Conformance
We look for functional bugs in RTL before it is taped out, by checking the design against its specification.
The engineering is done by AI coding agents (Claude and Codex), directed by ChipLab's founder, Nathan Chandrasekar. One agent finds and documents each bug. A different agent reproduces it independently before it reaches you.
What each bug report contains
- A minimal reproducer that runs on your RTL.
- A trace against a golden model or the specification, showing where the design diverges.
- The root cause in your RTL.
- A check against your known-issues list, so you are not sent what you already know.
Results are private to you. We do not publish client findings or file them anywhere.
How it works
Scope
You send the RTL, the spec or design notes, and your known-issues list. We confirm the design is eligible.
Test
Agents write tests and compare your design with a golden model or the spec, using open-source simulators (Verilator and Icarus Verilog).
Independent reproduction
A second agent reproduces each candidate bug from scratch. A bug it cannot reproduce is not reported.
Report
You receive each verified bug privately.
Delete
At the end of the engagement we delete your files.
Pilots
We are opening a small number of pre-tapeout pilots for professional teams, on published or EAR99 designs.
Terms
- To start, we take only published designs or EAR99 designs. We screen every client against the US Consolidated Screening List.
- RTL only. We do not take PDK files or NDA'd libraries.
- Your RTL is processed by the commercial AI models our agents run on, on plans where your data is not used for model training. We name each provider in the engagement letter.
- Your files are deleted at the end of the engagement.
- This is not sign-off and not a warranty that your RTL is free of bugs.
Open-PDK shuttle deadlines
Send RTL early enough to fix what we find before your deadline. Each platform publishes its own dates:
- Tiny Tapeout: shared shuttles on open PDKs
- wafer.space: GF180 runs
Public record
The same method, run in public on open-source cores. The Conformance Matrix runs the official RISC-V architecture tests across 11 open cores. The Findings page lists every public report.
| Project | Public issue |
|---|---|
| Hazard3RISC-V core in Raspberry Pi's RP2350 | #52 amominu.w and amomaxu.w store the signed min/max when Zbb is disabled. |
| Hazard3 | #51 A host write to data0 while an abstract command is busy changes data0; the spec says it must not. |
| Hazard3 | #50 sbcs.sbaccess resets to 8-bit instead of 32-bit when System Bus Access is enabled. |
| NEORV32 | #1663 A DMA byte-to-word transfer replicates the byte instead of zero-extending it. |
| NEORV32 | #1662 SLINK interrupt-enable bits sit at different positions in the RTL and the datasheet. |
| SERV | #173 Reading mie clears MTIE, which disables timer interrupts. |
| Microwatt | Under verification. |
Contact
Email [email protected]. Tell us what the design does, its rough size, your shuttle or tapeout date, and whether it is published or EAR99.