Play the problem. Inspect the result.
Project arcade
Six small games.
Each built around a real engineering decision.
AF INSIDE THE WORK
6 PLAYABLE
← → Choose · Enter Play · Tap a cabinet to play · 03 / 06Enter cabinet ↗
01 / BAKIANO
Fictional market puzzle
Basket Run
Move groceries into a limited basket. Solve four markets with pack sizes, stock, delivery fees and dated prices.
What it demonstratesComparable prices depend on units, dates and traceable conversion.
Inside Bakiano →SHELF Drag a pack or use Add
YOUR BASKET
Goods Delivery Total
Price trail / receipt
Need a hint?
Four fictional markets. Pack sizes, stock, shared delivery fees and dated currency observations determine the answer. This puzzle illustrates Bakiano’s data checks; it does not query its private pipeline. Visit Bakiano ↗
02 / GCP INGESTION LAB
Executable pipeline simulation
Packet Factory
Build a working conveyor line. Place machines, branch duplicate packets and keep every delivery safe through five shifts.
What it demonstratesReceipt placement and observation time decide whether retries are safe.
Inside GCP Ingestion Lab →SHIFT 1 / 5
First shift
WAREHOUSE TARGETREADS 0 / 1MERGES 0 / 1PARTS 4 / 7
Conveyor: move a packet in the arrow direction. Choose a tool, then a floor tile.
INCOMINGGate: ↑ fresh · ↱ receipt found
Choose a floor tile to build.R rotate · Del erase
Finish the conveyor, then release a packet.
WAREHOUSE / milkEmptyNo completion receipts
OUT / RETURNS / RETRIES0 / 0 / 00 conveyor steps
Need a building hint?
Execution trace (0 steps)
An original conveyor puzzle based on the repo’s receipt, commit, and observation-time rules ↗. Fictional shipments; counters belong to this deterministic browser simulation. No GCP calls. Inspired by the build-and-run factories of Infinifactory ↗ and the programming puzzles of Human Resource Machine ↗.
03 / CATALOG MATCH LAB
Runs the repository’s matcher
Catalog Customs
Find the verified match in increasingly crowded product lineups. Sometimes none qualifies. Three mistakes close the desk.
What it demonstratesShared words do not establish product identity. Missing evidence deserves a hold.
Inside Catalog Match Lab →Catalog Match · inspection campaign
8 checkpoints / 35 declarationsCatalog Customs
Start at the stamp desk. Graduate to crowded shelves of almost-identical products. Find the one that actually fits — or prove that none does. Three citations close your shift.
Enable JavaScript to open the inspection desk.
Select one declaration. You can change your selection until you commit.
Inspect every declaration
The pinned repository matcher computed these checks. Open a declaration to see the attributes behind its verdict.
End-of-shift manifest
35 fictional declarations; real Catalog Match rules, with no cloud calls or model decisions. Replays shuffle shelf positions. Desk-inspection mechanic inspired by Papers, Please ↗. Original interface and artwork.
Free inspection: try your own product names
A / SHARED WORDS
A deliberately simple reference rule.
B / IDENTITY CHECKS
| Check | Requested | Candidate | Result |
|---|
Fictional names. These offline rules support a small vocabulary; unfamiliar details stay uncertain. This is not a trained model or Bakiano’s production matcher. Inspect the exact code ↗
04 / BIGQUERY QUERY GUARD
Executable scan simulation
Partition Sweep
Select the data blocks that produce the requested answer. Deliver every required row without exceeding the scan budget.
What it demonstratesProjection and partitions reduce reads. A small output alone does not.
Inside BigQuery Query Guard →PARTITION SWEEP01 / 03
Cut the baggage
Switch dates and fields off to shape your read. Then run the plan.
Planned read
Budget
Rows returned
| Dates ↓ Fields → | ||||
|---|---|---|---|---|
Will read◆ Required blockNumbers = scan units
Keep the answer. Cut the read.
Match the requested rows and fields within the budget.
Inspect the plan and returned rows
Dates are partition metadata. Product is also read for the fixed ordering, even if you omit it from the output. LIMIT changes returned rows after scanning. This toy ignores compression, clustering, metadata shortcuts, caching and billing minimums.
What this demonstrates in the repository
The Query Guard checks an estimate before execution and passes a byte cap to BigQuery. This puzzle uses fictional blocks to make query shape visible; it does not execute the guard or BigQuery.
In the repository’s recorded price-card example, selecting three fields instead of all fields changed the estimated read from 18.71 MB to 6.30 MB. Lowering LIMIT from 20 to five left the estimate at 18.71 MB. These are recorded observations, not this game’s scores.
Inspect the recorded evidence ↗An original miniature of the “same output, fewer resources” puzzle loop in Opus Magnum. No copied art or levels. Fictional scan units are not MB, execution time or billed cost.
05 / IMAGE TO PIXELS
Pictures made by the repository’s renderer
Pixel Shuffle
Put the picture back together. Tap two tiles to swap; later pictures add a few sideways pieces. Aim for fewer moves.
What it demonstratesThe repository renderer creates every pixel picture. The puzzle checks the image itself, accepting visually identical pieces.
Inside Image to Pixels →PIXEL SHUFFLE
Moves 0Aim for In place
Made with Image to Pixels
Free inspection: compare rendering methods
Both outputs obey the same color budget. Dithering trades a patterned texture for shading; it is not automatically better for every image.
Select this project to render the comparison.
Runs a pinned copy of the real renderer in a browser worker. Inspect the code ↗ · Open Image to Pixels ↗
06 / VESTI
Synthetic attribute puzzle
Complete the rack
Choose a third garment so each attribute is either all the same or all different across the rack.
What it demonstratesSeparate attributes let you reason about an item instead of its whole scene.
Inside Vesti →Vesti · attribute puzzle
3 racks / 2 tries eachComplete the rack
Choose the third garment. For each feature — shape, color and pattern — the three cards must be all the same or all different. Two matching and one different fails.
Enable JavaScript to complete the rack.
?Your choice
Choose one from the rail
2 tries left
Rail complete
Synthetic attribute puzzle — not segmentation or AI retrieval. It makes multiple constraints inspectable; Vesti's image pipeline remains in development. Feature-rule mechanic inspired by SET ↗. Original garment drawings.
PLAY → INSPECT → READ THE CODERead all projects →