A confusion matrix tells you what someone perceived. It cannot tell you how they got there — whether they processed the two dimensions one after another or at the same time, whether they waited for both or stopped as soon as one was settled.
That isn't a limitation of the method. It's a limitation of the data. Below, three of the five architectures produce the same confusion matrix — not similar, the same. No amount of accuracy data, and no cleverer statistics, could ever separate them. But they finish at visibly different times. Response times are not a nice extra here; they are the only channel the information is in.
Counts alone — just the aggregate 4×4 table.
| A1/B1 | A1/B2 | A2/B1 | A2/B2 | |
|---|---|---|---|---|
| A1/B1 | 84 | 7 | 8 | 1 |
| A1/B2 | 6 | 85 | 1 | 8 |
| A2/B1 | 9 | 1 | 84 | 6 |
| A2/B2 | 1 | 9 | 7 | 83 |
Counts + response times — every trial, individually, with a timestamp.
| stimulus | response | rt (s) |
|---|---|---|
| A1/B1 | A1/B1 | 0.61 |
| A1/B2 | A1/B2 | 0.44 |
| A2/B1 | A1/B1 | 0.98 |
| A2/B2 | A2/B2 | 0.52 |
| … one row per trial, hundreds of rows | ||
That's a genuinely different data-collection commitment, not just "the same experiment, analysed more thoroughly" — every trial needs its own timestamp, not just its outcome. The per-trial table gets compressed down to 5 quantiles per cell (80 numbers total) before it reaches the model; that compression is shown live in the The response times card below once you run an experiment.
Exact response probabilities implied by the space.
One curve per stimulus, over the times of correct responses. Each curve is scaled by that stimulus's accuracy, so the area under it is P(correct) — a curve that is both low and fast is the signature of guessing, not of skill.
The small ticks under each curve are the five quantiles GRIN actually reads (10th, 30th, 50th, 70th, 90th). That is the entire RT input: 5 numbers per cell, 80 in total. Everything the model concludes about processing, it concludes from those ticks.
Architecture and the accumulator parameters are things only the RT model can see at all — there's no "counts-only version" of those to compare against. But the GRT question underneath — separable? independent? — is something BOTH models answer. Same simulated data, same question, two models: this is that comparison, live, not just asserted.
The same perceptual space, the same accumulator settings, run under all five architectures. Look at the accuracy column, then look at the timing column.