CASE 1 M,N,K game research used a Rust and CUDA AlphaZero system to test whether a smaller-board checkpoint could improve learning on a larger board. The checkpoint loaded successfully, but the campaign never established a benefit over matched scratch training, and several promising settings produced a worse and more expensive combined result.
Autoresearch case studies
Field reports on using autoresearch in applied projects, including what worked, what failed, and how the research infrastructure evolved.
10 cases 2026-02-04 - 2026-08-14
CASE 2 Tactics.md used autoresearch to develop and tune its hand-written Greedy and Heuristic agents across more than 24 million simulated games, then test learned action search. One learned model reached 86.7% held-out accuracy but scored only 10% against the simpler Greedy agent because useful actions were pruned before deeper evaluation.
CASE 3 Tactics.md used Codex to test 102 interventions on the CPU pipeline generating self-play training games, accelerating turn search and quantized neural evaluation while uncovering eight correctness defects. Narrow component benchmarks repeatedly misled; only full-turn replay and end-to-end self-play could distinguish real speedups from changed game behavior.
CASE 4 A Gradient Bang headless client completed live trading, combat, fleet, and exploration workflows and reached contemporaneous visible ranks of 29 in exploration and 27 in trading. Parallel probe agents then began consuming one another's completion events.
CASE 5 Tactics.md used autoresearch to train neural models that evaluated positions inside its game-tree search, testing architecture size, training targets, quantization, and self-play across 435 result bundles. Under the same 10 millisecond move budget, a model that evaluated each position 73 times more slowly still won because its judgments were better, while the model with the best validation loss played worst among matched candidates.
CASE 6 Options market-making research used GPT-5.5 for implementation, Gemini 3.5 Flash for candidate generation and reflection, and Fable 5 for judgment and stopping decisions. An independent audit found gaps between the research record and implemented system. One candidate failed a causal correction and its replacement survived bounded validation, but the historical return disappeared in a short fresh window with stronger quote competition.
CASE 7 FlySim trained flight-control policies in a public JAX simulator, then a private overlay tested a frozen policy in a live external flight application. When a half-second handoff clipped 14% of controls, replaying the exact safety-filtered inputs showed that the simulator underrepresented the live aircraft's roll and yaw response. Retraining fixed the clipping without weakening the live safety limits.
CASE 8 A party-management strategy game used six model families as independent playtesters who managed parties, edited character tactics, and revised strategies against a deterministic battle simulator. Their disagreements prevented premature game-design changes, exposed tool-assisted play that violated the rules, and recorded which mechanics the models described as clear, confusing, or engaging.
CASE 9 FreeTranscribe.org doubled transcription throughput while sharing one RTX 3090 between live production and autoresearch. Error counts stayed unchanged across 295 broad cases and fell from 105 to 103 on a reviewed long-form set.
CASE 10 Quantitative-trading autoresearch tested 1,807 unique policy-operation variants, then ran 93 mechanism-first research cycles across two long-running campaigns and completed 73 economic evaluations while sealed future data stayed untouched. A staged universe analysis screened 858 markets, showed that blindly expanding a 20-market policy failed, and grew the forward data regime from 20 to 42 markets with richer hash-bound evidence.