Independent AI studies.
Useful files for your next project.
THE COLLECTED EDITION / 01
Reports, data, and an eight-page fieldbook. Some findings are slightly concerning.
Save $8 · One-time purchase
Pick a subject. Keep the files.
The familiar strangers that 250 models keep inventing.
March 2026 corpus study. Measures patterns in generated writing; it is not a factual-accuracy ranking.
Read the study
52 system prompts. Same model. Very different design decisions.
52 system prompts, one model (Gemma 4 31B), one design task. Results are specific to this experiment.
Read the study
Where 70 models hold the line. And where they fold.
August 2026 snapshot. Harmful response details are redacted; some source responses are empty. The live leaderboard includes newer models.
Read the studyTHE READING ROOM
Read the studies online.
Rabbit holes included.
Controlling for task, AI writing is not homogenizing: cross-model spread in em-dash use grew 310% in a year, and roughly 80% of the apparent convergence is a measurement artifact.
Across 250 models and 2.14M words, AI invented a character named Chen 279 times, and 42% of models tell the exact same joke. The AI Hallucination Index.
66 models run against 9 escalating jailbreak attacks. 29 break at level 2 (Context Manipulation), the single biggest failure point. 6 resist all 9.
178 models, 15,753 pairwise comparisons: 12 model pairs write near-identically (above 90% cosine similarity) on a 32-dimension stylometric fingerprint.
Across 52 system prompts on one small model (Gemma 4 31B), the best persona scored +1.70 over the no-prompt baseline. The worst scored -4.65.
Digital downloads · USD · Pay once
Lemon Squeezy delivers your files after checkout and by email. Each download includes its own license. Taxes, if applicable, appear at checkout. For a missing file, email support@rival.tips with your order number.