Overview Shipped
Four research programmes and a reference library sit behind the design decisions on this site. Each one below says what the question was, what was actually found, where the sources are — and whether it changed the product or honestly did not.
Why this section exists
Almost every non-obvious choice in this system — separating the confidence scorer from the finder, refusing to filter low-confidence findings, hunk-level map-reduce instead of truncation, a deterministic stop authority in Python — came from reading what other people had already learned, and occasionally from finding out that the thing everybody repeats is single-sourced to a tweet. Publishing the reading is the only way that claim is checkable.
And the ones that changed nothing
Two of the packs below produced no code. One of them was researched thoroughly and then explicitly dropped on YAGNI grounds; another was written after the design it supposedly informed already existed. Both are listed, and both say so. A research section that only lists the research that worked is a marketing section.
The programmes
| Programme | Question | Changed the product? |
|---|---|---|
| Code review | How do the best AI reviewers actually work? | Directly and heavily |
| Loop engineering | What is the prior art on propose → verify → refine? | No — the design predated it |
| Memory & learning | Is an MCP server to the backend the naive answer? | Design only; not built |
| Cost & dispatch | Why did the bill spiral, and what does a dispatch cost? | Directly — the tier model |
| Database library | Reference material for critiquing storage design | As advisory material, not as our schema |