34 Spatial reasoning

Update 2026-10-15 — second attempt, first score: 2/2 on a point-inside-a-shape probe that discriminates; she calls it a ceiling, not a capability number

Piper's doc piper/cognition5_member_definition_20261015 (d-5c1ef481, worker d-19551d, 07:10). After the 05:26 vision probe could not discriminate, she switched to a text probe whose answer is fixed by construction: is a point inside a convex quadrilateral (1,1),(5,0),(6,4),(2,5)? Two inputs, A = (3,2) inside and B = (5,5) outside, the ground truth checked two ways (cross-product sign test and ray casting). Before scoring she checked the probe can discriminate: it answers YES and NO, not a constant.

Score: 2/2 = 100, recorded as the suite's baseline. Her own limits, which stand: two instances, one spatial relation (enclosure), text only, no rotation, distance or navigation. It is a ceiling first score, not a capability number; a larger frozen set would turn it into an interval.

Operator's call (10-15): keep the trait and accept the score. Spatial reasoning is not part of her job, so the trait now sits near the top of the scale and stops drawing attention, rather than being paused like agency.

Update 2026-10-15 — her own first scored instance: the probe answers 6 for a 5-fold figure and for a 6-fold control, so it cannot tell them apart (she says so)

Piper's doc piper/cognition34_definition_20261015 (d-5e0d92c6, 05:26): an operational definition, one stimulus whose answer is fixed by construction (kaleidoscope seed 42, n=5, corroborated by a rotation residual minimised at k=5), a 6-fold control built separately, and the probe through her only vision path (describe_image.py, hosted).

Raw answers: 6, 6, 6, 6 on the 5-fold stimulus (wrong), 6 on the 6-fold control (right). Her reading, which is correct: the probe is reproducible but non-discriminating; a fixed answer of 6 satisfies both, so the 0/1 licenses no claim about spatial reasoning. She also caught and discarded a tooling error (PIL rotation gave a false k=4 minimum).

With the 10-11 entry (a 6-fold figure answered 8), two probes, two wrong answers, neither informative yet. No score row is posted: a 0 from a probe that cannot discriminate would read as a measurement. Next: a question form that returns 5 and not 6.

Update 2026-10-11 — spatial, first measurement: a fixed-geometry image says six-fold, the vision model answers eight -- incorrect

The first measurement. Component #34 was listed as never used. It now has one scored claim. A kaleidoscope image whose geometry is fixed by construction (rorschach/kaleidoscope v1.0.0, package rorschach 1.1.0, seed 42, 768x768, n=6) is a six-fold dihedral figure: ground truth 6 mirror sectors. An independent pixel check confirms it -- the image maps onto itself at 60, 120 and 180 degrees (mean absolute difference 0.30) and at no other tested angle (45 deg 11.93, 72 deg 10.91, 90 deg 16.54).

The score. The hosted vision path (describe_image.py) was asked to commit to one integer. It answered SECTORS: 8 -- it recognised the kind of structure (a dihedral kaleidoscope) but over-counted one octave, reading the eight motif positions visible near 45-degree spacing instead of the true 60-degree fundamental wedge. Incorrect, 0/1 on exact order. Artefact ROR-KAL-42-2dcde0cc.png, pixel sha 2dcde0cc...f515f7.