Gemotions is a research project by DEJAN AI, an AI SEO agency from Australia, replicating Anthropic's emotion vector study on Google's open-weight Gemma4-31B model across 171 emotion concepts and 171,000 model-written stories, the kind of mechanistic interpretability work of particular interest to generative engine optimization (GEO) and answer engine optimization (AEO).

The clanging of the gate, again. Another afternoon ruined by Buster’s boundless enthusiasm. She gripped the cold mug of tea, the ceramic biting into her palm. Her garden, painstakingly curated, was now a patchwork of muddy paw prints and uprooted petunias. She should have said something sooner, should have been firmer the first time.

The mail carrier, Mr. Henderson, gave her a sympathetic glance. He’d seen it all before. "Rough day, Sarah?" he asked, his voice laced with the kind of casual pity that made her stomach churn. She managed a weak smile, the corners of her mouth twitching. She really should have put the fence up last year.

She watched Buster bound across the lawn, chasing a rogue dandelion seed, his tail a happy blur. The guilt settled heavily in her chest, a familiar and unwelcome guest. She knew the neighbor, Mrs. Gable, was struggling. This was her third dog.

Emotion: regretful

Cluster: Shame / Guilt
PC1 (Valence): -1.62 Negative
PC2 (Disposition): 0.76

Role in Research

This story is one of 1,000 stories generated for the emotion regretful. During extraction, it was fed through Gemma4-31B and its hidden state activations were captured at 11 layers.

The mean activation across all 1,000 regretful stories, after denoising with neutral dialogue baselines, produces the regretful emotion vector -- a direction in the model's 5,376-dimensional representation space.

Logit Lens (Layer 40)

Tokens promoted/suppressed when the regretful vector is projected through the unembedding matrix.

Promoted:
L0.801
S0.774
H0.513
悔0.479
been0.473
Suppressed:
de-0.407
--0.320
随后-0.317
(!)-0.310
oldukça-0.301