# AI-Prototype

Prototype workspace for Project '44 / Frame 44 descendant report pipeline.
First end-to-end test: producing a Tier 3 report for a hypothetical
descendant of a Calgary Highlanders soldier, using the regimental war
diary as the primary source.

## Conventions

- **Test customer name:** Bill Bloggins. Used as a placeholder across all
  prototype runs so test data is immediately distinguishable from real
  customer data. Never use a real name in prototype work.
- **File naming:** `<unit>_<yyyy-mm>.<ext>` where applicable.
- **Isolation:** This sandbox works only with copies of source data.
  Never connect this prototype to the master Postgres database.

## Folder structure

- `inputs/` — copies of source data (OCR .docx, original PDFs, position
  data). Read-only in spirit; never edited here.
- `prompts/` — prompts used to drive the AI runs, versioned (v1, v2…).
- `outputs/` — raw AI output from each run, saved verbatim.
- `reports/` — polished, human-reviewed deliverables built from outputs.
- `notes/` — project notes: unit selection, evaluation, follower
  feedback, decisions.

## Current test

- Unit: Calgary Highlanders (1600221)
- Window: mid-Sep 1944 through early Nov 1944 (the Scheldt)
- Tier: 3 (regiment known, dates fuzzy, one event known)
- See `notes/unit-selection.md` for the full scenario.

## Infrastructure

- https://deepinfra.com/dash used for running OCR via olmocr