Explainers

terminology library

Capability Ladders โ€” Partial Credit for Exploitation
draft ยท evals
Contamination and Decontamination
draft ยท data
Dataset Lineage
draft ยท data
Differential Oracles and Win Conditions
draft ยท training
Domain CPT, LoRA, and Forgetting
draft ยท training
MTP speculative decoding โ€” the detachable draft head
draft ยท infrastructure
Rejection Sampling and STaR โ€” the cheapest gain in post-training
draft ยท training
SFT vs DPO vs GRPO vs RLVR โ€” the four post-training regimes
draft ยท training
Small-model economics โ€” why a weekly-refreshed local model beats one giant
draft ยท strategy
The Purple Loop โ€” When the Attacker Trains the Defender
draft ยท training
Trajectories vs Writeups โ€” game tape vs victory speech
draft ยท training
ฯ„=0.9 structured compaction โ€” how long rollouts survive their own logs
draft ยท infrastructure