A real-data comparison
One dataset, two approaches
We took the same 20,000 respondents and the same 50-item measure and generated factor scores in two ways: a conventionally fitted model, and METER.
Conventional approach
Fit a confirmatory five-factor model directly to the dataset.
METER
Use the same five-factor structure, but score the dataset with a frozen pretrained model that has never been fitted or calibrated on these respondents.
The target was identical in both cases: each person’s score on each of the five factors.
The task
- Instrument
- 50-item IPIP personality measure
- Respondents
- 20,000
- Factors
- 5, specified in advance
- Target
- Person scores for each factor
The same respondents, items and factor structure were used for both approaches.
Two ways to get there
Conventional model
- Provide the factor structure
- Choose and fit an estimator
- Check estimation
- Extract factor scores
METER
- Provide the same factor structure
- One pretrained inference pass
- Factor scores and capability checks
METER estimates no new model parameters from the target dataset.
What disappears is the estimation, not the thinking.
The result
| Conventional model | METER | |
|---|---|---|
| Target-dataset fitting | Yes | No |
| Runtime | 662.1 s | 0.776 s |
| Person-score agreement | median r = 0.9791 | |
Agreement ranged from r = 0.9596 to 0.9839 across the five factors. In this evaluation METER closely reproduced the person scores from the separately fitted model without being fitted to the target dataset.
How to interpret this result
This is evidence of transfer, not superior accuracy. The conventional model and METER are being compared with one another; real data do not provide a known true factor score against which either method can be judged.
The runtime figures also apply to this specific benchmark. Simpler conventional scoring problems can already be extremely fast.
Where METER stops
METER did not discover the five-factor structure. It was supplied in advance.
For a multidimensional measure like this one, the current release also does not return item parameters, model-fit statistics or standard errors, and the two methods’ factor correlations differed by up to 0.15. Conventional model fitting remains appropriate when those outputs are required.
This boundary is deliberate: METER returns only capabilities supported by its current evidence base.
Why this matters
Traditional psychometric modelling usually starts again with each new dataset.
METER tests a different approach: can useful measurement capabilities be learned in advance and then transferred to new data without fitting another model each time?
This example shows that person scoring can transfer under the conditions tested here. If that generalizes across a sufficiently broad range of measurement problems, measurement becomes something that can increasingly be reused rather than rebuilt.