Keep the evidence categories separate.
An official statement is evidence that a claim was made. It is not automatically evidence that the claim is true. Records must preserve that distinction.
01 The distinction
An answer can sound right.
Make it earn the conclusion.
A fluent response can hide a weak inference. It can repeat a source as though it were independent confirmation, mistake an incentive for proof, or defend a forecast after its assumptions fail.
I design data around those pressure points. The task is to preserve distinctions, connect explanations to observable evidence, and state exactly what would require a different conclusion.
What is established?What is inferred?What would change it?
02 / Selected work
Inspect the
reasoning standard.
Three fictional institutional scenarios from my analytical work. Each makes the structure of the task visible: the claim, the competing explanation and the evidence that matters.
Edited previews adapted from a longer design portfolio. All three organizations are fictional. Explore the reasoning structure here; request a sample for your intended use.
Download the previews ↓03 The standard
Difficulty,
with a purpose.
The target is a model that can maintain analytical discipline when an answer becomes inconvenient. These are the standards I build a brief around.
Make alternatives do real work.
Competing explanations should change what evidence the analyst seeks. An objection that cannot affect the conclusion is decoration.
Carry the constraints through.
Long responses must retain their scope, time horizon and uncertainty. More detail should not quietly turn a conditional judgment into a certainty.
Define the conditions for revision.
A forecast needs observable counterevidence and a stated consequence for the conclusion. Missing evidence must remain distinct from disproof.
04 From brief to dataset
Specify the behavior.
Then design the data.
A focused commission begins with the failure you need to address and the standard by which you will judge the work.
Define the task.
Agree the domain, target behavior, difficulty, evidence boundaries and intended use. Identify what an unacceptable answer looks like.
Review a pilot.
Inspect a small set against explicit acceptance criteria. Reject or revise examples that collapse distinctions, overclaim or fail the brief.
Set the delivery terms.
Agree scope, record format, metadata, review requirements and usage rights before expanding the work. Judge model impact in your own evaluation.
Evaluate a pilot against your own tasks before scaling. Dataset design and measured model improvement are separate questions.
05 The offer
Buy the data.
Commission the difficult brief.
For teams that need more than plausible text: analytical scenarios with explicit constraints, evidence distinctions and a reason to revise the answer.
Analytical datasets
Dataset sales · By inquiryDiscuss access to work in institutional analysis, narrative analysis and scenario forecasting. Request a sample and establish the collection, packaging and rights relevant to your use.
Bespoke commissions
Your domain · Your acceptance criteriaBring an analytical failure your system struggles with. Commission a focused pilot of scenarios and responses designed around that failure, with scope and review criteria agreed in advance.
The right team
Research · Data design · EvaluationI am also open to joining a team where designing demanding synthetic data is central to the work. Show me the problem, the standards and the people building it.
06 Begin with the work
Bring the task.
Ask for a sample.
Tell me what you are building, where the reasoning breaks, and what would make a dataset useful. We can start with a concrete example.
ecuna.ec@gmail.com ↗