Why holotype exists¶
holotype was not designed in the abstract. It was built in response to a specific data-loss event and a specific publication requirement. Both are worth telling, because together they define exactly what the tool must guarantee.
The trigger¶
On the weekend of April 4–5, 2026, autonomous Claude Code sessions drove a 25-hour experiment on a thermoelectric measurement rig. The sessions produced the headline result of the project: ionic-gel equilibration takes 41 minutes, not the 60–120 seconds the lab had been using. That correction retroactively invalidated a chunk of the standard Seebeck-coefficient method for that material class. It is the kind of finding a paper is built around.
Weeks later, when it came time to cite the verbatim transcripts of those sessions in a planned Digital Discovery paper, they were gone. Claude Code's default cleanupPeriodDays: 30 had pruned them. What survived was Claude's own saved memory notes: an outline of what happened, not the record.
The realization¶
Memory notes and rsync backups together still don't deliver what a paper needs:
- Memory is a summary. It says what was concluded, not what was said and done, byte for byte.
- rsync preserves bytes but has no manifest, no hash chain, no semantic structure, and no way for a reviewer to prove the file they're reading is the file that was written.
For a forensic claim in peer review ("we ran this LLM, with these tools, on this code, and it produced these decisions"), an outline is not enough. You need a citable, hash-verifiable, immutable record. That record did not exist, and no existing tool produced it.
The publication requirement¶
Digital Discovery's Data Availability Statement requires LLM input/output logs, model identifiers, and generation dates. holotype is the substrate that makes such a statement satisfiable: not as a promise, but as a verifiable artifact a reviewer can check.
Why not an existing tool?¶
We looked. Each was built for a different job:
| Tool | What it does | Why it's the wrong fit |
|---|---|---|
| claude-vault (Rust, SQLite + FTS5) | Fast full-text search over sessions | Explicitly strips tool calls and system tags as "noise" before indexing. The right call for search, the wrong call for provenance. The tool calls are the experimental record. |
| claude-conversation-extractor (Python) | Export sessions, with a --detailed flag |
Closer, but search-first by design, not provenance-first. |
| Manual rsync | Copy the raw files | Preserves bytes, but no manifest, no hash chain, no semantic structure, no reviewer-facing verification. |
The single distinguishing line: the difference between a search index and a forensic record is whether you keep the tool calls. holotype fills the niche the search tools deliberately leave empty.
The test every future change must pass¶
Does this change strengthen or weaken the chain from a paper's citation back to a hash-verifiable transcript?
Backups, search, and convenience come second. The April 2026 incident is the load-bearing memory: any design move that re-opens that hole is rejected.