On July 28th we committed a scoring rubric to git with no scores in it: what to measure, and why, before a single data point existed to plug in. That commit exists specifically so nobody could point at the eventual ranking and ask whether the rubric was written to fit it.
Why bother
The easy way to pick ten chip design blocks is to argue backward from the one you already want to build. We'd been circling gf180mcu and sky130, two of the most active open PDKs, and it would have been simple to write a rationale that landed on our favorites. Instead we wrote the categories first: demand, feasibility, and how much friction each block would put on klayout-tools per unit of effort. Then we commissioned a round of deep research into the actual candidates. The git history is the receipt: the rubric commit and the scored commit, methodology first, thirteen minutes apart. Those commits live in an internal repository, though, so you are taking our word on the timestamps; the receipts you can check yourself are the block repos below.
What the research found
Demand turned out to be checkable, not guessable. Vidatronic licensed almost exactly our analog shortlist at the 180nm class: LDO, bandgap, temperature sensor, clocking. That lineup now survives only on the Internet Archive, the vendor's own domain having since been pointed somewhere else entirely. That's one vendor's product line, not a market study, but it's real evidence that these categories get licensed rather than built in-house.
The sharper signal was absence. Silicon Creations, the category leader in PLLs with over a thousand production licenses across nodes from 3nm down to 180nm, does not serve gf180mcu at all. Nobody had bothered to fill that gap.
The other thing worth checking was whether "open PDK" meant "already solved." It didn't. Free sky130 analog IP exists too, bandgaps and SAR ADCs and LDOs among them, and it's simulation-qualified, not measured on real silicon, with one narrow exception we found and discounted the relevant score for. Free and simulated is not the same claim as free and proven, and a design that someone still has to tape out to trust is a real gap, not a sales pitch.
The ladder
The ten blocks got ordered as a difficulty ladder, not a priority list: bandgap, then LDO, then temperature sensor and power-on-reset, then oscillator and PLL, then a SAR ADC, then a TRNG, gf180mcu first at each step with the sky130 port trailing. Ten counts blocks on a specific PDK, not circuit types: bandgap, LDO, PLL, and the SAR ADC each take a gf180mcu and a sky130 slot inside the first ten, which is eight; the temperature sensor and power-on-reset share the ninth; the TRNG is the tenth. The oscillator sits just outside the cut as first alternate. Same instinct as kicad-tools' first board being a single LED rather than the motor controller. Bandgap won the top slot because it's the smallest circuit that still has real analog teeth, a reference voltage that has to hold across corners, and because almost every other block in the kit needs one anyway.
Week one, block one
gf180-bandgap went from an empty repo to a schematic in xschem, an error amplifier with an offset and mismatch budget, a characterization of the vertical-PNP pair mismatch that budget depends on, and a testbench suite indexed to the spec table, one verdict per covered line. Corner analysis then found that the error amp's offset budget and PSRR were short of spec: a topology problem, not a tuning problem. Not a clean pass, but a place where the rubric's "designable with current tools" assumption met an actual circuit and had to argue its case.
That argument is settled now, and it went the way the corner analysis said it would. The shortfall issue closed on August 1st when the five-transistor error amp was replaced with a telescopic-cascode OTA, a topology change, not a knob turn. Layout landed the same day, DRC came back clean and LVS matched by August 3rd, and the verification suite re-ran against the extracted layout; the wave-one note carries the dated trail. The block isn't finished: the drawn PNP array currently disagrees with the schematic about its effective ratio, tracked in #87.
The repo went public on July 31st, after an audit pass for anything that shouldn't leave the building, alongside its sky130 sibling. The next five cleared the same gate by August 4th, which put seven of the ten in public view. The remaining three, the sky130 ports of the LDO, the PLL, and the SAR ADC, sit further up the ladder and stay private until their own audits clear.
What we don't know yet
The rubric assumed feasibility and friction would behave consistently across a kit of related analog blocks. LDO was the fast check on that, close enough in circuit family to bandgap to tell us whether "designable now" was a real prediction or got lucky once. The answer came back in a shape we hadn't thought to score: the agents designed gf180-ldo, measured it, and then found the loop-stability numbers were not measuring stability, a subtler failure than a circuit that doesn't work. Two rungs in, the feasibility calls are holding and the surprises come from a direction the rubric never asked about: not whether agents can design the circuit, but whether they know what they're measuring once they have. Whether that stays true through a SAR ADC, a TRNG, and the first real silicon is the part we can't check from here yet.