Experiments that do not need permission to be useful.
Benchmarks, prototypes, tiny utilities, comparisons and things that may never become a polished product.
Graph-aware coding context
Comparing agent answers with and without a generated code graph to see whether structure actually improves repository understanding.
JEV routing simulation
Comparing routed model tiers with direct fixed-model baselines, including explicit token/runtime penalties when a lightweight model is mismatched to a harder task.
Agent harness benchmarks
Real intentionally broken projects, issue discovery, repair quality, token usage and generated tests.
Tiny developer tools
Small utilities that solve one annoyance well enough to keep around, even if they never deserve a full product page.
AI product trials
Testing claims made by AI products against actual workflows instead of repeating feature lists.
Useful dead ends
Ideas that looked promising and then failed the “would I actually use this?” test.
Graph-aware coding context
The experiment compares the same repository-understanding task with and without graph-derived context. The useful question is not whether a graph looks impressive, but whether it helps an agent find the right files, boundaries and request path with less wandering.
JEV routing simulation
JEVArena makes routing decisions visible instead of hiding them behind a single score. The simulator separates decision time from model execution, compares routed tiers against direct models, and applies transparent capability-mismatch multipliers when a selected lightweight model is asked to carry a substantially harder workload.
Agent harness benchmarks
Benchmarks use intentionally broken projects so an agent has to discover issues, reason about them, make repairs and prove the result with tests. Token usage and failure modes matter as much as a final “success”.
Tiny developer tools
Some ideas are useful long before they deserve a roadmap. This is where small utilities, prototypes and one-purpose helpers get to exist without pretending to be products.
AI product trials
The test starts from a task, not a feature checklist. If a product claims to build, reason, automate or save time, I want to see what survives contact with an actual workflow.
Useful dead ends
Not every prototype deserves continuation. The useful part is recording why it stopped: weak reliability, bad ergonomics, duplicated value, cost, or simply failing the “would I keep using this?” test.