# Scheme scope + "should we OCR the whole corpus?" — evidence-based answer

**Date:** 2026-08-15 | Additional Azure DI spend for this analysis: **$0.22** (40 docs / 116 pages)

## Question
Is there benefit to OCR-ing and indexing the entire ~23,561-document 2025-2026 corpus, or do we
already have what's needed?

## Answer: NO whole-corpus OCR needed — we have what's required.
The **index metadata alone** (grantor/grantee/type/date, which we scraped for free) already isolates
the fraud fingerprint. Whole-corpus OCR (~$210, ~23k docs) would add negligible value for this case.
Instead a **tiny targeted OCR** conclusively closed the remaining gap:

### What we tested
1. **All "INVALID DEED" annotations** in the window → exactly **2** (#4583923 Ingalls, #4583924
   Schermer) — both already captured.
2. **Self-conveyance QCDs** (grantor = grantee): 935 — but almost all are legitimate estate-planning
   transfers to the owner's own **family trust**.
3. **Exact self-to-self, non-trust QCDs** (the true land-patent fingerprint): **42**. Removing the 2
   known INVALID DEED docs left **40**; we downloaded + OCR'd all 40.
4. **Scanned those 40 for the distinctive fraud markers** ("land patent", "patent account", "Bagnell
   v. Broderick", "successor in interest", "highest evidence of title", "Collins v. Bartlett", and
   the fraud notary **Hicks / #194569**).

### Result
- **Zero** of the 40 contain the distinctive land-patent language or the Hicks notary. (A few contain
  generic quitclaim boilerplate like "remise" or "Willamette Meridian" — present on ordinary deeds.)
- **Conclusion:** the only land-patent quitclaim frauds recorded in Island County in the 2025-2026
  window are the **three already identified** (Shroyer #4582708, Ingalls #4583923, Schermer #4583924),
  plus their supporting notices/declarations. No additional victims/properties from this specific
  scheme are hiding in the corpus.

## Why the index (not full OCR) is sufficient here
- The fraud is **self-indexed** under the perpetrators' own names and flagged by the County
  ("INVALID DEED"), so a **name/pattern search of the metadata index finds it** — no image text needed.
- Full-text OCR only helps for deeds that reference a property **without** naming the target in the
  index. That risk is real in general, but for THIS scheme the perpetrators recorded under their own
  names, so the index already captures them.

## When whole-corpus OCR *would* be worth it (not now)
- If you later suspect frauds recorded under **strawman names with mis-tied legals** (the classic
  "wild deed"), a bounded content index would help — the pipeline is built and ready
  (`scripts\ocr_di.py`, `push_to_search.py`, Azure index `island-county-deed`). Cost to do the full
  2025-2026 window later would be ~$210. Not justified by current evidence.

## Running DI cost
Cumulative additional Document Intelligence spend across the whole engagement: **~$0.22**
(see `08_self-index\azure\page_tally.json`) — vs. the $500 cap. The originating tip + free index made
the expensive bulk OCR unnecessary.