NARC-tiny is a set of puzzles made by a program, not written by hand. The program read 98,113 stories from TinyStories, a public corpus of short children's stories written by language models (Eldan and Li, 2023). From each story it kept the events that can be drawn on a grid: someone or something appears, leaves, moves, goes up, comes down, is picked up, dropped, lost or handed over, a mood changes, a thing is transformed. A grammar then writes that chain of events twice, once as grids and once as plain sentences.

The grids. Each event is the change from one grid to the next. A character is a coloured bar, and its height is its mood: one cell sad, two neutral, three happy. An object is one cell, two if it is big, and grey if it is broken. The bottom row is the ground, and its colour is the place. One grid is hidden. The program hides a grid only where the visible grids leave several one-event completions open (at least three in this sample), the story picks one of them, and every drawing convention the answer needs is shown in another grid.

The two texts. Each puzzle is tested with two clues. The story is the TinyStory as written. The grammar text is the same chain of events written back as plain sentences by the grammar's templates, so it agrees with the grids by construction.

The sample. The program produced 21,864 puzzles. Three models (qwen3.8-27b, gemma-4-31b, gpt-oss-120b) judged whether every extracted event is asserted by its story, and all three passed 11,049 puzzles. The 200 puzzles here come from those 11,049: four to seven grids, 8 to 33 puzzles for each of the eleven kinds of hidden event. The sample covers the kinds of hidden event about evenly. The full set does not: in the 11,049, four of five puzzles hide a mood change.

N = NARC (solved only with both)   G = Grids sufficient   L = Text sufficient   × = Unsolved. The dots in each row show the better of the two texts for each model; open a row for both.

Order sensitivity: each NARC cell is re-tested with the grids in three shuffled orders (same text).   Strong = fails every shuffle   Partial = solves some   Weak = solves every shuffle.

Narrative sensitivity: each NARC cell is re-tested three times with the text replaced by an alphabetized list of its key terms (same grids).   Narrative = fails every keyword trial   Partial = solves some   Lexical = solves every keyword trial.

Scope: 200 puzzles · 11 models · one sample per cell at temperature 0. Grids only, story alone, story + grids and grammar text + grids: September 11, 2026 to October 6, 2026. Shuffled orders: September 15, 2026 to October 7, 2026. Both were run by a separate script that uses this site's prompts and grader; it kept the last few hundred characters of each response and counted an answer with no readable grid as not solved. Grammar text alone: September 29, 2026 to October 6, 2026. Keyword lists: September 29, 2026 to October 7, 2026. These two use this site's pipeline, which leaves an unreadable answer ungraded. The puzzles and the first two sets of runs are part of the paper under review (September 2026). The stories are from TinyStories (CDLA-Sharing-1.0 licence).