ttywisp

my current stack

yearly: 4-strain flu shot

weekly:
testosterone enanthate 250mg
tirzepatide 10mg

daily:
rosuvastatin 5mg
oxandrolone 20mg
escitalopram 20mg
10,000 ui d3
1.5mg methylcobalamin
1.5mg methylfolate
15mg p5p
1200mg nac
plus iron, iodine and a bunch of other shit that’s not as important supplementation-wise

Holy fuck, you’re stupid.

God damn.

weaponized autism time.

this app store sudoku is trash. i intercepted its traffic with a MITM proxy, replayed its API requests, and built a Go grader using HoDoKu-inspired solving techniques. then i graded all five difficulty levels for 90 daily puzzles: 450 grids total.

the results:

- its “players online” counter is made up. an old puzzle sat at 664–665 “online” players for 40 minutes while its “played” count went up by one.
- of the 90 puzzles the app labeled Hard, my grader rated 86 Beginner, 3 Expert, and 1 Diabolical.
- of the 90 puzzles the app labeled Master, its highest difficulty, my grader rated 40 Beginner.

the ten easiest Master puzzles I analyzed are all solvable with full houses and singles. four Hard puzzles took XY-wings or an XY-chain.

results attached.

AI-written report

Benchmarked two zero-output-token "System-1" decision APIs, TypeSafe's jev-1.14.0 and Google's gemini-decision-1 (via OpenRouter), across 21 Zelda games. The experiment covered 840 counterbalanced pairwise choice evaluations and three passes through a 126-question, 5-axis score rubric.

All 1,218 Gemini evaluations cost $0.026, including 840 pairwise choices and 378 rubric scores. Each 126-question rubric pass cost $0.004. A single pass recovered >92% of pairwise tournament rank variance, reducing ranking complexity from O(N²) pairwise comparisons to O(N) rubric scoring.

On exact repeats, Gemini recorded 0.000 mean absolute error and 0/126 decision flips. Jev recorded 0.034 MAE and 7/126 flips. Jev also quantized probabilities to 0.01 increments, with exact 0.00 values making unadjusted log-ratio calculations undefined. Gemini returned fine-grained probabilities without the observed quantization.

Both APIs exhibited positional bias. Jev favored the second choice option by 4.37%; Gemini favored opt_1 by 4.63%. Rearranging JSON key insertion order left Gemini's results unchanged. Counterbalancing requires swapping assignments between neutral option labels.

Pairwise structure was overwhelmingly transitive. Subtracting target entropy from cross-entropy yielded KL divergence. Hodge decomposition attributed 97.4% of pairwise variance to a 1D transitive gradient, leaving 2.6% in cyclic structure. Adding cycle terms produced zero cross-validated improvement. Jev's tie behavior fit Davidson constant-tie scaling. Gemini's tie mass declined toward the tails, with an ILR slope of -0.366. An Ordered Logit model explained 95.4% of its KL deviance.

The five rubric axes had mean inter-axis correlations of 0.725 for Jev and 0.400 for Gemini. Jev placed Tears of the Kingdom first in its pairwise ranking. Gemini separated the scoring criteria more distinctly: GameCube Wind Waker ranked #3 in atmosphere and #19 in pacing, while A Link to the Past, Majora's Mask, and Ocarina of Time occupied the overall top three.

For this benchmark, Gemini's rubric reproduced most of its pairwise ranking at linear cost, with stable, high-resolution probability outputs and greater separation between scoring axes. Jev's 0.01 quantization permanently discarded probability resolution, and its rubric dimensions remained substantially more correlated.

…uation-dashboard.mail-22c.workers.dev

Kkkk essa mulher é foda. Nunca aposte contra ela.