MARCELOCanonical item
The shared identity of a title: platform, region, edition and other attributes that define what the product actually is.
01 · Flagship case study · Field tested
Retro Game Vision & Resale System
A production-minded pipeline for turning Japanese second-hand store photos into traceable buying decisions. The difficult part is not recognizing a game. It is deciding whether this exact physical copy is worth buying without silently mixing incompatible evidence.
SYSTEM / FIELD STATEVALIDATED
Canonical items750+
Physical observations1,000+
Market observations1,800+
Integrity modelSQLite
PUBLIC DEMO RUNNABLE
The problem
A price tag is not a valuation.
I started with a practical question while looking through retro-game stores in Japan: which copies are actually worth buying for resale?
A title can look profitable until the details change. The disc may be missing. The copy may be a budget reissue. The market result may be for a different platform. A nearby price label may belong to another game. A high retail asking price may be mistaken for a sold value.
The project grew from photo reading into a system for preserving evidence, resolving identity, filtering incompatible market data and producing decisions that can explain themselves.
Architecture
Evidence first. Decision last.
The pipeline deliberately keeps recognition, physical-copy state, market evidence and economics separate. Each stage can refuse to continue when the evidence is not good enough.
01 / INPUTStore photoimmutable source evidence
→
02 / EVIDENCEHash + regiondedupe and spatial lineage
→
03 / VISIONOCR + candidatestitles, labels, positions
→
04 / IDENTITYCanonical itemplatform, region, edition
→
05 / MARKETCompatible evidencesold, ask, buyback, guide
→
06 / MODELEconomicsfees, ROI, max buy
→
07 / OUTPUTDecisionBUY / MAYBE / RESEARCH / SKIP
RULE
No downstream stage is allowed to manufacture certainty that the evidence layer does not have.
Data model
The separation that makes the system trustworthy.
Physical observation
One exact copy at one store, on one visit, at one visible price, with its own condition and completeness state.
Market observation
One sold result, active ask, retail ask, buyback quote or guide observation. Evidence types stay semantically distinct.
Evidence asset
The source image or document supporting a claim, linked through hashes and image regions instead of being discarded after OCR.
Hard rules
The system is designed to say “I don’t know.”
01Different platform?
Reject the comp. A Saturn market result cannot price a PlayStation copy.
BLOCK02Different edition?
Do not silently borrow standard-edition pricing for a budget or collector release.
BLOCK03Different completeness?
Disc-only and complete-copy evidence remain separate unless an explicit rule allows otherwise.
BLOCK04Obscured price?
Store NULL. Research time is not substituted for an unreadable source value.
UNKNOWN05Weak market evidence?
Return RESEARCH rather than forcing a confident-looking purchase decision.
DEFERFailure → redesign
The useful mistakes changed the architecture.
Whole-shelf OCR attached the wrong price to the wrong game.
Reading a dense shelf as a flat block of text lost the spatial relationship between titles and nearby labels.
REDESIGN
Keep the original image immutable, derive crops and grid regions, and preserve explicit spatial lineage back to the source.
Headline market value looked useful until copy state disagreed.
A complete-copy comp can turn a damaged or incomplete physical copy into a fake bargain.
REDESIGN
Physical-copy state became a separate entity, with compatibility gates before any market evidence reaches economics.
Repeat imports risked duplicating evidence and changing history.
A sourcing tool that cannot safely re-run becomes difficult to trust during iterative research.
REDESIGN
Hash-based evidence identity, idempotent writes, integrity checks and checkpoint validation became part of the workflow.
Field validation
Built against messy store data, not a clean demo CSV.
A recent validation batch illustrates how the system behaves when the evidence is imperfect.
uploaded photos10
unique evidence assets9
distinct physical copies49
directly readable prices48
obscured prices guessed0
compatible market observations added55
resulting screen1 BUY · 7 MAYBE · 41 RESEARCH
FK CHECK0INTEGRITYOKREPEAT IMPORTNO-OP
Engineering surface
What the project demonstrates.
Python
Typed domain models, modular decision logic, CLI workflows, hashing, validation and testable business rules.
SQLite / data modeling
Normalized entities, foreign keys, append-only evidence, idempotent writes and integrity checks.
Computer vision workflow
Image preservation, OCR candidate extraction, crop/grid derivation and evidence-region tracking.
Data engineering
Deduplication, canonical resolution, provenance, evidence typing, compatibility and freshness handling.
Decision systems
Deterministic fees, profit, ROI, confidence gates, maximum-buy thresholds and explicit failure states.
Reliability
Regression tests, checkpoint validation, repeat-import checks and CI-safe publication boundaries.
Public / private boundary
Inspectable without publishing the operational sourcing system.
The public repository is deliberately runnable. A reviewer can inspect the architecture, execute the synthetic demo, exercise the SQLite layer and run the tests. Real store photos, production market datasets, private adapters and operational thresholds remain private.
PUBLIC
Typed Python models
PRIVATE
Production database and sourcing history
PUBLIC
Synthetic end-to-end demo
PRIVATE
Real store photographs and market evidence
PUBLIC
Compatibility tests + CI
PRIVATE
Operational thresholds and source adapters
Repository
The public implementation is built to be read, run and challenged.
Architecture notes, data-model documentation, ADRs, regression tests and the synthetic demo are all available in the repository.