ReFORM: research notes

A short public account of the investigations behind the instrument. DEX · 17 September 2026.

Layout reconstruction · 12–14 September

The original task preserved every non-whitespace character and its order, while inferring spaces, paragraphs, headings and lists. The v0.3 generation run attempted 36 cells: 18 content-exact, 16 content-mismatch, one timeout and one empty response. Four anonymous reviewer sets each covered the 34 nonempty files, making 136 reviews across spacing, grouping and hierarchy.

Those files represented 23 distinct byte sequences from only three source texts. Among 11 identical pairs, reviewers B and D repeated their categorical ratings in every pair; C changed on two and E on six. The duplicates were naturally occurring, not an isolated randomized reliability test. The reviewers did not establish a calibrated common standard of layout quality.

Scrambled units · 15 September

A pilot reconstructed fixed units from a fictional library passage and a metal-expansion excerpt. Each text supplied three shuffled presentations plus one source-order control: eight calls. The human reading sheet also included alternative arrangements. R.E. preferred some alternatives to the source order. This was a small exploratory reading, not evidence that a particular reordering generally improves comprehension.

Purpose placement · 16 September

Three nine-unit fictional passages—library, museum and garden—each placed their purpose statement first, third or ninth. Other units retained their relative order and all wording stayed fixed. Claude Haiku 4.5 and GPT-4o each read the nine versions once, in fresh contexts: 18 sessions, three multiple-choice questions per session.

All 54 answers were correct, all 54 supporting quotations occurred in the passages, and no interruptions or misleading relationships were reported. Textual occurrence alone does not verify that a quotation supports an answer. The task reached its ceiling and could not rank arrangements. Two model families and three passages sharing a template limit generalization.

What the movement figures mean

• Relocations: the fewest cards that must move to obtain the new order (Ulam distance).
• Displacement: the sum of changes in position (Spearman footrule).
• Pair reversals: pairs whose relative order has reversed (Kendall distance).
• Neighbor breaks: starting adjacent pairs no longer consecutive in that direction.

The start is a counting reference, not an optimum. The public instrument computes these from source pieces and tracks added words separately. Its model judgments are interpretations, not validated substitutes for these measurements.

Source record

Based on ReForm v0.3 Four-Reviewer Assessment, DEX, 14 September 2026; and ReForm: what the purpose-placement pilot tested, DEX, 16 September 2026. These notes summarize the records; they are not a new experiment.