01 / A question from your work
What do you need
to distinguish?
- You combine sources into one table. Were they observed over the same times and conditions?
- You rely on a summary to see whether something changed. Does it retain the member or process differences your task needs?
- You need to distinguish two conditions, but the readings look alike. Could changing when or how you observe help?
Leaving things out is necessary. What counts is whether the omitted difference matters for your question. An existing summary may already be enough.
We trace results back through records, build tools, and explore observations that could make unresolved differences distinguishable. Our projects are at different stages, not one finished product. See current fields.
02 / A shared row
Every ten minutes in the table.
Measured every ten minutes?
In the UCI Appliances Energy Prediction dataset, house energy records share rows with airport weather data. The source describes different routes into the table. Here are two:
Two source histories. One table.
When your comparison depends on timing
To compare changes over the same physical interval, first establish which measurement window each value represents. Sharing a row establishes the pairing, not that time relationship.
This case does not establish incorrect timestamps or invalidate the dataset’s original modelling task.
What remains unresolved
A populated temperature cell is not, by itself, evidence of a fresh measurement at that row’s timestamp. The reviewed material does not establish the timezone, daylight-saving treatment, the start/end anchoring of the energy window, or the precise weather interpolation and its supporting observations.
03 / A changed target
A positive claim count.
A zero training target.
In our reconstruction of a public insurance example, 9,116 records—26.8% of all positive-count records—have their claim count reset to zero after a missing amount is filled with zero.
How an absence changes a training target
- After matching the filesCount is positiveNo matching amount row
- Fill the missing amountAmount becomes 0This zero was filled, not recorded in the amount file.
- Apply a later ruleCount becomes 0If amount is zero, reset the positive count.
- Construct the targetFrequency target is 0The original positive count no longer reaches this target.
If your task needs absence and zero kept separate
If your question distinguishes “no claim” from “no matching amount record”, the resulting zero target alone cannot answer it. The original files remain available to retrace where it came from.
The example explains the reset; this reconstruction does not establish an error or real pricing effects. No model was trained.
Source versions and reconstruction limits
Separately authored reconstruction on freMTPL2 (OpenML 41214 and 41215, version 1), following scikit-learn 1.7.2. We did not run the original example. Why these records have no matching amount row remains unresolved. This describes changed fields and targets, not model performance or real accident counts.
04 / One piece of work
Start small.
Leave something others can use.
This could begin with one table, one measurement chain or one comparison. First agree what needs to be distinguished, then trace what the available records can support.
An unchanged method or an unresolved result can still be useful. We would keep the grounds for either, so the next person can see why the work took this direction.
Working together may change both the question and the method we first had in mind.
05 / Ways to take part
What might we build together?
Research and engineering
Help define a problem, design an observation or experiment, or build and maintain a tool or small prototype. Challenging our interpretation is part of the work.
Industry and laboratories
Bring a distinction that matters in practice. Help establish what existing methods already handle, what conditions can be observed, and what improvement would actually mean.
Financial support
Support identifiable work: who contributes time, what is investigated or built, which outputs are expected, and when the next decision is reviewed.
Employment, project collaboration, research support and equity investment are different relationships. Roles, periods, amounts and compensation arrangements are not yet fixed.
06 / How we work together
Let the work change our minds.
We would retain enough for others with appropriate access to inspect the work. When a result differs from what we expected, we would examine the record, the method and the expectation itself. The available evidence may not be enough to explain the difference.
We do not want failures renamed as successes, or a rejected hypothesis to erase findings that still stand. Commercial expectations cannot decide in advance what an observation must find.
We work with AI systems. Questions, evidence selection and method choices remain open to examination alongside the answer. See method and what we cannot see.
Before work begins, we would agree responsibilities, resources, compensation or expected returns, access, use and disclosure, and how the collaboration can be revised or ended.
07 / Who to speak with
Pingxin Wang
For an initial conversation about this invitation, contact Pingxin Wang at LFR Resonance Systems LTD via info@lfrfrequency.com.
08 / Begin a conversation
One decision, comparison
or unresolved question.
Tell us what you need to distinguish, what you can currently observe, and what would change if those situations turned out to be different.
You do not need a prepared dataset or agreement with our framing. Please do not send confidential or proprietary data in a first message.
Start with Bring a question, or write to us. To build, research or support the work, tell us what you would like to contribute.
info@lfrfrequency.com