Innovation Prior

Every paper here is rewritten as a first-person discovery trace — the method re-derived from the pain points and prior art of its time, hitting walls and self-correcting until it lands on the idea and the code.

Finished papers are written theorem-first: the answer arrives polished, with the struggle edited out. Innovation Prior runs that backwards. For each paper we reconstruct the chain of thought of how the method was actually arrived at, starting from the state of the field before it existed.

Each method comes with three artifacts:

• Context — the pre-method ground: the research question, the prior art, and exactly what hurt at the time.
• Reasoning — the long-form first-person trace of the derivation, walls and self-corrections included. This is the main event.
• Answer — the distilled summary plus reference code, the place the trace lands.

Pick a paper from the sidebar to begin — / focuses search.

atomic traces, generated by the paper-to-reasoning skill. Switch to Trajectories (0) for iterative research lines that climb an MLS-Bench task's baseline ladder — initial context → each baseline + my feedback → a stronger method. Agentic (0) replays those same research lines as a tool-using agent transcript (think → edit code → run_experiment → observe). Training data browses the full SFT corpus these traces are assembled into.