GessoEvaluate a workflow dataset by tracing complete cases, checking what the records actually show and establishing the rights for your intended use. File counts alone do not establish useful scale. For agent post-training, evaluation or RL environments, the useful question is whether the data contains the context and evidence your task requires.
Gesso sources and prepares connected business records for AI labs. This guide explains what to request when assessing a dataset, including the evidence we discuss in buyer packages. Availability, volumes and permitted uses depend on the specific opportunity.
Define the workflow your model needs
Describe a task the model or agent should perform, the information available at the start and the outcome you want to evaluate. For example, an investigation task might require an initial support request, access to relevant working records and evidence of how the issue was resolved. A collection of final responses serves a different purpose from a history of the investigation.
Tell the provider which domains, languages, tools, formats and kinds of outcome evidence matter. Specify whether you want historical source records, connected examples with annotations, or constructed tasks and environments. Each requires different preparation and acceptance criteria.
Gesso’s current sourcing areas are industrial software, technical industrial distribution, HVAC and refrigeration, accounting and CAS, materials testing and failure analysis, and coatings, adhesives and sealants. Share your requirements to discuss fit and relevant opportunities.
Trace one case from request to outcome
Ask for a sample that lets your team follow one workflow across the underlying records. The following fictional support case uses invented labels.
| Stage | Record in the fictional case | What the buyer checks |
|---|---|---|
| Starting request | Support case S-104 reports a scheduling problem | The original problem and the context available at the beginning |
| Investigation | Issue E-22 records reproduction and a proposed fix | Source links, ordered work and the basis for the decision |
| Change | Release R-8 records the tested change | Whether the change was proposed, tested or released |
| Outcome | A later reply on S-104 records the customer’s observation | The source of the outcome and any remaining uncertainty |
Check whether consistent replacement labels preserve those links after de-identification. Ask which source files support each step and which steps are missing. A generated summary can help navigation, but the package should identify it separately from the original records.
Count connected cases and document gaps
Request counts at the level your task uses. A package may contain source files, extracted records, unique content and connected cases; those are different measures. Ask how duplicates, automated messages, broken references and excluded material affect the reported volume.
Useful questions include:
- How many cases contain both a starting request and a recorded outcome?
- Which cases have intermediate decisions, attachments or cross-tool context?
- How were repeated exports, quote revisions and related visits grouped?
- Which counts are seller-reported inventories, and which have been validated after processing?
- What is missing, and how does that affect the intended task?
A source inventory can support an early proposal. It should remain distinguishable from processed, validated delivery counts.
Check rights and privacy evidence
Ask who controls the source material, what authority supports licensing it and what the license permits. Review uses, restrictions, exclusivity and the treatment of third-party material for the actual dataset.
Privacy preparation needs its own evidence. Request a description of de-identification, exclusions, review coverage and remaining limitations, including attachments and metadata. Gesso arranges independent privacy review before buyer access; the proposed delivery should explain how that review applies to the material being offered. Our data-protection page describes the process.
Rights and privacy answer different questions. A record can require exclusion because of a contract or trade secret even after personal identifiers are removed. NIST’s De-Identification of Personal Information also explains that de-identification involves a balance between useful data and re-identification risk. Ask what was assessed for the particular release rather than relying on a broad label such as “anonymized.”
Keep evaluation answers out of the starting context
Historical outcome evidence can support task design, but it does not automatically provide reliable scoring. Decide which information belongs in the agent’s starting context and which belongs in a separate answer or evaluation set.
In the fictional support case, the final customer confirmation would reveal information the agent did not have at the beginning. A task built around investigating the original problem should control when that information becomes available. Check attachments, filenames, summaries and linked records as well as the main prompt for answers that would make the task trivial.
If an environment is required, agree the tool interfaces, starting state, reset behavior and scoring criteria separately. Ask how the evaluator handles multiple valid solutions, unresolved historical cases and source records that do not establish a clear outcome.
Agree a sample and acceptance criteria
Before a larger delivery, agree a sample that tests the properties your team needs: readable source content, preserved links, usable context, documented rights and outcome evidence. Identify the sample-selection method and how well it represents the proposed delivery.
Write acceptance criteria in terms your team can check. Examples include required fields, valid case links, allowed formats, treatment of duplicates and the evidence needed for an outcome label. Document known gaps and how corrections or replacements will be handled.
Dataset documentation helps make those decisions repeatable. Google Research’s Data Cards paper describes structured documentation of dataset origins, collection and annotation methods, intended uses and other information needed across a dataset’s lifecycle. For a Gesso opportunity, ask for the buyer package and discuss how its evidence maps to your acceptance criteria.
To begin, send your sourcing requirements or book a Gesso Data Sourcing call. Include the workflows, starting context and outcome evidence you need, without sending confidential research or private source records.