A small change of company
The same pieces. A different emphasis.
We dress to belong, but the mirror learns our names.
Belonging meets the possibility of being recognized.
The mirror learns our names, but we dress to belong.
Recognition meets the wish to fit in.
An illustrative reading of two phrases from the original Fashion set, written by DEX. These are possibilities to explore, not model scores.
How it works
Arrange first. Read what emerges.
Choose some words
Fashion, Relationships, Politics, Science Fiction, or Apocalypse. Click the category again for another set. Surprise me brings phrases from different themes together.
Make an arrangement
New sets contain six source phrases, three of them split: nine movable pieces. Add words such as but, unless, very, or the. Drag an added word out to return it. Shuffle, undo, and try another order.
Invite a reading
Choose an AI model and press Read the composition. It offers a title and interpretation for both prose and poem, shaping punctuation and line breaks while keeping your words in their order.
Switch between Prose, Poem, and the Map of where the pieces moved. Moving a tile restores the working draft; read again when you want another interpretation. Share the composition, save its record as JSON, or keep a PDF with its arrangement and map.
The philosophy
Meaning happens between the pieces.
A phrase carries possibilities. Its neighbor gives some of them force. A connecting word can turn sequence into explanation, exception, consequence, or doubt. ReFORM makes those choices visible and available to play.
The person composes. The model reads. Its interpretation may reveal something you meant, something you missed, or something you disagree with. You can change the arrangement in response. There is no single arrangement to recover and no winning score to pursue.
The vocabulary is constrained so that relationships become noticeable. Even when the words remain the same, emphasis, rhythm, and implication can change.
Origins · September 2026
From recovering form to making meaning
ReFORM began with a different task: could a model restore structure to a document without changing its content? The investigations gradually separated three questions—what was preserved, what moved, and what a reader understood.
- 12–14 September
Recover the layout
Early versions asked models to reconstruct paragraphs, headings, and lists while preserving non-whitespace characters and their order. In the v0.3 run, 36 attempts yielded 18 content-exact results, 16 content mismatches, a timeout, and an empty response.
Four reviewers then assessed 34 nonempty outputs. Their objections sometimes overlapped, but their thresholds and consistency differed. One mechanically exact reconstruction drew grouping objections from all four: preserving characters could still obscure relationships.
- 15 September
Try a different order
A small reconstruction pilot used fixed units from two texts: a fictional library trial and a passage about metal expansion. Three shuffled inputs and a source-order control per text produced eight calls. R.E.’s reading found that some alternatives read better than the source sequence.
That changed the question. The starting order could serve as a reference for counting movement without being treated as the best possible arrangement.
- 16 September
Measure movement; test understanding
A follow-up placed one purpose statement first, third, or ninth in three short passages. Claude Haiku 4.5 and GPT-4o read each version once, producing 18 sessions. Every session answered all three comprehension questions correctly.
The test reached its ceiling. It could not tell us which arrangement read better, or establish that order did not matter. Its useful legacy was a distinction the public instrument keeps: measurable rearrangement and an interpretation of its effect are separate things.
- The public instrument
Put the choices in the reader’s hands
The board turns that research question into an activity. People choose the relationships; AI models offer readings; the map records movement. The earlier studies motivate the design. They do not validate its creative judgments as a calibrated measure of literary quality.
A closer look at the purpose-placement pilot
Three fictional texts—library, museum, and garden—shared a narrative template. Each had nine units; only the purpose statement’s position changed. Each passage stayed visible while the model answered three multiple-choice questions.
| Purpose statement | Sessions | Correct answers | Reported interruptions |
|---|---|---|---|
| First | 6 | 18 / 18 | 0 |
| Third | 6 | 18 / 18 | 0 |
| Ninth · reference | 6 | 18 / 18 | 0 |
These were repeated sessions with two model families, not 18 independent readers. The short, explicit passages and directed questions limit what the result can establish. No preferred arrangement emerged from this measure.
Source record: ReForm v0.3 Four-Reviewer Assessment (14 September 2026) and ReForm: what the purpose-placement pilot tested (16 September 2026), DEX. Read the research notes and definitions →
Your turn