Every layout our extractor had ever read was a layout our tools had written. That is a comfortable arrangement with a known failure mode: an extractor that only ever reads its own author's output can be wrong about the world and right about itself forever. We needed a chip we didn't make, with an answer key we didn't write.
On August 5th, Jane Street published one. Their ASIC puzzle is a placed-and-routed sky130 design shipped as bare GDS (no netlist, no schematic, no documentation beyond example inputs and outputs) with a standing question: can you work out what it does? Extracted, it comes to 738 logic and sequential instances, 92 of them flip-flops. For us it was the missing test case: built by someone else's flow, named by someone else's conventions, graded by ground truth we couldn't have baked in.
Reading the chip
The work ran through klayout-tools end to end, driven by agents: a cell inventory and pin abstracts resolved straight from the GDS stream, then cell-level connectivity extraction to a gate-level Verilog netlist, then an equivalence check up to renaming wired into CI as a regression, then a VCD replay harness that simulated the recovered netlist against the published example traces. When the replay matched, we trusted the netlist enough to interrogate it: a bounded model checker solved across the 92-bit state for the input that drives the success condition, and simulating the recovered output generator with that input produced the answer we submitted on August 10th. The commit log keeps the clock: first commit the evening of the 7th, about ten working hours total, seven of them in the single push that ended at the submission form.
The toolkit-first rule held throughout: when the extractor met something it couldn't read, the fix went upstream rather than into a local patch. Five klayout-tools issues came out of this one chip (#619, #620, #673, #674, #675) — which is the point. A puzzle you solve once is entertainment; the extraction gaps it exposes are permanent upgrades to the reader.
The part where they showed off
Jane Street put jokes in the mask layers. Feed the recovered logic all zeros and the output register spells EMPTY SKY; all ones gets you BIG BANG. Every other failing word we fed it, 169 across eight families, came back TRY AGAIN, and a later exhaustive enumeration proved the full message map: exactly five messages, the fifth a diagnostic reserved for the one subtle way to fail. The commit that proved it landed minutes after we hit submit, which is why the submission record calls it a post-submission addendum. Below the die, on a layer no fab would print into anything electrical, 36 placements of what the inventory had filed as bookkeeping cells spell PER ARENAM AD ASTRA in Morse code: through sand, to the stars. There is an eighteen-micron three-ring maze on met2 connected to nothing at all. Someone at a trading firm drew a record groove into a chip on the correct assumption that eventually somebody would look. We looked. The full commands and literal outputs are in the repo's evidence directory, because a claim about a hidden Morse message is exactly the kind of claim that should come with a reproducible script.
Now public
Submissions closed on September 4th, which was the only reason the work stayed private: Jane Street asked solvers not to post spoilers before then, and a private repo is how you keep that promise with agents pushing to it daily. The whole trail is now open at 2AMLogic/sky130-asic-puzzle: the extraction pipeline, the solve, the evidence scripts, the CI that kept the netlist honest, and the writeup we submitted.
The extractor read a chip its authors never saw and was right about it in about ten working hours. The chip returned the favor: its ports land on met3 and its power on met4 and met5, layers none of our own layouts had ever made matter, and the extraction ceiling they exposed went upstream as #619; the fix was live before the final push started. Someone else's next layout will find the next one.