research vaults · Better3

What it is

A research group reads hundreds of papers and keeps them in a folder. A vault turns that folder into something that can be asked. Each paper gets a one-page summary with structured frontmatter; from those summaries a script regenerates everything else: a master index, thematic topic files, a reverse citation index, a method index, a BibTeX file and a browsable dashboard. Nothing derived is edited by hand, so nothing drifts.

three vaults

fig. 01 — each vault is one subject, one folder, one protocol

How it works

Papers arrive as extracted text, one folder each. A first pass drafts a controlled taxonomy of eight to twelve themes from the corpus. Then each paper is summarised against that taxonomy: takeaway, methods, claims, what it is useful for, and which earlier papers it revises. The regenerator reads all summaries and rewrites the derived files.

vault build · defence systems

fig. 02 — the regenerator is the only thing that writes derived files

tree DefenseSystem_Vault/

fig. 03 — derived files are regenerated; summaries and the drafting log are written by hand

Two indexes, because research and drafting are different jobs

The subject index answers “what does the vault know about X”. The argument index answers “can I write this sentence, and what backs it”. A vault with only the first is a library; the second is what makes it usable while writing a paper or a grant. Both come from the same summaries, so the cost is one regenerator run.

retrieval protocol

fig. 04 — step four is what keeps an old summary from announcing a claim that a newer paper overturned

Asking a vault

A one-line command puts a question to the vault and prints the answer. It is read-only by design: the agent gets Read, Grep and Glob and nothing else, so a question arriving from a chat bot can never modify the vault.

ask-vault

fig. 05 — read-only tools; the answer cites the summaries it used

Design decisions

Summaries are the source; everything else is derived. One regenerator, no hand-edited indexes, no drift.
Supersession is explicit. Connections are written into newer papers only, so the reverse index is precomputed and consulted every time.
Read-only retrieval. The bridge that answers questions cannot write; the drafting log is appended by a person.

What is not done

Vaults take extracted text, one folder per paper; raw PDFs are converted first. The taxonomy is drafted once and revised by hand. Contested claims are recorded only when a paper in the corpus revises another, so a young vault has an empty claims file, and says so.