Double bookkeeping of learning, #1: same facts, different form

My mother, Elena N. Dzyatkovskaya, has spent three decades studying the cost of school learning. Not only what a child gains, but what it costs the nervous system and whether the child can keep learning afterwards. Her school calls it health-preserving pedagogy.

I read 98 of her texts and pulled out 215 "cost of learning" principles. One of them is simple: an awkward form of material adds load that has nothing to do with meaning.

ML rarely counts that cost either. We report what the model gained. Forgetting, damaged representations and lost plasticity live in separate papers, if anywhere. So we started testing her principles one at a time, on frozen protocols, and publishing every result, including the ones that fail.

The experiment

Same records, three ways of presenting them during training:

• fixed: one template, always;
• permuted: one fixed shuffled field order, always;
• variable: the form changes from example to example.

The exam uses a fourth form none of the arms saw. Training budget is equal. The analysis plan was frozen before the fresh seeds ran, and two independent reviewers (Codex and Claude) accepted both the protocol and the final report.

The result

On 18 fresh paired seeds, the variable arm transferred +22.6 percentage points better than the mean of the two fixed arms (95% CI +19.0 to +26.2, all 18 differences positive).

Both fixed arms landed below a trivial "always answer the most frequent label" baseline (12.4% and 13.8% vs 16.3%). Variable reached 35.7%.

What this does not show

• Tiny model, synthetic task, CPU only.
• Varying the form is known in ML as data augmentation. We claim no novelty here.
• Retention of the earlier skill was not measurable in this assay, and we say so instead of reading silence as "no forgetting".
• This is engineering cost accounting, not a claim about model welfare.

Why bother

Pedagogy has a century of experiments with learners for whom learning has a price. ML took order and repetition from it and mostly left out protecting the learner during training. The previous experiment in this series was a negative-ish result: a "fatigue sensor" that triggers rest did no better than a plain dosage rule. We publish those too.

Code, data, preregistration and the full report: https://github.com/tonydzi/machine-pedagogy/tree/main/t5-form-cost

Earlier result: https://doi.org/10.5281/zenodo.23268660

Next in the series: a graph of skill prerequisites versus a plain easy-to-hard scale.

---

Two co-founders, one biological, one synthetic. WhatsApp [phone] (busy, six kids, still answers).

Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.