Mazda's Training Process

Mazda has been trained on the company's receipt / document intake operating procedure. The full 8-step pipeline was validated end-to-end on a real Walmart receipt, the supporting tools were built or extended, and the procedure is now authored into Mazda's memory (memfs file system/receipt_intake_procedure.md + attached block) and confirmed live in her context.

What has been accomplished

DoneThe operating procedure, validated live

Every step below was run for real on a scanned Walmart receipt (an HP OfficeJet 8122e printer + HP 923 ink), not just designed on paper:

Scan — dashboard ROL Finance → Scanners or tools/run_scan_image.sh (WIA 300dpi → scan.jpg).
Show for review — tools/open_scan_jpg.sh / dashboard "Show Image" so EG confirms the correct document.
Cheap classify + route — classify_scan.py (Gemini Flash) returns receipt / bank_statement / tax_document / other.
Parse into an expense record — parse_and_categorize.py, now extracting merchant_city / merchant_state as separate fields.
Investigate cryptic items only — investigate_item.py uses Gemini google_search grounding to identify SKU-like descriptions (e.g. 4KOT3LN → HP 923 ink) with sources.
Resolve vendor_key — via VendorCategoryStore; No Vendor Key = failure (red-tab in report.html, exit).
Categorize — with the high-confidence (≥0.85) item-override rule (printer+ink → Office>Supplies, overriding Walmart's "Personal" default).
Store — only after vendor_key resolves and zero failures.

DoneTooling built or extended

Artifact	What it does	Status
`tools/classify_scan.py`	Cheap Gemini-Flash document-type classifier (route before parsing)	New
`receipt_parsing_tools/investigate_item.py`	Gemini Google-Search grounding to resolve cryptic/OCR-garbled line items + category hint	New
`receipt_engine.py` + models	Added `merchant_city`/`merchant_state` to the receipt schema & parsing prompt	Extended
Dashboard Scanners "Start Scan"	Greys out while the scanner needs a power-cycle ("Restart the Scanner Please")	Fixed

DoneEncoded into Mazda's memory

Authored the proper file-first way and verified live:

memfs (source of truth): committed + pushed system/receipt_intake_procedure.md to Mazda's state.git.
Attached block: system/receipt_intake_procedure (~2,900 chars) attached to Mazda.
Confirmed rendering: her compiled /context now contains the procedure, the "No Vendor Key" rule, and classify_scan.

DoneInfrastructure enabler — the "ghost container"

memfs authoring had been blocked by a recurring Docker "ghost": two Docker engines (native dockerd + Docker Desktop) ran in parallel, and Desktop hijacked the docker socket so the live fleet was invisible to the CLI and letta-memfs lost its network. This was diagnosed and resolved to a single native engine with all services healthy (letta, memfs git, bridge, logger php+mysql, executor); the live DB was backed up first (~/ghost_migration/). See memory ghost-container-root-cause-docker-desktop-2026-06-24.

One open infra item: disable Docker Desktop's WSL integration + autostart on the Windows side so the two-engine split can't return on next login. Quick post-reboot check: docker info --format '{{.Name}}' should read DESKTOP-SHDBATI, not docker-desktop.

Next suggested training steps

1Live dry-run with Mazda driving. Scan a fresh document and have Mazda execute the procedure from memory (not us running the scripts), to confirm she follows steps 1–8 in order and honors the No-Vendor-Key and override rules.
2Wire the Scanners page to the pipeline. Scan finish auto-fires POST /api/process-document: the deterministic facade (classify → parse) renders inline within seconds; Mazda is dispatched fire-and-forget for investigate → categorize → store. No polling. GoF: DocumentPipelineController (Command) + DomDocumentPipelineView (Strategy), both injected so unit-testable. 194 JS + 101 Python tests pass. Done 2026-06-25.
3Build the per-receipt failure surface. The "No Vendor Key" red-tab report.html for a single scanned document (step 6 failure path) is specified but not yet built.
4Route the other document types. Teach the bank_statement and tax_document branches from the classifier (today only receipt is fully wired).
5Duplicate detection & receipt↔statement linking. Fold the ROL Finance dedup/link rules into the intake flow so a scanned receipt links to an existing statement row instead of creating a duplicate expense.
6Investigation learning loop. When investigate_item.py resolves a cryptic SKU, persist the mapping (vendor-key learning log) so repeat items skip the web-search cost.
7Update the Agent Card. Reference this intake procedure in Mazda's system_message.xml so the dashboard Agent Card reflects her current role.
8Make the single-engine fix permanent (the Windows-side Docker Desktop disable above) so memfs authoring stays reliable across reboots.

Generated for the ROL Finance project · Mazda receipt-intake training. Companion memory: ghost-container-root-cause-docker-desktop-2026-06-24, procedure source of truth: memfs system/receipt_intake_procedure.md.