, before and after.
- Baseline
- The profile of the approved slice before any change.
- Transformation log
- Every change, in order, traceable to its rule.
- Final measures
- The agreed tests run on the output, with failures listed.
- Release documents
- The dataset card, data dictionary, and manifest.
Common concerns
What data owners ask before they sign.
- Meaningful repeats
- Deduplication never removes repeated events that carry meaning, such as recurring faults.
- Synthetic additions
- Anything generated is labeled synthetic. Guesses are never turned into factual labels.
- Anonymity
- We describe the tests run and their limits. We make no claim of universal anonymity or legal clearance.
- Ownership
- The agreement states your rights to the output and gives us no implied right to commercialize your data.
Who does what
Who decides, who advises, and who does the work.
Your team
- Authorize the sources and access
- Agree the specification and tests
- Review the quality report
- Accept against the criteria
Copia
- Profile and report the baseline
- Carry out the agreed transformations
- Run the acceptance tests
- Package and document the release
Afterward
Where clients go next.
Each route stands on its own. You choose, and nothing starts without your approval.
License through Copia
Use the prepared asset in a managed licensing process.
Sell it yourself
Take the outputs to a buyer you already know, with or without direct-sale support.
Keep it internal
Prepared, documented data is useful for your own teams whatever you decide.
Before the first call
What to have ready.
- A high-level description of each source
- Who can authorize access, and on what terms
- Any buyer specification you already hold
- Known problems: gaps, duplicates, or format changes
Related services
What “accepted” means
Measured against criteria you agree to in advance.
- Deliverables are tested against the criteria in the statement of work, with method and denominator reported.
- Material failures trigger the specified correction process and remedy.
- Your rights to the prepared output, our pre-existing tools, and third-party dependencies are stated in the agreement.
- Acceptance is not approval by an AI developer, legal permission to license the data, or a commitment by anyone to buy it.
Scope a preparation project
Tell us about the dataset.
We reply with the questions that shape scope, then a proposal with the fee and the acceptance tests.
- An engineer reads every request
- A reply by email within two business days of receipt
- No obligation until scope and price are agreed
Not ready to scope this? Discuss data licensing first: a free conversation about which route fits your dataset.
Where this fits
- Check Dataset readiness check
- Describe Fit discussion
- License Managed Data Licensing
- Prepare Data Quality & Packaging
- Sell direct Direct-Sale Enablement
Not ready yet? Dataset readiness check. After this: Managed Data Licensing or Direct-Sale Enablement.
Questions
Does preparation increase what a buyer will pay?
We do not describe quality improvements as an increase in license price. We measure quality against an agreed specification and report the result.
What is the difference between an assessment and a project?
A readiness assessment is a diagnostic report and remediation plan, including a recommendation to go no further where that is warranted. A preparation project carries out specific transformations and acceptance tests on a defined dataset.
Can we use the outputs with another provider?
Your agreement specifies your rights to the prepared deliverables and any tooling restrictions. A preparation-only scope does not appoint Copia as your broker.
Do you keep any rights to our cleaned data?
No implied right. The agreement states that we do not commercialize client data because we created the cleaned version.
How is a project priced?
Against sources, record types, complexity, manual review, and any buyer specification. Never against a future sale value.