Gemini 4 hackathon · 5 Sigma track

Agents that finish the hardware design.

routr puts teams of Gemini 4 agents on the two jobs that stall chip and board work: routing the last connections on a circuit board, and sizing an analog circuit until it passes at every corner. Every result here is scored by an independent checker against a published or standard baseline.

6/10hard PCBWorld boards routed clean. Best published: 4/10
4/10PCBRouteAgent, the published LLM routing agent
7/7op-amp sizing runs that pass all 15 PVT points
1/18Bayesian optimisation and random search, same budget
Part 1 · PCB routing

A circuit board with one unconnected wire is scrap.

PCBWorld's hard tier: 10 real boards. A board passes only when every connection is routed and KiCad's rule checker finds zero errors. Boards that pass, out of 10:

Freeroutingclassic autorouter, published
3/10
PCBRouteAgentLLM routing agent, published
4/10
routrrouter + repair + Gemini 4 agents
6/10
classic router alone (3)+ our scripted repair, no model (+2)+ Gemini 4 step agents (+1: Chickadee (0400))
Every hard-tier board and who fixed it
BoardResultFixed byOther arms
Chickadee (0400)cleanGemini step agents
Managed Agent: cleanrandom control: clean
temperature-sender (0423)cleanclassic router
uedaino (0448)cleanclassic router
data-manager (0473)not cleannot clean yet
agents: 4 → 3 open
kitspace f-91w (0496)not cleannot clean yet
negative (0520)cleanscripted repair
ozinverterkicad (0546)cleanscripted repair
Brushless ESC (0571)cleanclassic router
RGBMatrixPanel CPLD (0599)not cleannot clean yet
agents: 60 → 31 open
HybridChargeController (0626)not cleannot clean yet
agents: 1 → 1 open
How the crew works

The router does the bulk. Gemini agents finish the board, one checked step at a time.

Classic router

Freerouting

Routes about 97% of each board, then leaves a few connections open.

Lead

gemini-4-carbon

Looks at a picture of the board and writes one strategy note for the crew.

Four step agents

gemini-4-hydrogen

Careful, bold, layer-first and power-aware, each on its own copy of the board.

KiCad acts

router + rule check

Every step is one action: route, route with a via, rip up, roll back. KiCad does the copper.

Judge

PCBWorld evaluate_one

Scores every saved board. The best board is the one submitted.

Each step:observe fresh facts→decide one action→KiCad routes→rule check judges→keep the best or roll backup to 20 steps per agent

Control arm

A random policy with the same actions and the same step budget. No model.

Blocker screen

gemini-decision-1

Names the routed nets that wall off an open connection (whole-plan arm).

Managed Agent specialist

gemini-4-hydrogen, long-running

One agent per stuck board, with coarse tools on the same engine, for boards nothing else fixes.

Real runs

Watch the agents work a board

Every line below is the agent's own reasoning from the run log. Hover the chart for each step; filter the feed by agent.

hard-tier test board

Chickadee (0400)

Gemini agents: 1 open → 0 · clean pass in 5 stepsrandom control: 1 open → 0 · clean pass in 13 steps18 Gemini calls · 31,366 tokens
board before
before · classic router · 1 open
board after
after · careful · 0 open

Open connections after each step

a rip-up pushes it up first
024605101520stepclean
careful 0bold 0layer-first 0power-aware 0random control

What Gemini decided

straight from llm_calls.jsonl
Leadgemini-4-carbonstrategy note from the board picture

Move +BATT first: it has only 4 pads, crosses the open GNDPWR path twice on top copper (layer 0), and sits right at the DRC error (red X) near (211.5, 104.1). Reroute or nudge +BATT to clear the rule violation at that pad. Leave all other nets and the existing 30-pad GNDPWR network alone. Use a via bridge to bottom copper (layer 1), which is completely open around x = 205–212, y = 104: either drop GNDPWR onto bottom copper to duck under +BATT, or bridge +BATT onto bottom copper.

carefulstep 1route net 21 → 1 open

Net 2 (GNDPWR) has a single short open connection crossed by net 7 (+BATT) on layer 0. Let's try routing net 2 directly with shove mode first.

boldstep 1route_via net 21 → 1 open

Net 2 (GNDPWR) has a short open connection crossed by net 7 (+BATT) on layer 0. Try route_via on net 2 first to see if a via bridge can jump over +BATT.

layer-firststep 1route_via net 21 → 1 open

Net 2 (GNDPWR) is crossed by net 7 (+BATT) on layer 0; following the layer-first strategy, try route_via on net 2 first.

power-awarestep 1route_via net 21 → 1 open

Net 2 (GNDPWR) has a single short open connection crossed by net 7 (+BATT) on layer 0. Let's try route_via on net 2 to bridge over +BATT without ripping up any nets.

carefulstep 2route_via net 21 → 1 open

route on net 2 with mode 1 failed, so try route_via to bridge over the +BATT crossings on layer 0.

boldstep 2ripup net 71 → 4 open

route_via on GNDPWR (net 2) failed because +BATT (net 7) walls it off on layer 0. Rip up +BATT (only 4 pads) to clear space for GNDPWR.

layer-firststep 2ripup net 71 → 4 open

route_via on net 2 (GNDPWR) failed to close the connection. Net 7 (+BATT) crosses the path twice on layer 0 and only has 4 pads, so I will rip up net 7 to clear the way.

power-awarestep 2ripup net 71 → 4 open

route_via on GNDPWR failed because +BATT (net 7, 4 pads) blocks it on layer 0. Rip up +BATT so GNDPWR can be routed cleanly.

carefulstep 3ripup net 71 → 4 open

GNDPWR (net 2) is blocked by +BATT (net 7) on layer 0, and both route and route_via on net 2 failed. Rip up +BATT (net 7, only 4 pads) to clear the path for GNDPWR.

boldstep 3route_via net 24 → 3 open

Net 7 (+BATT) was ripped up and net 2 (GNDPWR) now has no crossings, so route net 2 using route_via.

layer-firststep 3route net 24 → 3 open

Net 7 (+BATT) is ripped up and net 2 (GNDPWR) now has no crossings, so route net 2 first.

power-awarestep 3route net 24 → 3 open

Net 2 (GNDPWR) now has no crossings on its open connection after ripping up +BATT, so we should route it first.

carefulstep 4route net 24 → 3 open

Now that +BATT (net 7) is ripped up, GNDPWR (net 2) has no crossings. Let's route GNDPWR on its current layer.

boldstep 4route_via net 73 → 0 open · clean

GNDPWR (net 2) is now closed; route +BATT (net 7) using via bridges to clear the crossing nets.

layer-firststep 4route_via net 73 → 0 open · clean

GNDPWR is now closed, and +BATT (net 7) has 3 open connections; try route_via on net 7 to bridge onto bottom copper.

power-awarestep 4route_via net 73 → 0 open · clean

GNDPWR (net 2) is now routed. Next, route +BATT (net 7) using route_via so it can bridge onto bottom copper to clear the crossing nets.

carefulstep 5route_via net 73 → 0 open · clean

GNDPWR (net 2) is now routed, leaving only +BATT (net 7) with 3 open connections. Let's use route_via on net 7 to bridge over the crossing nets.

hard-tier test board

data-manager (0473)

Gemini agents: 4 open → 3 openrandom control: 4 open → 3 open81 Gemini calls · 232,801 tokens
board before
before · classic router · 4 open
board after
after · careful · 3 open

Open connections after each step

a rip-up pushes it up first
0102005101520stepclean
careful 3bold 3layer-first 3power-aware 3random control

What Gemini decided

straight from llm_calls.jsonl
Leadgemini-4-carbonstrategy note from the board picture

Move first: Reroute 2-pad nets /U2_RX and /CAN_TX first; their bottom-layer traces are the only obstacles blocking clean bottom paths for /U0_RX and /U2_TX. Leave alone: Keep high-fanout power and control nets (GND, +3V3, +5V, /RESET, /RTS, /DTR) intact. Via bridges: Use a bottom-layer via bridge for Net-(R5-Pad2) to pass directly beneath top-layer GND and Net-(R4-Pad2). Once /U2_RX and /CAN_TX are cleared, bridge /U0_RX and /U2_TX on bottom copper under top-layer congestion. Route /SD_DAT0 on bottom copper under +3V3 and /CAN_RX, adding short top-layer via bridges over bottom GND.

carefulstep 1route net 584 → 4 open

Net 58 (Net-(R5-Pad2)) is the shortest open connection (9.23 mm) and only crossed by two nets; try routing it directly with shove mode first.

boldstep 1route_via net 584 → 3 open

Net 58 is the shortest open connection (9.23mm) with only two crossing nets; try route_via with shove mode first.

layer-firststep 1route_via net 584 → 3 open

Following layer-first style, try route_via on net 58 first as it is short and only crossed on layer 0.

power-awarestep 1route_via net 584 → 3 open

Net 58 is short (9.23 mm) and only crossed on layer 0, so route_via with shove should be able to bridge over it.

carefulstep 2route_via net 584 → 3 open

Net 58 failed to route on its current layer with shove mode, and its crossings are only on layer 0, so a via bridge to another layer should easily complete it.

boldstep 2route_via net 403 → 3 open

Try routing open net 40 (/U2_TX) using via bridges with shove mode before ripping up any blocking nets.

layer-firststep 2route_via net 403 → 3 open

Try route_via on open net 40 (/U2_TX) first, following the layer-first strategy.

power-awarestep 2route_via net 403 → 3 open

Net 40 (/U2_TX) is a short 2-pad signal net; try route_via with shove mode first to see if it can bridge across the obstacles without ripping anything up.

carefulstep 3route net 403 → 3 open

Net 40 (/U2_TX) is open; try routing it directly with shove mode first before attempting vias or ripups.

boldstep 3ripup net 39, 413 → 5 open

Rip up 2-pad nets /U2_RX (39) and /CAN_TX (41) to clear the bottom-layer obstacles blocking /U2_TX (40) and /U0_RX (42).

layer-firststep 3route_via net 423 → 3 open

Try route_via on /U0_RX (net 42) next to see if a via bridge can close it before ripping any obstacles.

power-awarestep 3ripup net 39, 413 → 5 open

Rip up 2-pad nets /U2_RX (39) and /CAN_TX (41) as advised to clear their bottom-layer obstacles blocking /U0_RX and /U2_TX.

carefulstep 4route_via net 403 → 3 open

Single-layer routing for net 40 (/U2_TX) failed, so try a via bridge to complete it across layers.

boldstep 4route_via net 405 → 4 open

Now that /U2_RX and /CAN_TX are ripped up, try routing /U2_TX (net 40) with a via bridge.

layer-firststep 4route_via net 613 → 3 open

Try route_via on the remaining open connection /SD_DAT0 (net 61) before ripping any blocking nets.

power-awarestep 4route_via net 395 → 5 open

Net 39 (/U2_RX) was just ripped up to clear bottom-layer obstacles; reroute it with route_via so it can use via bridges around top and bottom congestion.

carefulstep 5ripup net 393 → 4 open

Net 40 (/U2_TX) failed to route with both route and route_via. Following the lead engineer's note, I will rip up net 39 (/U2_RX) first, as it is a 2-pad net blocking both /U2_TX and /U0_RX on the bottom layer.

boldstep 5route_via net 424 → 3 open

With /U2_RX and /CAN_TX cleared and /U2_TX routed, route /U0_RX (net 42) using via bridges.

layer-firststep 5ripup net 39, 413 → 5 open

route_via failed on all three open nets. Ripping up 2-pad nets /U2_RX (39) and /CAN_TX (41) as advised to clear bottom-layer obstacles for /U0_RX and /U2_TX.

power-awarestep 5route_via net 425 → 4 open

With /U2_RX and /CAN_TX ripped up, try routing /U0_RX (net 42) using route_via to bridge under top-layer congestion.

carefulstep 6route_via net 394 → 4 open

Net 39 (/U2_RX) was just ripped up; try routing it with vias using shove mode.

boldstep 6route_via net 393 → 3 open

With /U2_TX and /U0_RX now routed, attempt to route /U2_RX (net 39) using via bridges.

layer-firststep 6route_via net 425 → 4 open

With /U2_RX and /CAN_TX cleared, try route_via on /U0_RX (net 42) to bridge it on bottom copper.

power-awarestep 6route_via net 404 → 3 open

Now that /U2_RX and /CAN_TX are cleared and /U0_RX is routed, try routing /U2_TX (net 40) using via bridges.

carefulstep 7ripup net 414 → 5 open

Net 41 (/CAN_TX) is a 2-pad net crossing /U2_RX, /U2_TX, and /U0_RX on the bottom layer; ripping it up will clear the path for those nets.

boldstep 7route_via net 413 → 3 open

Net 39 failed to route via, so let's try routing net 41 (/CAN_TX) with route_via in shove mode.

layer-firststep 7route_via net 404 → 3 open

With /U2_RX and /CAN_TX ripped up and /U0_RX routed, try route_via on /U2_TX (net 40) to complete it using a bottom-layer via bridge.

power-awarestep 7route_via net 393 → 3 open

Nets 42 and 40 are now routed, so let's try routing net 39 (/U2_RX) with via bridges.

carefulstep 8route_via net 425 → 4 open

With /U2_RX and /CAN_TX cleared, try routing /U0_RX (net 42) using route_via to bridge on bottom copper under top-layer congestion.

boldstep 8route_via net 613 → 3 open

route_via failed for nets 39 and 41 in the current state; try routing /SD_DAT0 (net 61) with via bridges as suggested.

layer-firststep 8route_via net 393 → 3 open

Net 42 and 40 are now routed, so try route_via on open net 39 (/U2_RX) next.

power-awarestep 8route_via net 413 → 3 open

Net 41 (/CAN_TX) was ripped up earlier and has not been routed yet; try route_via with shove mode.

carefulstep 9route_via net 414 → 4 open

Net 42 (/U0_RX) is now routed, so let's try routing net 41 (/CAN_TX) with a via bridge.

boldstep 9ripup net 403 → 4 open

Net 40 (/U2_TX) heavily crosses nets 41 (/CAN_TX), 61 (/SD_DAT0), and 39 (/U2_RX), blocking their via bridges. Rip up net 40 to clear space.

layer-firststep 9route_via net 413 → 3 open

Net 39 (/U2_RX) failed route_via, so let's try route_via on the next open connection, net 41 (/CAN_TX), before ripping anything.

power-awarestep 9route_via net 613 → 3 open

Nets 39 and 41 just failed route_via. Let's try routing net 61 (/SD_DAT0) with route_via.

carefulstep 10route_via net 614 → 4 open

Net 61 (/SD_DAT0) has not been attempted with route_via yet; try bridging it with vias.

boldstep 10route_via net 394 → 4 open

With /U2_TX (40) just ripped up, try routing /U2_RX (39) using route_via in walk-around mode (mode=2) to see if it can now find a path.

layer-firststep 10ripup net 403 → 4 open

route_via failed on /CAN_TX (41) and /U2_RX (39); /U2_TX (40) crosses /CAN_TX and /SD_DAT0 7 times each, so rip it to clear a path.

power-awarestep 10route_via net 393 → 3 open

route_via with shove (mode 1) failed for net 39 (/U2_RX), so try walk-around (mode 2) to see if it can route around the obstacles.

carefulstep 11route_via net 404 → 3 open

Nets 39 (/U2_RX) and 41 (/CAN_TX) are now ripped up, which may have cleared the path for net 40 (/U2_TX) to route with a via bridge.

boldstep 11route_via net 414 → 4 open

Net 41 (/CAN_TX) is only crossed by top-layer nets, so routing it via bottom copper with walk-around mode should succeed now that net 40 is cleared.

layer-firststep 11route_via net 394 → 4 open

Net 40 (/U2_TX) was just ripped up; try routing net 39 (/U2_RX) with route_via now.

power-awarestep 11ripup net 403 → 4 open

Net 40 (/U2_TX) is a 2-pad signal net that crosses /CAN_TX (41) 7 times, /SD_DAT0 (61) 7 times, and /U2_RX (39) once. Ripping up net 40 will clear the main obstacle for these three open connections.

carefulstep 12route_via net 393 → 3 open

Net 40 (/U2_TX) was just routed, changing the congestion around net 39 (/U2_RX). Try routing net 39 with vias now.

boldstep 12route_via net 394 → 4 open

With /U2_TX (40) ripped up, try routing /U2_RX (39) using route_via in shove mode (mode=1).

layer-firststep 12route_via net 414 → 4 open

Try route_via on /CAN_TX (net 41) in the current state before ripping any more nets.

power-awarestep 12route_via net 414 → 4 open

Net 40 was just ripped up, so try routing net 41 (/CAN_TX) with vias now that net 40's copper is out of the way.

carefulstep 13route_via net 613 → 3 open

route_via with mode 1 failed for nets 39, 41, and 61; try route_via with mode 2 (walk around) on net 61 (/SD_DAT0).

boldstep 13ripup net 43, 374 → 6 open

Rip up 2-pad nets /U0_TX (43) and /CAN_STB (37) that cross /U2_RX (39) to clear space for routing.

layer-firststep 13route_via net 614 → 4 open

Try route_via on /SD_DAT0 (net 61) since we haven't tried it in the current board state.

power-awarestep 13route_via net 394 → 4 open

Net 40 (/U2_TX) was ripped up, which may have cleared space for net 39 (/U2_RX) to be routed with a via bridge.

carefulstep 14route_via net 413 → 3 open

Net 41 (/CAN_TX) has not been tried with route_via in walk-around mode (mode=2) in the current 3-open state.

boldstep 14route_via net 416 → 6 open

With /CAN_STB and /U0_TX ripped up, try routing /CAN_TX (41) using via bridges.

layer-firststep 14route_via net 404 → 3 open

Net 40 (/U2_TX) was ripped up in step 7 and has not been retried with route_via in the current state.

power-awarestep 14restore_best4 → 3 open

Restore the best board state with 3 open connections so we don't waste steps re-routing net 40.

carefulstep 15route net 393 → 3 open

Try routing /U2_RX (net 39) on its current layer with shove mode before ripping up any blocking nets.

boldstep 15route_via net 436 → 6 open

Net 43 (/U0_TX) was just ripped up and has a relatively short 20.14mm path; try routing it with via bridges now that space is cleared.

layer-firststep 15route net 393 → 3 open

route_via failed on net 39 (/U2_RX) in 90 trials; try single-layer shove routing with route mode=1.

power-awarestep 15ripup net 39, 413 → 5 open

Rip up 2-pad nets /U2_RX (39) and /CAN_TX (41) as advised by the lead engineer so they can be rerouted out of the way of /U0_RX and /U2_TX.

carefulstep 16ripup net 403 → 4 open

Net 40 (/U2_TX) crosses both /CAN_TX (net 41) and /SD_DAT0 (net 61) 7 times, blocking route_via. Ripping up net 40 should clear a path for them.

boldstep 16restore_best6 → 3 open

We are at 6 open connections with only 5 steps left; restore_best returns us to the 3-open state.

layer-firststep 16ripup net 433 → 4 open

Net 39 (/U2_RX) failed route_via and route; ripup 2-pad net 43 (/U0_TX) which crosses it on both layers.

power-awarestep 16route_via net 425 → 4 open

With /U2_RX and /CAN_TX ripped up, try routing /U0_RX (net 42) with vias to clear its open connection.

carefulstep 17route_via net 394 → 4 open

Net 40 (/U2_TX) is now ripped up, which may have cleared the path for net 39 (/U2_RX). Try route_via on net 39 with walk around mode.

boldstep 17ripup net 39, 413 → 5 open

Nets 39 (/U2_RX) and 41 (/CAN_TX) are blocking nets 40 (/U2_TX) and 42 (/U0_RX) on the bottom layer. Ripping them up will clear space to route 40 and 42.

layer-firststep 17route_via net 434 → 4 open

Net 43 (/U0_TX) is open and we haven't tried route_via on it yet; let's try bridging it with vias.

power-awarestep 17route_via net 404 → 3 open

With /U2_RX and /CAN_TX ripped up and /U0_RX routed, try routing /U2_TX (net 40) with via bridges.

carefulstep 18restore_best4 → 3 open

Ripping up /U2_TX (40) made things worse (4 open connections instead of 3) and /U2_RX (39) still could not route. Restore the best board state with 3 open connections.

boldstep 18route_via net 425 → 4 open

With /U2_RX and /CAN_TX ripped up, /U0_RX (net 42) should now have a clear bottom-layer path via a bridge.

layer-firststep 18restore_best4 → 3 open

Restore the best board seen so far (3 open connections) since ripping up and routing net 43 failed.

power-awarestep 18route_via net 393 → 3 open

Nets 42 and 40 were just rerouted, so try route_via on net 39 (/U2_RX) with shove mode to close it.

carefulstep 19route_via net 423 → 3 open

Net 42 (/U0_RX) has not been attempted yet in this state; try route_via with shove mode to bridge it.

boldstep 19route_via net 404 → 3 open

With /U0_RX routed and /U2_RX and /CAN_TX out of the way, try routing /U2_TX (40) with via bridges.

layer-firststep 19route_via net 423 → 3 open

Try route_via on /U0_RX (net 42) to close it using a via bridge.

power-awarestep 19route_via net 413 → 3 open

Net 39 failed route_via in the current state; try route_via on net 41 (/CAN_TX) now that nets 42 and 40 are routed.

carefulstep 20route_via net 423 → 3 open

With one step left, try routing /U0_RX (net 42) with a via bridge using walk-around mode (mode 2) since shove mode (mode 1) just failed.

boldstep 20route_via net 393 → 3 open

With /U0_RX and /U2_TX now routed, try routing /U2_RX (net 39) using route_via in shove mode.

layer-firststep 20route_via net 403 → 3 open

Net 40 (/U2_TX) is open; try route_via on it to close the connection with a via bridge.

power-awarestep 20route_via net 613 → 3 open

route_via on nets 39 and 41 just failed in this state, so try route_via on the remaining open net 61 (/SD_DAT0) with shove mode.

hard-tier test board

RGBMatrixPanel CPLD (0599)

Gemini agents: 60 open → 31 openrandom control: 60 open → 50 open81 Gemini calls · 301,884 tokens
board before
before · classic router · 60 open
board after
after · power-aware · 31 open

Open connections after each step

a rip-up pushes it up first
025507505101520stepclean
careful 31bold 33layer-first 33power-aware 31random control

What Gemini decided

straight from llm_calls.jsonl
Leadgemini-4-carbonstrategy note from the board picture

First, reroute simple 2-pad nets /CPLD_A3 and /DRVL to clear local congestion and DRC errors, then finish all unblocked +5V and GND segments. Leave existing 84-pad GND traces, 5-pad /FB, and /INTVCC alone—ripping up massive ground networks or sensitive regulator/feedback traces risks cascading failures. Use via bridges where single-layer blockers remain: route +5V (84.6, 94.9 to 93.2, 94.7) on bottom copper to pass under top-layer GND (and /CPLD_A3), and bridge GND onto top copper to cross over bottom-layer /INTVCC and /FB (x=64–74, y=90–101).

carefulstep 1route net 160 → 54 open

Net 1 (+5V) has multiple open connections with no crossings, so routing it on its current layer with shove mode should close several connections.

boldstep 1route_via net 160 → 53 open

Net 1 (+5V) has multiple open connections, most with no crossings. Routing it with via bridges should close several connections cleanly.

layer-firststep 1route_via net 160 → 53 open

Following layer-first style, try route_via on net 1 (+5V) to close its open connections using layer changes.

power-awarestep 1route net 160 → 54 open

Route the +5V power net (net 1) first, as most of its open connections have no crossings.

carefulstep 2route_via net 154 → 53 open

Net 1 (+5V) has only 1 open connection left after mode 1 routing; try route_via to bridge it.

boldstep 2route_via net 253 → 53 open

Net 2 (GND) has many open connections, some with no crossings. Using route_via with shove mode should close several of them.

layer-firststep 2route_via net 253 → 53 open

Net 2 (GND) has many open connections; following the layer-first style, try route_via on net 2 before ripping anything.

power-awarestep 2route net 254 → 37 open

GND (net 2) has many open connections with no crossings; routing it with shove mode should close several connections.

carefulstep 3route net 253 → 36 open

Net 2 (GND) has many open connections, several with no crossings. Try routing net 2 with shove mode first to close as many as possible.

boldstep 3ripup net 6553 → 54 open

Rip up 2-pad net /DRVL (net 65) as advised to clear local congestion blocking GND segments.

layer-firststep 3ripup net 6553 → 54 open

route_via on GND (net 2) made 0 progress. /DRVL (net 65) is a simple 2-pad net crossing GND in multiple places, so ripping it up will clear local congestion as advised.

power-awarestep 3ripup net 6537 → 38 open

Rip up 2-pad net /DRVL (net 65) as advised to clear local congestion blocking GND segments.

carefulstep 4route_via net 236 → 36 open

GND (net 2) still has 10 open connections blocked by single-layer traces like /INTVCC and /FB; use route_via to bridge GND across them.

boldstep 4route_via net 6554 → 54 open

Net 65 was just ripped up to clear local congestion; reroute it cleanly using route_via.

layer-firststep 4route_via net 6554 → 54 open

Reroute net 65 which was just ripped up to clear local congestion.

power-awarestep 4route net 6538 → 37 open

Route net 65 that was just ripped up to clear local congestion.

carefulstep 5ripup net 6536 → 37 open

GND (net 2) is blocked by the 2-pad net /DRVL (net 65) at two locations, and route_via on GND failed. Ripping up /DRVL (net 65) will clear local congestion so GND can be routed.

boldstep 5route net 6554 → 53 open

route_via failed on net 65, so try single-layer shove routing to reconnect it.

layer-firststep 5route net 6554 → 53 open

Net 65 was ripped up and route_via failed to close it, so try single-layer routing with shove mode.

power-awarestep 5route_via net 137 → 36 open

Complete the remaining +5V connection using route_via so it can change layers if blocked on its current layer.

carefulstep 6route net 6537 → 36 open

We just ripped up net 65 to clear congestion; now let's reroute net 65 cleanly.

boldstep 6route net 253 → 37 open

GND (net 2) has many unblocked open connections; routing it on its current layer with shove mode should close the unblocked segments.

layer-firststep 6route net 253 → 37 open

We haven't tried single-layer routing on GND (net 2) yet; let's route all unblocked GND segments on their current layer.

power-awarestep 6route_via net 236 → 36 open

GND (net 2) has several open connections, some unblocked and others crossed by bottom-layer /INTVCC and /FB. Using route_via on net 2 will route the unblocked segments and bridge over the bottom-layer blockers.

carefulstep 7route_via net 736 → 35 open

Net 7 (/LAT) is a 2-pad signal net crossed by GND and +3V3; trying route_via to bridge across the blockers.

boldstep 7route_via net 237 → 37 open

GND (net 2) now has only 11 open connections, several of which are blocked by bottom-layer /INTVCC and /FB; trying route_via on net 2 will bridge GND onto top copper.

layer-firststep 7route_via net 237 → 37 open

GND (net 2) now has only 11 open connections left, several of which cross /INTVCC and /FB on layer 2. Trying route_via on net 2 to bridge those crossings onto top copper.

power-awarestep 7route_via net 736 → 35 open

route_via on GND failed; try route_via on 2-pad signal net /LAT (net 7) to clear an open connection.

carefulstep 8route_via net 435 → 35 open

Net 4 (/R1) is a 2-pad signal net with an open connection; try routing it with a via bridge.

boldstep 8route_via net 737 → 36 open

route_via on GND made no progress; let's try routing /LAT (net 7) with vias to close its open connection.

layer-firststep 8route_via net 337 → 36 open

route_via on net 2 failed to make any connections; try route_via on net 3 (+3V3) per layer-first style.

power-awarestep 8route_via net 335 → 34 open

Route power net +3V3 (net 3) using route_via to bridge across GND and signal traces.

carefulstep 9route_via net 835 → 34 open

Net 8 (/B1) is a 2-pad signal net that is open and crossed by several traces; try routing it with vias.

boldstep 9route_via net 436 → 36 open

Net 4 (/R1) is a 2-pad net with one open connection crossed by several nets on both layers; route_via can bridge across them.

layer-firststep 9route_via net 436 → 36 open

Try route_via on open net 4 (/R1) before ripping any blocking nets.

power-awarestep 9route_via net 434 → 34 open

GND route_via failed earlier, and we have several 2-pad signal nets open. Let's try routing /R1 (net 4) with vias.

carefulstep 10route_via net 934 → 33 open

Net 9 (/CLK) is a 2-pad signal net that is open; trying route_via to bridge across the blocking traces.

boldstep 10route_via net 836 → 35 open

Net 8 (/B1) has not been tried with route_via yet; let's attempt to route it using via bridges.

layer-firststep 10route_via net 736 → 35 open

Following layer-first style, try route_via on open net 7 (/LAT) before ripping any blockers.

power-awarestep 10route net 234 → 34 open

GND (net 2) has 10 open connections, several with no direct crossings, so trying walk-around mode (mode=2) may resolve some of them.

carefulstep 11route_via net 1033 → 33 open

Try routing 2-pad net 10 (/D) with vias to reduce open connections.

boldstep 11route_via net 335 → 34 open

Net 3 (+3V3) has an open connection that hasn't been tried with route_via yet; let's attempt to route it using vias.

layer-firststep 11route_via net 835 → 34 open

Net 8 (/B1) has an open connection and has not been tried with route_via yet; try via bridging before ripping any blockers.

power-awarestep 11route_via net 834 → 33 open

Net 8 (/B1) is an open 2-pad signal net crossed by several traces on both layers; try routing it with vias.

carefulstep 12route_via net 1233 → 33 open

Net 12 (/C) is a 2-pad signal net that hasn't been attempted yet; try routing it with vias.

boldstep 12route_via net 934 → 33 open

Net 9 (/CLK) is a 2-pad net that has not been attempted yet; route_via may close it by bridging over blockers.

layer-firststep 12route_via net 934 → 33 open

Net 9 (/CLK) is open and hasn't been tried with route_via yet. Trying route_via first per layer-first style.

power-awarestep 12route_via net 1233 → 33 open

Net 12 (/C) has not been attempted yet and has fewer crossings than /CLK or /R1; try routing it with vias.

carefulstep 13route_via net 333 → 32 open

Net 3 (+3V3) has an open connection that hasn't been tried with route_via yet; let's try bridging it.

boldstep 13route_via net 1033 → 33 open

Try routing /D (net 10) with via bridges to clear one of the long open signal connections.

layer-firststep 13route_via net 1033 → 33 open

Following the layer-first strategy, try route_via on open net 10 (/D) before ripping anything.

power-awarestep 13route net 233 → 33 open

GND (net 2) has 10 open connections, several with no crossings, and we haven't tried routing it with shove (mode 1) yet.

carefulstep 14route net 232 → 32 open

GND (net 2) has several unblocked open connections (crossed_by is empty); routing on its current layer with shove mode should close them.

boldstep 14route_via net 1233 → 33 open

Net 12 (/C) has not been attempted yet and is crossed by nets on both layers, so route_via with shove mode is the best first attempt.

layer-firststep 14route_via net 1233 → 33 open

Following the layer-first style, try route_via on open net 12 (/C) before ripping any blocking nets.

power-awarestep 14route_via net 933 → 32 open

GND routing failed in both modes and route_via; try route_via on /CLK (net 9) which hasn't been attempted yet.

carefulstep 15route_via net 232 → 32 open

Single-layer routing for GND (net 2) made 0 connections. Trying route_via to bridge GND onto top copper over bottom-layer blockers like /INTVCC and /FB.

boldstep 15route net 233 → 33 open

GND (net 2) has several open connections with no crossings that failed under shove mode; try walk-around mode (mode=2) to close them.

layer-firststep 15ripup net 6533 → 34 open

route_via failed on all remaining open nets (2, 4, 10, 12). Net 65 (/DRVL) is a small 2-pad net that crosses two open GND connections, so ripping it up will unblock GND.

power-awarestep 15route_via net 1032 → 32 open

Net 10 (/D) has not been attempted yet; try route_via with shove mode to bridge across the blocking traces.

carefulstep 16route_via net 1332 → 31 open

Net 13 (/G2) has not been tried yet; let's attempt route_via in shove mode to see if it can bridge across.

boldstep 16ripup net 6533 → 34 open

Net 65 (/DRVL) is a simple 2-pad net that blocks two GND connections. Ripping it up will clear space to route GND, and then we can re-route /DRVL.

layer-firststep 16route_via net 6534 → 34 open

Net 65 was just ripped up to clear congestion; route it now using via bridges if needed.

power-awarestep 16route_via net 1332 → 31 open

Net 13 (/G2) is an open 2-pad signal net that has not been attempted yet; try routing it with via bridges.

carefulstep 17route_via net 1431 → 31 open

Net 14 (/B) has not been attempted with route_via yet; try bridging it across the blocking traces.

boldstep 17route_via net 6534 → 34 open

Reroute net 65 that was just ripped up using route_via with shove mode to clear local congestion.

layer-firststep 17route net 234 → 33 open

Several GND (net 2) connections have no crossings, and route_via failed earlier, so try same-layer routing on net 2.

power-awarestep 17route_via net 231 → 31 open

GND (net 2) has 10 open connections, several of which are blocked by bottom-layer /INTVCC, /FB, /DRVL, /TCK, and +3V3. Using route_via on net 2 should bridge these blockers.

carefulstep 18route net 231 → 31 open

GND (net 2) has several unblocked open connections, and shove mode failed to route any of them. Trying walk-around mode (mode=2) to see if it can complete the unblocked GND segments.

boldstep 18restore_best34 → 33 open

Ripping up net 65 increased open connections to 34 and route_via failed to close it, so restore the best board seen so far (33 open connections).

layer-firststep 18route_via net 233 → 33 open

Net 2 (GND) still has 10 open connections, several of which are crossed by single-layer traces; try route_via on net 2 to bridge over them.

power-awarestep 18route_via net 1431 → 31 open

Net 14 (/B) has not been attempted yet; trying route_via with shove mode to see if it can find a path.

carefulstep 19route_via net 431 → 31 open

Net 4 (/R1) is open and has not been tried with route_via yet; let's attempt a via bridge to route it.

boldstep 19route_via net 233 → 33 open

GND (net 2) has 11 open connections, many of which are blocked by single-layer traces on layer 2. Using route_via can bridge GND over them.

layer-firststep 19route_via net 1333 → 33 open

Net 13 (/G2) has not been tried with route_via yet; trying route_via to see if it can bridge across the blockers.

power-awarestep 19route_via net 231 → 31 open

route_via on GND (net 2) with mode 1 failed, so try mode 2 (walk around) to bridge over the single-layer blockers.

carefulstep 20route_via net 231 → 31 open

With 1 step left, try route_via on GND (net 2) with walk-around mode (mode 2) to close any of its remaining open connections without increasing open count.

boldstep 20route_via net 433 → 33 open

Net 4 (/R1) is an open 2-pad connection that hasn't been attempted with route_via in the current state.

layer-firststep 20route_via net 433 → 33 open

Try route_via on net 4 (/R1) as it has not been attempted yet in this state.

power-awarestep 20route_via net 431 → 31 open

Net 4 (/R1) is an open 2-pad signal net that has not been attempted with route_via yet; try routing it with vias in shove mode.

hard-tier test board

HybridChargeController (0626)

Gemini agents: 1 open → 1 openrandom control: 1 open → 1 open81 Gemini calls · 196,076 tokens
board before
before · classic router · 1 open
board after
after · router · 1 open

Open connections after each step

a rip-up pushes it up first
0204005101520stepclean
careful 2bold 1layer-first 1power-aware 1random control

What Gemini decided

straight from llm_calls.jsonl
Leadgemini-4-carbonstrategy note from the board picture

To connect /COM+ and clear the rule error at [156.1, 53.4]: 1. Move first: Reroute small bottom-layer nets /SW_LD (2 pads) and /S_LA_V (3 pads) out of the path, along with Net-(Q4-Pad1) (4 pads, 3 crossings) and Net-(Q6-Pad3) (4 pads) near the endpoints if needed, resolving the short/DRC violation at the red X. 2. Leave alone: Keep large top-layer nets /GND (38 pads) and /LA+ (10 pads) untouched. 3. Via bridge: Drop /COM+ onto the sparse bottom layer via a bridge to pass cleanly underneath /GND, /LA+, Net-(Q4-Pad1), and Net-(Q6-Pad3).

carefulstep 1route net 321 → 1 open

Net 32 (/COM+) is the only open connection. Try routing it directly with shove mode first before attempting vias or ripups.

boldstep 1route_via net 321 → 1 open

Only net 32 (/COM+) is open. Try route_via with shove mode first to see if it can bridge over the crossing nets using vias.

layer-firststep 1route_via net 321 → 1 open

Net 32 (/COM+) is the only open connection. Following the layer-first strategy, try route_via on it first to see if a via bridge can bypass the crossing nets.

power-awarestep 1route_via net 321 → 1 open

Net 32 (/COM+) is the only open connection; try route_via with shove mode first to see if it can bridge across the obstacles.

carefulstep 2route_via net 321 → 1 open

Net 32 failed to route on its current layer with shove mode. Try route_via to bridge across the blocking nets.

boldstep 2ripup net 15, 581 → 4 open

Rip up small bottom-layer nets /SW_LD (15) and /S_LA_V (58) to clear space on the bottom layer for /COM+ (32) to bridge.

layer-firststep 2ripup net 501 → 4 open

route_via on net 32 failed; rip up Net-(Q4-Pad1) (net 50) which crosses the path 3 times.

power-awarestep 2ripup net 15, 58, 501 → 7 open

route_via on /COM+ (net 32) failed because small signal nets block both the top and bottom layers. Rip up the small blocking nets /SW_LD (15), /S_LA_V (58), Net-(Q4-Pad1) (50), and Net-(Q6-Pad3) (54) while leaving /GND and /LA+ untouched.

carefulstep 3ripup net 151 → 2 open

route_via on net 32 failed, so rip up the smallest blocking bottom-layer net 15 (/SW_LD, 2 pads) to clear space for a via bridge.

boldstep 3route_via net 324 → 4 open

With /SW_LD and /S_LA_V ripped up, try routing /COM+ (net 32) using a via bridge.

layer-firststep 3route_via net 324 → 4 open

Net 50 (Net-(Q4-Pad1)) is now ripped up, so try route_via on net 32 (/COM+) to see if it can bridge across.

power-awarestep 3route_via net 327 → 7 open

We ripped up the blocking small nets (15, 58, 50). Now we route power net /COM+ (32) first using route_via to bridge underneath /GND and /LA+.

carefulstep 4route_via net 322 → 2 open

Now that /SW_LD (net 15) is ripped up, try routing /COM+ (net 32) using a via bridge.

boldstep 4ripup net 50, 544 → 10 open

route_via on net 32 (/COM+) still failed; rip up the small crossing nets 50 (Net-(Q4-Pad1)) and 54 (Net-(Q6-Pad3)) to clear space near the endpoints.

layer-firststep 4ripup net 15, 584 → 7 open

route_via on net 32 (/COM+) failed even after ripping net 50. Rip small bottom-layer nets 15 (/SW_LD) and 58 (/S_LA_V) to clear the bottom layer for a via bridge.

power-awarestep 4ripup net 61, 547 → 12 open

Net 32 (/COM+) failed to route via the bottom layer, likely blocked by small signal nets 61 (/S_LA_I) on layer 2 and 54 (Net-(Q6-Pad3)) on layer 0. Ripping up these small nets to clear the path.

carefulstep 5ripup net 582 → 4 open

Net 58 (/S_LA_V) is on layer 2 and crosses net 32 (/COM+), blocking a bottom-layer via bridge. Ripping up net 58 will clear the path.

boldstep 5route_via net 3210 → 9 open

With nets 15, 58, 50, and 54 ripped up, try routing net 32 (/COM+) using a via bridge.

layer-firststep 5route_via net 327 → 7 open

Now that /SW_LD (15) and /S_LA_V (58) are ripped up, try route_via on /COM+ (32) to bridge underneath the top-layer nets.

power-awarestep 5route_via net 3212 → 11 open

With nets 61 and 54 ripped up, try route_via on /COM+ (net 32) to bridge under /GND and /LA+.

carefulstep 6route_via net 324 → 4 open

Both /SW_LD and /S_LA_V have been ripped up to clear the bottom layer; now try route_via on /COM+ (net 32).

boldstep 6route_via net 549 → 6 open

Net 32 (/COM+) is now routed with 0 rule errors. Let's re-route the ripped-up nets starting with net 54 (Net-(Q6-Pad3)), which has short connections.

layer-firststep 6ripup net 547 → 10 open

route_via on /COM+ (32) still failed after ripping up 50, 15, and 58; Net-(Q6-Pad3) (54) crosses near the endpoint and is small (4 pads), so rip it up next.

power-awarestep 6route_via net 5411 → 8 open

Net 32 (/COM+) is now connected. Next, route the short net 54 (Net-(Q6-Pad3)) using route_via to handle layer crossings.

carefulstep 7ripup net 504 → 7 open

Ripping up Net-(Q4-Pad1) (net 50) which has 3 crossings across /COM+ on layer 0 to clear the path for /COM+.

boldstep 7route_via net 506 → 3 open

Net 32 (/COM+) and Net 54 are now routed. Next, route Net 50 (Net-(Q4-Pad1)) using route_via to clear its 3 open connections.

layer-firststep 7route_via net 3210 → 9 open

With nets 50, 15, 58, and 54 ripped up, try route_via on /COM+ (net 32) to see if it can now bridge cleanly.

power-awarestep 7route_via net 508 → 5 open

Net 32 (/COM+) and net 54 (Net-(Q6-Pad3)) are now routed. Next, route net 50 (Net-(Q4-Pad1)) using route_via to clear its 3 open connections.

carefulstep 8route_via net 327 → 7 open

Net 50 has been ripped up, so let's try routing /COM+ (net 32) with a via bridge now.

boldstep 8route_via net 153 → 2 open

Net 32, 54, and 50 are routed. Now route /SW_LD (net 15) using route_via.

layer-firststep 8route_via net 509 → 6 open

Net 32 (/COM+) is now connected. Following the layer-first strategy, try route_via on the open connections of net 50 (Net-(Q4-Pad1)).

power-awarestep 8route_via net 155 → 4 open

Net 15 (/SW_LD) is a 2-pad signal net with one open connection crossed by power/ground nets on both layers; route_via can bridge around them.

carefulstep 9ripup net 547 → 10 open

Net-(Q6-Pad3) (net 54) still crosses /COM+ near its endpoint; ripping it up as suggested before retrying /COM+.

boldstep 9route_via net 582 → 1 open

Net 58 (/S_LA_V) is the only remaining open net; route_via with shove mode should complete both its short and long connections.

layer-firststep 9route_via net 546 → 3 open

Net 54 (Net-(Q6-Pad3)) has three short open connections; try routing it with vias first.

power-awarestep 9route_via net 584 → 3 open

Net 58 (/S_LA_V) has 2 open connections and hasn't been routed yet after ripup; try route_via with shove mode.

carefulstep 10route_via net 3210 → 9 open

Now that Net-(Q6-Pad3) (net 54) is ripped up, try routing /COM+ (net 32) with a via bridge to see if the path is clear.

boldstep 10ripup net 611 → 3 open

Net 58 (/S_LA_V) has 1 open connection left and route_via could not close it. Net 61 (/S_LA_I) has 5 crossings on both layers and only 3 pads, so ripping it up will clear space to route net 58.

layer-firststep 10route_via net 153 → 2 open

Net 15 (/SW_LD) is open; try route_via first to see if it can complete by changing layers.

power-awarestep 10route_via net 613 → 1 open

Net 61 (/S_LA_I) has not been routed yet since it was ripped up, and one of its two open connections has zero crossings. Let's route_via net 61 with shove mode.

carefulstep 11route net 549 → 7 open

Net 32 (/COM+) is now connected; next, route the short open connections of Net-(Q6-Pad3) (net 54).

boldstep 11route_via net 613 → 1 open

Net 61 was just ripped up to clear its path; try routing it with via bridges.

layer-firststep 11route_via net 582 → 1 open

Net 58 (/S_LA_V) has 2 open connections and 0 rule errors; try route_via on net 58 first.

power-awarestep 11route_via net 581 → 1 open

Only net 58 (/S_LA_V) remains open with zero rule errors; try routing it with vias using shove mode.

carefulstep 12route net 327 → 6 open

Net 32 (/COM+) has a tiny 0.66mm open connection with no crossings, so routing it directly with shove mode should close it.

boldstep 12ripup net 13, 141 → 3 open

Net 58 (/S_LA_V) failed to route_via, and it is crossed by two small 2-pad nets: net 13 (/LED_BTN) on layer 2 and net 14 (/SW_LA) on layer 0. Ripping up both small nets will clear space to route net 58.

layer-firststep 12ripup net 611 → 3 open

Net 58 (/S_LA_V) has 1 open connection left and route_via just failed to close it in 90 trials. Net 61 (/S_LA_I, 3 pads) crosses its path 5 times on both layers, so ripping up net 61 should clear the way.

power-awarestep 12route_via net 581 → 1 open

Net 58 (/S_LA_V) is the only remaining open connection; route_via with shove mode failed, so try walk-around mode.

carefulstep 13route net 506 → 3 open

Net 32 (/COM+) is now routed; start re-routing the ripped-up nets beginning with net 50 (Net-(Q4-Pad1)) using shove mode.

boldstep 13route_via net 583 → 3 open

We just ripped up nets 13 and 14 to clear space; now route net 58 (/S_LA_V) using route_via.

layer-firststep 13route_via net 613 → 1 open

Net 61 (/S_LA_I) was just ripped up and now has 2 open connections; try route_via on it to complete its connections across layers.

power-awarestep 13ripup net 13, 141 → 3 open

Net 58 (/S_LA_V) is the last open connection and route_via failed in both modes. Ripping up small 2-pad signal nets 13 and 14 that cross it on layers 2 and 0 should clear a path.

carefulstep 14route net 153 → 3 open

Net 32 and Net 50 are routed; now try routing the open connection on small net 15 (/SW_LD) using shove mode.

boldstep 14restore_best3 → 1 open

Ripping up nets 13 and 14 did not help route net 58. Restore the best board (1 open connection) to save steps and try clearing a different blocker for net 58.

layer-firststep 14ripup net 501 → 4 open

Net 58 (/S_LA_V) is the last open connection and is crossed twice by small net 50 (Net-(Q4-Pad1)); rip up net 50 to clear the path.

power-awarestep 14route_via net 583 → 3 open

Nets 13 and 14 were ripped up to clear space; route net 58 (/S_LA_V) first with route_via mode 1 before re-routing the smaller 2-pad nets.

carefulstep 15route_via net 153 → 2 open

Net 15 (/SW_LD) failed to route on its current layer with mode 1; try route_via to bridge across the blocking nets.

boldstep 15ripup net 15, 581 → 4 open

Rip up the two small bottom-layer nets /SW_LD (15) and /S_LA_V (58) to clear space on the bottom layer for /COM+ (32) to use a via bridge.

layer-firststep 15route_via net 504 → 1 open

Net 50 was just ripped up to clear the path for net 58. Re-route net 50 using route_via first, as it previously routed all 3 connections cleanly.

power-awarestep 15restore_best3 → 1 open

Ripping up nets 13 and 14 did not allow net 58 to route. Restore the best board (1 open connection) so we can target a net that actually crosses net 58.

carefulstep 16route net 582 → 1 open

Net 58 (/S_LA_V) has two open connections, one of which is a short 1.8 mm uncrossed segment. Try routing net 58 with shove mode first.

boldstep 16route_via net 324 → 4 open

With /SW_LD and /S_LA_V ripped up, route /COM+ (net 32) using route_via to bridge underneath the top-layer nets.

layer-firststep 16route_via net 581 → 1 open

Net 50 was just re-routed, potentially clearing the path for net 58. Try route_via on net 58 again.

power-awarestep 16route_via net 321 → 1 open

Only net 32 (/COM+) is open; try route_via with shove mode to bridge it onto the bottom layer.

carefulstep 17route_via net 581 → 1 open

Net 58 (/S_LA_V) has 1 open connection left and 0 rule errors; try route_via with shove mode to bridge across the blocking nets.

boldstep 17ripup net 504 → 7 open

Net 50 crosses both net 32 (/COM+) and net 58 (/S_LA_V) three times. Ripping up net 50 should clear the path for net 32.

layer-firststep 17route net 581 → 1 open

route_via on net 58 failed, so try routing net 58 directly on its current layer with shove mode.

power-awarestep 17ripup net 15, 581 → 4 open

Ripup the small bottom-layer nets 15 (/SW_LD) and 58 (/S_LA_V) to clear the bottom layer for /COM+ (net 32) to pass via a two-via bridge.

carefulstep 18ripup net 611 → 3 open

Net 58 (/S_LA_V) failed to route with both route and route_via. Net 61 (/S_LA_I) crosses it 3 times on both layers and has only 3 pads, so ripping it up should clear the path.

boldstep 18route_via net 327 → 7 open

Net 50 (Net-(Q4-Pad1)) is now ripped up alongside nets 15 and 58, so try routing /COM+ (net 32) with a via bridge.

layer-firststep 18route_via net 581 → 1 open

Net 58 (/S_LA_V) failed with shove mode (mode=1) for both route_via and route; try route_via with walk-around (mode=2) to see if it can find a path.

power-awarestep 18route_via net 324 → 4 open

With /SW_LD and /S_LA_V ripped up, route /COM+ (net 32) first using route_via to bridge under /GND and /LA+.

carefulstep 19route_via net 613 → 2 open

Net 61 (/S_LA_I) was just ripped up and has 2 open connections; try route_via to close them across layers.

boldstep 19restore_best7 → 1 open

With only 2 steps left and 7 open connections, restore the best board (1 open connection) and try one final action if possible.

layer-firststep 19route net 581 → 1 open

Only 2 steps left, so ripping up another net would leave it unrouted; trying route with mode 2 (walk around) on net 58 as it has not been attempted yet.

power-awarestep 19restore_best4 → 1 open

We only have 2 steps left and 4 open connections after ripping up nets 15 and 58. Restoring the best board brings us back to 1 open connection, leaving 1 step to try routing net 32 in walk-around mode.

carefulstep 20route net 612 → 2 open

Net 61 (/S_LA_I) still has 1 open connection after route_via; try routing it on its current layer with shove mode.

boldstep 20route_via net 321 → 1 open

With only one step left, try routing net 32 (/COM+) with route_via in walk-around mode (mode=2) to see if it can find a path.

layer-firststep 20stop1 → 1 open

Only 1 step remains and all route/route_via attempts on net 58 have failed; ripping up another net would leave it unrouted with 0 steps left, so stopping is best.

power-awarestep 20route_via net 321 → 1 open

Only 1 step left; try route_via on /COM+ (net 32) with walk-around mode (mode=2) since shove mode failed.

development board (not scored in the tier)

bt-tnc (0492)

Gemini agents: 4 open → 1 openrandom control: 4 open → 3 open55 Gemini calls · 122,753 tokens
board before
before · classic router · 4 open
board after
after · bold · 1 open

Open connections after each step

a rip-up pushes it up first
02040051015stepclean
careful 3bold 1layer-first 2random control

What Gemini decided

straight from llm_calls.jsonl
Leadgemini-4-carbonstrategy note from the board picture

Reroute SCK first; its bottom trace blocks N-0000071, N-0000072, and SPK. Nudge 2-pad nets N-0000014, N-000006, and N-0000059 only where pad clearances require. Leave existing GND and VCC trunks, N-0000024, MOSI, and RESET alone to avoid cascading rip-ups. Use bottom-copper via bridges to tunnel the open GND link under N-0000024, and to route N-0000071 and N-0000072 under top traces once SCK moves. Route SPK with a mid-span layer change: bottom copper under GND, VCC, and MISO, then a via to top copper over MOSI and RESET.

carefulstep 1route net 54 → 4 open

Net 5 (GND) is a very short open connection (2.72 mm) crossed by only one trace; try routing it with shove mode first.

boldstep 1route_via net 54 → 4 open

Net 5 (GND) has a very short 2.72mm open connection crossed only by net 19 on layer 0; routing it with vias should easily bridge over net 19.

layer-firststep 1route_via net 54 → 4 open

Following layer-first style, try route_via on the shortest open connection (net 5, GND) which is only crossed on layer 0.

carefulstep 2route_via net 54 → 4 open

Net 5 (GND) is very short (2.72mm) and direct routing failed, so try routing it with a via.

boldstep 2ripup net 194 → 6 open

Net 5 (GND) is blocked by net 19 (N-0000024), and route_via failed to close it. Ripping up net 19 will clear the path for net 5.

layer-firststep 2route_via net 314 → 3 open

Following layer-first style, try route_via on the next open connection, net 31.

carefulstep 3route net 324 → 3 open

Net 32 is a short open connection (7.04 mm) with only two crossings; try routing it directly with shove mode.

boldstep 3route net 56 → 5 open

Net 19 was ripped up to clear space around net 5 (GND). Now try routing net 5 with shove mode.

layer-firststep 3route_via net 323 → 2 open

Following layer-first style, try route_via on open net 32 before ripping anything.

carefulstep 4route net 313 → 3 open

Net 31 is open and has not been attempted yet; try routing it with shove mode first.

boldstep 4route_via net 195 → 3 open

Net 19 was ripped up to let GND route; now route net 19 with route_via so it can complete both connections and bridge over GND if needed.

layer-firststep 4route_via net 382 → 2 open

Try route_via on net 38 (SPK) first following the layer-first style before ripping any blocking nets.

carefulstep 5route_via net 313 → 3 open

Net 31 failed to route on its current layer, so I will try route_via to see if a via bridge can close it.

boldstep 5route_via net 323 → 2 open

Net 32 is the shortest open connection (7.04 mm) and has not been tried yet; attempting route_via with shove mode.

layer-firststep 5ripup net 372 → 4 open

route_via failed on both GND (5) and SPK (38). Per the lead engineer note, SCK (37) blocks SPK, so rip up SCK to clear the path.

carefulstep 6ripup net 373 → 5 open

SCK (net 37) is blocking N-0000071 and SPK on the bottom layer, and route_via on N-0000071 just failed. Rip up SCK so we can reroute it out of the way.

boldstep 6route_via net 312 → 1 open

Net 31 is open and has not been tried yet; try route_via with shove mode to bridge across the crossing nets.

layer-firststep 6route_via net 374 → 2 open

SCK (net 37) was just ripped up; try routing it with route_via in shove mode.

carefulstep 7route_via net 375 → 3 open

SCK (net 37) was just ripped up; rerouting it with vias to clear both top and bottom layer crossings.

boldstep 7route_via net 381 → 1 open

Only net 38 (SPK) remains open, and it needs a mid-span layer change to cross GND/VCC/MISO on bottom and MOSI/RESET on top.

layer-firststep 7route_via net 382 → 2 open

SCK was rerouted in the previous step; try route_via on SPK (38) with walk-around mode (2) to see if a valid via path now exists.

carefulstep 8route_via net 383 → 3 open

SCK has been rerouted, so let's try routing SPK (net 38) using route_via to bridge across the remaining traces.

boldstep 8ripup net 371 → 3 open

SPK (net 38) failed route_via and is blocked by SCK (net 37) on bottom copper. Rip up SCK first to clear space for SPK.

layer-firststep 8route_via net 52 → 2 open

route_via on net 5 (GND) with mode 1 failed earlier; try mode 2 (walk around) to tunnel under N-0000024.

carefulstep 9route_via net 313 → 3 open

route_via with mode=1 failed for net 31; trying walk-around mode (mode=2) to see if it can find a via bridge without shoving.

boldstep 9route_via net 383 → 3 open

SCK (net 37) is now ripped up, so route SPK (net 38) first using route_via to bridge across GND, VCC, MISO, MOSI, and RESET.

layer-firststep 9ripup net 72 → 3 open

Both route_via modes failed for SPK (38) and GND (5). MISO (7) is a small 2-pad net crossing SPK, so ripping it up may clear the path for SPK.

carefulstep 10ripup net 123 → 4 open

Net 31 (N-0000071) failed route_via in both modes; ripping up 2-pad net 12 (N-0000014) which crosses it may clear the way for a via bridge.

boldstep 10route_via net 373 → 1 open

SCK (net 37) was ripped up to clear space for SPK, and we haven't tried routing SCK yet. Let's route_via SCK first.

layer-firststep 10route_via net 73 → 2 open

Net 7 (MISO) was just ripped up; try route_via on it now to close it using a layer change around SCK.

carefulstep 11route net 124 → 3 open

Net 12 (N-0000014) was just ripped up and has no crossings, so route it on its current layer with shove mode.

boldstep 11ripup net 71 → 2 open

SPK (net 38) failed route_via even with SCK ripped up. Ripping up MISO (net 7, only 2 pads) on layer 0 will clear space for SPK to route.

layer-firststep 11route net 52 → 2 open

route_via failed for GND (net 5) in both modes; try single-layer route with shove (mode 1) to see if it can complete or shove N-0000024.

carefulstep 12route_via net 53 → 3 open

GND (net 5) is a short 2.72mm connection crossed only by N-0000024; try route_via with walk-around mode (mode=2) to tunnel under it.

boldstep 12route_via net 382 → 2 open

MISO (net 7) is currently ripped up, which clears space for SPK (net 38). Let's try routing SPK with vias now.

layer-firststep 12route_via net 52 → 2 open

Net 5 (GND) failed route_via with mode 2 (walk around) and route with mode 1; trying route_via with mode 1 (shove) to place the via bridge under N-0000024.

carefulstep 13route_via net 53 → 3 open

GND (net 5) failed route_via in walk-around mode (mode 2); try shove mode (mode 1) to place a via bridge under N-0000024.

boldstep 13route_via net 72 → 1 open

Net 7 (MISO) was ripped up and is only crossed by net 37 (SCK) on layer 2, so routing it with a via should easily complete it.

layer-firststep 13route net 382 → 2 open

route_via failed for both net 5 and net 38; try route with shove (mode 1) on net 38 (SPK) to see if it can push through.

carefulstep 14route_via net 383 → 3 open

route_via on SPK (38) has only been tried in shove mode (1); try walk-around mode (2) to see if it can find a via bridge.

boldstep 14ripup net 7, 371 → 4 open

SPK (net 38) is still blocked even after ripping up SCK or MISO individually. Ripping up both MISO (7) and SCK (37) at once will clear the top-layer bottleneck so SPK can route with a via.

layer-firststep 14route net 382 → 2 open

route_via and shove-mode route failed for SPK (38), so try walk-around mode (2) on SPK.

carefulstep 15ripup net 293 → 4 open

Net 31 (N-0000071) is blocked by small 2-pad nets including net 29 (N-000006). Ripping up net 29 may allow net 31 to route.

boldstep 15route_via net 384 → 4 open

Both MISO (7) and SCK (37) are ripped up, clearing the path for SPK (38). Route SPK with a via bridge now.

layer-firststep 15route_via net 382 → 2 open

SPK (net 38) failed route_via in walk-around mode (mode 2) and failed single-layer route in both modes; try route_via in shove mode (mode 1).

carefulstep 16route_via net 294 → 3 open

Net 29 (N-000006) was just ripped up and is crossed by net 12 on layer 0, so use route_via to bridge across it.

boldstep 16restore_best4 → 1 open

SPK (net 38) failed to route with mode 1 even with MISO and SCK ripped up. Restoring the best board (1 open connection) so we can try mode 2 on SPK.

layer-firststep 16ripup net 372 → 4 open

SPK (38) and GND (5) are both stuck with route and route_via failing. SCK (37) crosses SPK on layer 0 and is a small 3-pad net, so ripping it up may allow SPK to route.

carefulstep 17route_via net 313 → 3 open

Net 31 (N-0000071) is still open after rerouting net 29, and route_via with shove (mode 1) has not been tried on it yet.

boldstep 17route_via net 381 → 1 open

Only 2 steps left, so ripping up nets is not viable. Try route_via on SPK (net 38) with walk-around mode (mode 2) since shove mode failed.

layer-firststep 17route_via net 374 → 2 open

SCK (net 37) was just ripped up; try routing it with via support to close its two open connections.

carefulstep 18route net 53 → 3 open

With 1 step left and all route_via attempts failing on the 3 open nets, try shoving on GND (net 5) to close its short 2.72mm gap.

boldstep 18route net 381 → 1 open

route_via failed in both modes for net 38 (SPK), so try same-layer routing with shove mode on the final step.

layer-firststep 18route_via net 382 → 2 open

SPK (net 38) failed route_via in mode 1, and we have 1 step left, so try route_via in walk-around mode (mode 2).

development board (not scored in the tier)

nixiebrd (0617)

Gemini agents: 5 open → 3 openrandom control: 5 open → 5 open55 Gemini calls · 214,041 tokens
board before
before · classic router · 5 open
board after
after · bold · 3 open

Open connections after each step

a rip-up pushes it up first
02040051015stepclean
careful 5bold 3layer-first 4random control

What Gemini decided

straight from llm_calls.jsonl
Leadgemini-4-carbonstrategy note from the board picture

Move first: Rip up and reroute 2-pad nets /AN5, /AN3, and /AN7, which block multiple open /AN* connections across both layers. Next, consolidate multi-layer crossings on /K18, /K13, /K14, and /K16 onto the bottom layer (blue) to clear top-layer routing channels. Leave alone: Do not disturb high-fanout nets GND (33 pads), +3V3 (18 pads), Net-(K1-Pad3), /RCLK, or /SCLK. Via bridges: Route /AN2, /AN4, and /AN6 primarily on the sparser top layer (red) over bottom-layer /K17, /K19, /KDP1, and GND, using short bottom-layer via bridges only to duck under /AN1. Route +3V3 on top over /SCLK and Net-(U11-Pad14), bridging onto bottom solely to cross /RCLK.

carefulstep 1route_via net 15 → 5 open

Net 1 (+3V3) has two open connections, including a very short 3.9 mm one; try routing it with vias first.

boldstep 1route_via net 15 → 5 open

Net 1 (+3V3) has two open connections with relatively few crossings; trying route_via with shove mode to close them.

layer-firststep 1route_via net 15 → 5 open

Following the layer-first style, try route_via on net 1 (+3V3) first to see if it can bridge across the blocking traces.

carefulstep 2ripup net 595 → 6 open

Net 1 (+3V3) could not be routed with route_via. Following the lead engineer note to move first on blocking 2-pad nets, I will rip up net 59 (/AN5) which blocks both /AN4 and /AN6 across both layers.

boldstep 2ripup net 59, 57, 615 → 8 open

Rip up 2-pad nets /AN5 (59), /AN3 (57), and /AN7 (61) which block multiple open /AN* connections across both layers.

layer-firststep 2route_via net 565 → 4 open

Layer-first style: try route_via on open net 56 (/AN2) before ripping anything.

carefulstep 3ripup net 57, 616 → 8 open

Rip up 2-pad nets /AN3 (57) and /AN7 (61) as advised to clear multi-layer blockers for the open /AN* connections.

boldstep 3route_via net 608 → 8 open

With /AN5, /AN3, and /AN7 ripped up, /AN6 is only blocked by /AN1 on the top layer and can be routed using a via bridge.

layer-firststep 3route_via net 584 → 4 open

Following the layer-first style, try route_via on open net 58 (/AN4) before ripping up any blocking nets.

carefulstep 4route_via net 618 → 8 open

We ripped up /AN5, /AN3, and /AN7 as advised. Now let's route the shortest of them, /AN7 (net 61), using route_via to bridge across /AN1 on layer 0.

boldstep 4route_via net 618 → 8 open

Try routing /AN7 (net 61), the shortest open /AN* connection, using via bridges now that /AN3 and /AN5 are out of the way.

layer-firststep 4route_via net 604 → 4 open

Try route_via on net 60 (/AN6) before ripping any blocking nets, as it hasn't been tried yet.

carefulstep 5route_via net 598 → 8 open

route_via on net 61 failed, so try routing net 59 (/AN5) with via bridges to clear its open connection.

boldstep 5route_via net 568 → 8 open

Try routing /AN2 (net 56) with via bridges in walk-around mode (mode 2) to see if it can find a path over the bottom-layer crossings.

layer-firststep 5ripup net 59, 61, 574 → 7 open

route_via failed on nets 1, 58, and 60; ripping up 2-pad blocking nets /AN5 (59), /AN7 (61), and /AN3 (57) as advised.

carefulstep 6route_via net 608 → 8 open

/AN5 and /AN7 failed route_via in mode 1; try route_via on /AN6 (net 60) with walk-around mode 2 to duck under /AN1.

boldstep 6ripup net 97, 958 → 16 open

Rip up multi-layer nets /K18 (97) and /K16 (95) to clear top-layer routing channels for the open /AN* connections.

layer-firststep 6route_via net 617 → 7 open

Net 61 (/AN7) was just ripped up and is the shortest open /AN* connection (21.89 mm); try routing it with vias first.

carefulstep 7route net 618 → 8 open

route_via failed on /AN7, /AN5, and /AN6; try routing the shortest open connection /AN7 (net 61) on its current layer with shove mode.

boldstep 7route_via net 5816 → 16 open

After ripping up /K18 and /K16 to clear space, try routing /AN4 (58) using via bridges.

layer-firststep 7route_via net 597 → 7 open

Try route_via on /AN5 (net 59) now that /AN3 and /AN7 are ripped up.

carefulstep 8ripup net 978 → 12 open

Net 97 (/K18) crosses /AN3 and /AN4 on both layers; ripping it up will allow rerouting it onto the bottom layer to clear top-layer channels.

boldstep 8route net 9516 → 13 open

Net 95 (/K16) was ripped up to clear space; let's re-route it using shove mode so it consolidates cleanly before we tackle the /AN* nets.

layer-firststep 8route_via net 577 → 7 open

Try route_via on /AN3 (net 57) before ripping up any crossing nets.

carefulstep 9route_via net 9712 → 8 open

Net 97 (/K18) was just ripped up; try routing it with via bridges to clear its open connections.

boldstep 9restore_best13 → 5 open

We are at 13 open connections with 10 steps left, whereas the best board had only 5 open connections. Restoring the best board to work from a cleaner state.

layer-firststep 9ripup net 97, 92, 937 → 19 open

route_via failed on /AN3, /AN4, /AN5, /AN6, and /AN7. Following the lead engineer note, rip up /K18, /K13, /K14, and /K16 to clear top-layer routing channels.

carefulstep 10route_via net 588 → 8 open

Try routing /AN4 (net 58) with via bridges, as we haven't attempted it yet since ripping up /AN3, /AN5, and /AN7.

boldstep 10route net 15 → 4 open

Net +3V3 (1) has a short 3.9mm open segment crossed only on layer 2 and a 38.63mm segment; try routing it directly with shove mode.

layer-firststep 10route_via net 9219 → 15 open

Net 92 (/K13) was just ripped up and has multiple open connections; let's route it with route_via.

carefulstep 11route_via net 568 → 7 open

Net 56 (/AN2) has not been attempted with route_via yet; trying it now to see if a via bridge can close it.

boldstep 11route_via net 14 → 4 open

Net 1 (+3V3) has only 1 short 3.9mm open connection left, and all its blockers are on layer 2, so route_via should be able to bridge it on layer 0.

layer-firststep 11route_via net 9315 → 14 open

Route /K14 (net 93) with route_via after it was ripped up in step 9.

carefulstep 12route_via net 17 → 7 open

Net 1 (+3V3) has two short open connections that may be solvable with a via bridge, and we haven't tried routing it yet.

boldstep 12route_via net 604 → 4 open

Net 60 (/AN6) is currently open and hasn't been tried with route_via in this board state; trying route_via to bridge over blocking nets.

layer-firststep 12route_via net 114 → 14 open

Net 1 (+3V3) has two open connections that have not been attempted yet; try route_via on it.

carefulstep 13route_via net 577 → 7 open

Net 57 (/AN3) was ripped up in step 1 and has not been routed yet; try route_via with shove mode.

boldstep 13ripup net 594 → 5 open

Ripping up 2-pad net 59 (/AN5) which blocks both /AN4 and /AN6 on both layers.

layer-firststep 13route_via net 9714 → 10 open

We ripped up net 97 (/K18) in step 7 and haven't rerouted it yet; try route_via on net 97.

carefulstep 14ripup net 557 → 8 open

Net 55 (/AN1) is a small 2-pad net on layer 0 that blocks /AN4, /AN5, /AN6, and /AN7; ripping it up will clear top-layer access.

boldstep 14route_via net 595 → 5 open

Net 59 (/AN5) was just ripped up; try routing it with via bridges now.

layer-firststep 14route_via net 6110 → 10 open

Now that /K18 and /K13 have been rerouted, retry route_via on /AN7 (net 61) which is the shortest open /AN* connection.

carefulstep 15route_via net 588 → 8 open

With /AN1 (net 55) ripped up, /AN4 (net 58) has fewer obstacles on the top layer; try route_via on net 58.

boldstep 15route net 595 → 5 open

Net 59 (/AN5) was just ripped up and route_via failed; try direct routing with shove mode.

layer-firststep 15route_via net 6010 → 10 open

Net 60 (/AN6) is open and has not been tried with route_via in the current state; trying it next per layer-first style.

carefulstep 16route_via net 598 → 8 open

Net 59 (/AN5) has not been tried with route_via yet; try routing it with a via bridge to close one of the open /AN* connections.

boldstep 16restore_best5 → 4 open

Ripping up /AN5 (59) left us with 5 open connections and routing it failed; restore_best will return to the 4-open-connection state.

layer-firststep 16restore_best10 → 4 open

Only 3 steps remain and we currently have 10 open connections, worse than the 7 open connections seen earlier. Restore the best board state first.

carefulstep 17restore_best8 → 5 open

We are at 8 open connections while the best state had 7; restoring the best board first.

boldstep 17route_via net 584 → 4 open

With only 2 steps left, ripping up a net would leave it open, so try route_via on open net 58 (/AN4) in the current 4-open state.

layer-firststep 17route_via net 584 → 4 open

Try route_via on /AN4 (net 58) to see if it can bridge over the crossings without ripping anything.

carefulstep 18route_via net 605 → 5 open

With 1 step left after restoring the best state (5 open connections), try route_via on net 60 (/AN6) which is a relatively short 35.15mm open connection.

boldstep 18route_via net 564 → 3 open

Net 56 (/AN2) has not been attempted with route_via yet; try routing it with vias on the last step.

layer-firststep 18route_via net 604 → 4 open

Last step: try route_via on net 60 (/AN6) on the current 4-open board to see if a via bridge can close it.

development board (not scored in the tier)

sweet-sixteen (0658)

Gemini agents: 4 open → 1 openrandom control: 4 open → 2 open55 Gemini calls · 135,538 tokens
board before
before · classic router · 4 open
board after
after · careful · 1 open

Open connections after each step

a rip-up pushes it up first
02040051015stepclean
careful 1bold 1layer-first 1random control

What Gemini decided

straight from llm_calls.jsonl
Leadgemini-4-carbonstrategy note from the board picture

Leave alone: Multi-pad global nets GND, /VLED, and existing /~{OE} segments (though connect /~{OE}'s short unblocked 1.8mm gap immediately). Move first: Reroute local 2-pad nets around the (138, 125) congestion zone—specifically Net-(AFF7-Pad12) (blocking bottom copper) and top-layer nets Net-(AFF6-Pad13) and Net-(R6-Pad2) to clear space for /SDA_56 around /~{OE}. Via bridges: Bottom copper is largely open here. Drop Net-(AFF6-Pad10) to bottom copper to bypass Net-(AFF2-Pad12), Net-(AFF6-Pad12), and Net-(AFF6-Pad9). Route /SDA_56 and /SDA_67 horizontally on bottom copper to cleanly bridge under top-layer GND, /VLED, and AFF6/AFF7 pad nets.

carefulstep 1route net 1564 → 4 open

Net 156 is only 1.83 mm long with no crossings, so routing it directly with shove mode should close it cleanly.

boldstep 1route net 1564 → 4 open

Net 156 is a very short 1.83mm open connection with no crossings, so try routing it directly with shove mode first.

layer-firststep 1route_via net 1564 → 3 open

Following layer-first style, try route_via on the short open connection net 156 first.

carefulstep 2route_via net 1024 → 3 open

Net 102 (Net-(AFF6-Pad10)) is crossed only by top-layer nets; use route_via to bridge it on bottom copper.

boldstep 2route_via net 1564 → 3 open

Net 156 failed to route on its current layer despite no crossings, so it likely needs a layer change or via bridge.

layer-firststep 2route_via net 1023 → 2 open

Net 102 is only crossed on layer 0, so a via bridge to bottom copper should close it cleanly.

carefulstep 3route_via net 1563 → 2 open

Net 156 (/~{OE}) has a short 1.83mm unblocked gap, and since same-layer shove routing failed, trying route_via may complete it by changing layers.

boldstep 3route_via net 1023 → 2 open

Net 102 (Net-(AFF6-Pad10)) is only crossed by top-layer nets, so route_via should be able to drop it to bottom copper cleanly.

layer-firststep 3route_via net 1802 → 1 open

Following layer-first style, try route_via on open connection net 180 (/SDA_56) to bridge under top-layer obstacles.

carefulstep 4route_via net 1802 → 1 open

Try routing /SDA_56 (net 180) with a via bridge to see if it can cleanly bypass the top-layer crossings on bottom copper.

boldstep 4ripup net 116, 1662 → 4 open

Net-(R6-Pad2) (166) blocks both /SDA_56 and /SDA_67, and Net-(AFF7-Pad12) (116) blocks bottom copper for /SDA_67. Ripping up both clears the congestion zone.

layer-firststep 4route_via net 1811 → 1 open

Only net 181 (/SDA_67) remains open, and all its crossings are on layer 0. Try route_via to bridge it on bottom copper.

carefulstep 5route_via net 1811 → 1 open

Net 181 (/SDA_67) is the last open connection and is crossed only by top-layer nets, so a via bridge to bottom copper should complete it.

boldstep 5route_via net 1804 → 3 open

With nets 116 and 166 ripped up to clear the congestion zone, route /SDA_56 (net 180) using via bridges to bypass top-layer GND and /~{OE}.

layer-firststep 5ripup net 1661 → 2 open

route_via on net 181 (/SDA_67) failed, so rip up the crossing 2-pad net 166 (Net-(R6-Pad2)) to clear space for routing.

carefulstep 6ripup net 1661 → 2 open

Net 181 (/SDA_67) failed route_via in 90 trials; rip up the 2-pad net 166 (Net-(R6-Pad2)) that crosses it in the congestion zone to free up space for a via bridge.

boldstep 6route_via net 1813 → 3 open

Route /SDA_67 (net 181) using route_via to bridge under the top-layer nets on bottom copper, as recommended by the lead engineer.

layer-firststep 6route_via net 1662 → 2 open

Net 166 was just ripped up; try routing it with vias to close it cleanly.

carefulstep 7route_via net 1812 → 2 open

Net 166 was just ripped up to clear space around (138, 125). Try routing /SDA_67 (net 181) with via bridges now that the congestion is reduced.

boldstep 7route_via net 1663 → 3 open

route_via on net 181 failed; try route_via on net 166 (Net-(R6-Pad2)) which is shorter and only crossed by nets 156 and 180.

layer-firststep 7ripup net 1162 → 3 open

route_via failed on both open nets 166 and 181. Net 116 (Net-(AFF7-Pad12)) is a 2-pad net blocking bottom copper for net 181 (/SDA_67), so rip up net 116 to clear the path.

carefulstep 8ripup net 1162 → 3 open

Net 116 (Net-(AFF7-Pad12)) is on bottom copper (layer 2) and crosses /SDA_67 (net 181), blocking its via bridge. Rip up net 116 to clear bottom copper for /SDA_67.

boldstep 8route_via net 1163 → 3 open

Net 116 (Net-(AFF7-Pad12)) is open after being ripped up, and we have not tried routing it with vias yet. Try route_via on net 116.

layer-firststep 8route_via net 1813 → 3 open

Net 116 was blocking bottom copper around (138, 125); now that it is ripped up, try route_via on net 181 (/SDA_67) to bridge horizontally on bottom copper.

carefulstep 9route_via net 1813 → 3 open

Net 116 was ripped up to clear bottom copper; try route_via on net 181 (/SDA_67) now.

boldstep 9route_via net 1813 → 3 open

route_via with shove (mode 1) failed for net 181 (/SDA_67); try walk-around (mode 2) to see if it can route around the pad congestion and drop to bottom copper.

layer-firststep 9route_via net 1163 → 3 open

Net 116 (Net-(AFF7-Pad12)) is open and hasn't been tried with route_via yet; try routing it with a via bridge.

carefulstep 10restore_best3 → 1 open

Ripping up nets 166 and 116 did not allow net 181 to route via, so restore_best to return to the 1-open-connection state.

boldstep 10ripup net 1803 → 4 open

Net 180 (/SDA_56) is a 2-pad net crossing both open nets 166 and 181. Ripping it up will clear space to route 166 and 181.

layer-firststep 10route_via net 1663 → 3 open

Net 116 was ripped up in step 7, which may have cleared space near (138, 125) for Net 166. Trying route_via on Net 166.

carefulstep 11route_via net 1811 → 1 open

Net 181 (/SDA_67) is the only open connection left, and route_via with mode=1 failed. Try route_via with mode=2 (walk around) to see if it can find a valid via bridge.

boldstep 11route_via net 1814 → 4 open

Now that /SDA_56 (180) is ripped up, try routing /SDA_67 (181) with via bridges.

layer-firststep 11route_via net 1813 → 3 open

route_via with mode 1 failed on 181, so try route_via with walk-around (mode 2) to see if it can find a path on bottom copper.

carefulstep 12ripup net 1801 → 2 open

Net 181 (/SDA_67) failed route_via in both modes; ripping up 2-pad net 180 (/SDA_56) which crosses it and was recently routed may free the needed via site or path.

boldstep 12restore_best4 → 2 open

We are at 4 open connections and only have 7 steps left, whereas our best board had only 2 open connections. Restoring the best board gives us a much better starting point.

layer-firststep 12ripup net 1333 → 4 open

route_via failed on all three open nets. Rip up Net-(AFF8-Pad12) (133), a 2-pad net crossing Net-(AFF7-Pad12) (116) twice on both layers, to clear a path.

carefulstep 13restore_best2 → 1 open

Restore the best board state where /SDA_56 (net 180) is already routed, leaving only /SDA_67 (net 181) open.

boldstep 13ripup net 1162 → 3 open

Net 116 (Net-(AFF7-Pad12)) is on bottom copper blocking /SDA_67 (net 181). Ripping up net 116 alone gives us a chance to route 181 and 180 within the remaining 6 steps.

layer-firststep 13route_via net 1814 → 4 open

We just ripped up net 133, which may have cleared space for /SDA_67 (net 181). Let's try route_via on net 181.

carefulstep 14route net 1811 → 1 open

Net 181 failed route_via in both modes; try routing it on its current layer with shove mode.

boldstep 14ripup net 1663 → 4 open

Net 166 (Net-(R6-Pad2)) crosses both /SDA_56 (180) and /SDA_67 (181) on the top layer; ripping it up will clear space for both.

layer-firststep 14restore_best4 → 1 open

We went from 1 open connection to 4 open connections and route_via is still failing on net 181. Let's restore the best board (1 open connection) and try a different approach.

carefulstep 15ripup net 1001 → 2 open

Net 181 (/SDA_67) is still open and route/route_via failed; rip up Net-(AFF6-Pad7) (100) which crosses it to see if it frees a path or via site.

boldstep 15route_via net 1814 → 4 open

With 116 and 166 ripped up, try routing /SDA_67 (181) using route_via.

layer-firststep 15route_via net 1811 → 1 open

Net 181 (/SDA_67) is the only open connection and all its crossings are on layer 0; try route_via to bridge under them on bottom copper.

carefulstep 16route_via net 1812 → 2 open

Net 100 was just ripped up; try routing net 181 with vias now that space may be cleared.

boldstep 16route_via net 1804 → 3 open

With nets 116 and 166 ripped up, try routing /SDA_56 (net 180) using via bridges.

layer-firststep 16ripup net 1661 → 2 open

Net 181 (/SDA_67) failed route_via with mode 1. Rip up the local 2-pad blocker net 166 (Net-(R6-Pad2)) near the congestion zone so we can route 181 and then re-route 166 in the remaining steps.

carefulstep 17route net 1002 → 1 open

Net 100 has no crossings and was ripped up in step 9; routing it now should close it.

boldstep 17restore_best3 → 2 open

With only 2 steps left and all 3 open nets failing route_via, restoring the best board (2 open connections) is the optimal move.

layer-firststep 17route_via net 1812 → 2 open

Net 166 was just ripped up to clear space; now try route_via on net 181 (/SDA_67) in shove mode.

carefulstep 18route net 1811 → 1 open

Only 1 step left; try routing net 181 with mode 2 (walk around) since mode 1 failed.

boldstep 18route_via net 1802 → 1 open

Net 180 (/SDA_56) was successfully routed with route_via in step 9, so let's route it now to reduce open connections from 2 to 1.

layer-firststep 18restore_best2 → 1 open

With only 1 step left and 2 open connections, restoring the best board (1 open connection, 0 errors) is the best action.

Proof against a baseline

Did the agents matter?

Same actions, same step budget, random choices. Scored by PCBWorld's evaluate_one.

1/7boards brought to a clean pass by the agents
1/7by the random control
39/79open connections closed by the agents
15closed by the random control
BoardOpen at startAgents: endsteps best / allGemini callstokensRandom: endsteps best / all
bt-tnc (0492) dev41 open18 / 5455123k3 open18 / 54
nixiebrd (0617) dev53 open18 / 5455214k5 open– / 54
sweet-sixteen (0658) dev41 open18 / 5455136k2 open18 / 54
Chickadee (0400) hard test10 · clean5 / 171831k0 · clean13 / 42
data-manager (0473) hard test43 open20 / 8081233k3 open20 / 25
RGBMatrixPanel CPLD (0599) hard test6031 open20 / 8081302k50 open20 / 60
HybridChargeController (0626) hard test11 open– / 8081196k1 open– / 49

The hard boards nothing had fixed, every method, one scorer

Cell = share of connections routed · error-level rule violations left.

BoardFreeroutingScripted repairWhole-plan agentsManaged Agent specialistGemini step agents
Chickadee (0400)98.7% · 1 err98.7% · 1 errno boardclean passclean pass
data-manager (0473)95.7% · 4 err95.7% · 4 errno board97.8% · 2 err96.7% · 3 err
HybridChargeController (0626)99.5% · 1 err99.5% · 1 errno board99.5% · 1 errno board
RGBMatrixPanel CPLD (0599)57.4% · 60 errno boardno boardnot run78.0% · 31 err
clean boards0 of 40 of 40 of 41 of 41 of 4
Live

Route a board yourself

routr studio

Shaurya's live viewer: pick a board, press Route, watch the solver's versions arrive and KiCad's rule check judge each one. Replays play recorded runs at their original pace.

The studio is not running. Start it with web/serve.sh (or pcb/run_studio.sh 8766) and reload this page.

What still fails

Part 2 · Analog sizing

Size an op-amp that passes at every corner.

A SKY130 two-stage Miller op-amp, simulated in ngspice at 15 process, voltage and temperature points (5 corners × 3 temperatures). A design passes only if gain, bandwidth, phase margin, current and every transistor's saturation hold at all 15.

7/7routr runs that pass, within 40 candidates
1/18Bayesian-optimisation and random runs that pass, same budget
46/49device roles read right from a schematic photo (43/49 from the netlist alone)
15/15PVT points an independent sign-off agent matched
How the crew works

Read the circuit, hear the spec, size it, check it

Read the schematic

Decisions API

A photo of the schematic plus the netlist: every transistor gets a role (input pair, load, tail…) with a confidence.

Hear the spec

gemini-4-hydrogen

A spoken or typed spec becomes targets: gain, bandwidth, phase margin, current.

Architect

gemini-4-carbon

Reads the worst PVT failures and the Bode plot, then sends each worker a direction.

Four workers

gemini-4-hydrogen

Propose sizings in parallel. Every candidate is 15 ngspice decks.

Judge, then sign-off

ngspice · Managed Agent

Our judge scores every candidate; an independent agent re-simulates the winner from scratch.

pipeline diagram
Proof against a baseline

The crew against Bayesian optimisation

Same circuit, same 40-candidate budget (600 simulations), three seeds each. Optuna TPE and GP-BO are the standard black-box optimisers for this job.

Share of runs that pass, both specs

routr7/7 runsOptuna TPE0/6 runsOptuna GP-BO1/6 runsRandom search0/6 runs
MethodSpecRuns that passFirst pass at candidateMedianBest core current (µA)
routrBare spec4/410, 6, 6, 4619.2
Optuna TPEBare spec0/3–, –, –>40–
Optuna GP-BOBare spec1/3–, 39, –>4068.9
Random searchBare spec0/3–, –, –>40–
routrGuard-banded spec3/327, 18, 181833.5
Optuna TPEGuard-banded spec0/3–, –, –>40–
Optuna GP-BOGuard-banded spec0/3–, –, –>40–
Random searchGuard-banded spec0/3–, –, –>40–
Real runs

Every candidate the crew tried

Pick a run. Each dot is one sizing, simulated at all 15 points; hover for its worst case and why it failed.

passes all 15 PVT pointsfirst pass at candidate 27best 33.5 µA core current21 Gemini callsfull-library recheck: pass

Worst-case phase margin, every candidate

worst of 15 PVT points · hover a dot
02040608010203040candidatetarget 65°
passes every spec at all 15 pointsfails somewhere

Core current (µA)

line = lowest passing so far
05010010203040candidate

What the architect saw and asked for

architect.jsonl
Architectgemini-4-carbonround 0

The baseline fails all four specs plus tail/sink saturation. First, UGF is 9x higher than needed (101 MHz vs 11 MHz) because Cc=0.25 pF is far too small and I_tail=40 uA is excessive, collapsing phase margin to 27 deg. Second, DC gain is 54.6 dB (< 63 dB) due to short channel lengths (L_in=0.3 um, L_load=L_drv=0.7 um) and high bias currents. Third, i_core=107.6 uA exceeds the 90 uA limit.

→ Fixed-Ibias W-scaled low-power re-center: Keep Ibias_uA near 18-20 uA so the frozen wide-swing cascode bias network operates at its intended current density (fixing sink cascode margin), and cut i_core below 70 uA by scaling W_tail_um down to 1.5-2.2 um and W_sink_um to 2.5-3.2…

→ Reduced-Ibias moderate-sizing balanced design: Lower Ibias_uA to 10-14 uA while keeping W_tail_um around 2.5-3.5 um and W_sink_um around 3.5-4.5 um (targeting 50-75 uA total core current) to test how lower reference Vov affects sink and tail cascode margins across corners.

→ Bold ultra-low-current aggressive optimization: Aggressively target 30-45 uA core current right away while meeting the 11 MHz UGF floor.

→ High-gain wide-margin conservative anchor: Prioritize passing every PVT spec with wide safety margins at ~75-85 uA core current. Keep Ibias_uA = 20 uA (exact match to the frozen bias network), set W_tail_um = 2.0-2.5 um and W_sink_um = 3.5-4.0 um.

passes all 15 PVT pointsfirst pass at candidate 18best 41.6 µA core current25 Gemini callsfull-library recheck: pass

Worst-case phase margin, every candidate

worst of 15 PVT points · hover a dot
2040608010203040candidatetarget 65°
passes every spec at all 15 pointsfails somewhere

Core current (µA)

line = lowest passing so far
05010010203040candidate

What the architect saw and asked for

architect.jsonl
Architectgemini-4-carbonround 0

The baseline sizing fails four requirements simultaneously due to uncompensated high Stage-1 transconductance, short channel lengths, excessive bias mirror ratios, and crushed cascode headroom. First, UGF is 9x over-designed (101 MHz vs 11 MHz spec) because Cc (0.25 pF) is half of CL (0.5 pF) while Stage-1 gm is large at I_tail=40 uA, pushing UGF close to the Stage-2 output pole and collapsing phase margin to 27.1 deg at ff/125. Second, DC gain reaches only 54.6 dB at fs/125 (needs >= 63 dB) because L_in=0.3 um and L_load=L_drv=0.7 um suffer high output conductance at high current density.

→ Near-nominal Ibias with mirror downscaling and Miller/L boost: Keep Ibias_uA near the frozen bias network's 20 uA design point (18 to 22 uA) and cut worst-case i_core into the 62 to 78 uA range by scaling down W_tail_um to 1.6-2.4 um and W_sink_um to 2.5-3.4 um.

→ Lowered Ibias exploration to shrink cascode Vdsat: Test whether reducing Ibias_uA to 10-15 uA increases sink and tail bottom-cascode saturation margins at fs/-40 and fs/125 by lowering Vdsat across the frozen bias mirror.

→ Bold low-power high-gm/Id and high sink-to-tail ratio: Push aggressively into a low-current regime (target worst-case i_core 38-55 uA) by starving Stage 1 relative to Stage 2: use Ibias_uA=12-17 uA, W_tail_um=1.0-1.8 um (I_tail ~6-12 uA), and W_sink_um=2.4-3.6 um (I_sink ~22-36 uA).

→ High-Ibias cascode-headroom insurance and heavy compensation: Cover the opposite hypothesis on the frozen bias network: if its cascode gate generator relies on a resistor or high-current drop that under-biases sink Vds at -40 C when Ibias drops, test slightly higher Ibias_uA=21-27 uA while sharply…

passes all 15 PVT pointsfirst pass at candidate 18best 39.8 µA core current21 Gemini callsfull-library recheck: pass

Worst-case phase margin, every candidate

worst of 15 PVT points · hover a dot
2040608010203040candidatetarget 65°
passes every spec at all 15 pointsfails somewhere

Core current (µA)

line = lowest passing so far
05010010203040candidate

What the architect saw and asked for

architect.jsonl
Architectgemini-4-hydrogenround 0

The baseline design fails four specs due to under-compensation, short channel lengths, excessive bias current, and insufficient headroom on the wide-swing tail/sink cascodes. Specifically: (1) PM is 27.1 deg (vs >=65 deg) while UGF is 101.2 MHz (vs >=11 MHz) because Cc=0.25 pF is far too small relative to CL=500 fF and Rz=3 kohm does not adequately place the zero to cancel the non-dominant pole; we have ~9x UGF headroom to trade for PM. (2) DC gain is 54.6 dB at fs/125 (vs >=63 dB) because L_in=0.3 um and L_load=L_drv=0.7 um give low output impedances ro1 and ro2.

→ Balanced Headroom, Gain, and Miller Compensation: Reduce Ibias_uA to 13-16 uA to cut worst-case i_core safely below 90 uA and lower tail/sink Vdsat for cascode margin. Increase W_in_um to 15-30 um (keeping L_in_um at 0.3-0.4 um) to lower Vgs_in and lift the tail node out of triode.

→ Bold Long-Channel High-Gain and Low-Power Scaling: Make bold moves toward minimum core current and high intrinsic gain: push L_in_um to 0.5-0.8 um with W_in_um=30-60 um to simultaneously raise ro1 and reduce Vgs_in, and increase L_load_um and L_drv_um to 1.6-2.5 um with W_load_um=35-65 um…

→ Asymmetric Current Split Favoring Stage-2 Bandwidth: Explore shifting current budget from Stage 1 to Stage 2 so gm2 stays high (pushing the non-dominant pole higher for PM) while total i_core drops below 90 uA: reduce W_tail_um to 2.0-3.5 um while keeping W_sink_um at 4.0-5.5 um and setting…

→ Careful Ratio-Preserving Refinement Around Baseline: Stay close to the baseline sizing and make minimal, controlled adjustments to clear every violation cleanly: nudge Ibias_uA down just to 15-17 uA (targeting worst i_core around 75-85 uA and gently improving sink saturation above 10 mV),…

passes all 15 PVT pointsfirst pass at candidate 10best 21.1 µA core current25 Gemini calls

Worst-case phase margin, every candidate

worst of 15 PVT points · hover a dot
2040608010203040candidate
passes every spec at all 15 pointsfails somewhere

Core current (µA)

line = lowest passing so far
05010010203040candidate

What the architect saw and asked for

architect.jsonl
Architectgemini-4-carbonround 0

The baseline sizing fails all specs except UGF (which has a massive 10x excess at 101 MHz vs 10 MHz required). First, Phase Margin fails (27.1 deg at ff/125) because Cc=0.25 pF is far too small relative to Stage-1 gm, pushing UGF near the Stage-2 non-dominant pole. Second, DC gain falls short (54.6 dB at fs/125) due to short channel lengths (L_in=0.3 um, L_load=L_drv=0.7 um) and high bias currents degrading output resistance. Third, core current exceeds 100 uA (107.6 uA at ff/125) since Ibias=20 uA with W_tail=W_sink=5 um yields ~100 uA nominal before FF corner variation.

→ Balanced spec-compliance via mirror scaling at nominal Ibias: Keep Ibias_uA near 16-20 uA so the frozen wide-swing cascode bias operates near its design point, and cut i_core to ~55-75 uA by reducing W_tail_um to 2.0-3.0 um and W_sink_um to 3.0-3.8 um.

→ Ibias reduction with long-channel gain boost: Test lowering Ibias_uA directly to 9-13 uA (keeping W_tail_um in 3.5-5.0 um and W_sink_um in 4.0-5.0 um) to drop worst-case core current to 45-65 uA;

→ Low Stage-1 / strong Stage-2 asymmetric split: Skew current toward Stage 2 for high output transconductance and a high non-dominant pole while starving Stage 1 to lower UGF and save power: keep Ibias_uA at 16-20 uA, drop W_tail_um to 1.2-2.0 um, and set W_sink_um to 3.0-4.0 um (total…

→ Bold ultra-low-power sizing near 25-40 uA: Aggressively minimize core current right away toward 25-40 uA while maintaining UGF >= 10 MHz across PVT. Set Ibias_uA = 12-16 uA with W_tail_um = 1.0-1.8 um and W_sink_um = 1.6-2.5 um.

passes all 15 PVT pointsfirst pass at candidate 6best 19.2 µA core current21 Gemini callsfull-library recheck: pass

Worst-case phase margin, every candidate

worst of 15 PVT points · hover a dot
20406010203040candidate
passes every spec at all 15 pointsfails somewhere

Core current (µA)

line = lowest passing so far
05010010203040candidate

What the architect saw and asked for

architect.jsonl
Architectgemini-4-carbonround 0

The starting design has 10x excess UGF (101 MHz vs 10 MHz spec) because stage-1 gm is too high and Cc (0.25 pF) is too small, causing severe phase margin failure (27.1 deg at ff/125) as UGF approaches the non-dominant poles. Short channel lengths (L_in=0.3 um, L_load=L_drv=0.7 um) at high current yield insufficient output resistance, failing DC gain (54.6 dB at fs/125 vs 60 dB). Core current (107.6 uA at ff/125) exceeds the 100 uA limit.

→ Conservative All-PVT Feasible Baseline: Target a rock-solid passing design around 45-65 uA core current first. Lower Ibias_uA to 12-16 uA and set W_tail_um to 2.0-3.0 um and W_sink_um to 2.5-3.5 um to bring worst-case i_core safely below 65 uA.

→ Bold Ultra-Low-Power Weak-Inversion Push: Aggressively minimize core current toward 18-30 uA by exploiting the 10 MHz UGF floor with CL=500 fF. Cut Ibias_uA to 6-10 uA with W_tail_um = 1.0-1.8 um and W_sink_um = 1.8-2.8 um so I_tail is ~4-7 uA and I_sink is ~12-20 uA.

→ Nominal-Bias Mirror-Scaled Architecture: Keep Ibias_uA near the frozen bias network design point (16-20 uA) so wide-swing cascode gate biasing tracks its 20 uA nominal target, and instead reduce core current to 30-50 uA by scaling down W_tail_um to 1.0-1.8 um and W_sink_um to…

→ Stage-Ratio and LHP-Zero Compensation Optimization: Optimize the stage-2 to stage-1 gm ratio and Rz zero placement for >70 deg PM while holding i_core around 35-55 uA.

passes all 15 PVT pointsfirst pass at candidate 6best 23.5 µA core current21 Gemini callsfull-library recheck: pass

Worst-case phase margin, every candidate

worst of 15 PVT points · hover a dot
2040608010203040candidate
passes every spec at all 15 pointsfails somewhere

Core current (µA)

line = lowest passing so far
05010010203040candidate

What the architect saw and asked for

architect.jsonl
Architectgemini-4-carbonround 0

The baseline design fails four constraints due to excessive stage-1 bandwidth and over-currenting: (1) PM is 27.1 deg at ff/125 because Cc=0.25 pF yields an excessive 101 MHz UGF (10x the 10 MHz spec) that approaches the stage-2 non-dominant pole; (2) DC gain is 54.6 dB at fs/125 (<60 dB) due to short channels (L_in=0.3 um, L_load=L_drv=0.7 um) and high channel currents degrading ro;

→ Balanced All-Spec Feasibility: Keep Ibias_uA near its nominal 16-20 uA design point for predictable cascode gate biasing, and scale W_tail_um down to 2.0-3.0 um and W_sink_um to 2.5-3.5 um to drop worst-case i_core to 50-70 uA.

→ Bold Ultra-Low-Current Push: Exploit the 10x UGF margin (only 10 MHz needed into 500 fF) by targeting 20-35 uA total core current immediately.

→ High Stage-2 Ratio for Rock-Solid PM: Create a high gm_drv/gm_in ratio by making stage-1 current much smaller than stage-2 current: keep Ibias_uA at 15-20 uA, drop W_tail_um to 1.2-2.0 um (I_tail ~ 10-16 uA), and set W_sink_um to 3.0-4.0 um (I_sink ~ 30-48 uA) for ~50-65 uA…

→ Long-Channel High-Gain & Cascode Headroom: Focus on maximizing worst-case DC gain (>68 dB) and saturation margins across extreme corners while holding Ibias_uA strictly at 18-20 uA so the frozen cascode bias operates at its exact design current.

passes all 15 PVT pointsfirst pass at candidate 4best 19.7 µA core current21 Gemini callsfull-library recheck: pass

Worst-case phase margin, every candidate

worst of 15 PVT points · hover a dot
20406010203040candidate
passes every spec at all 15 pointsfails somewhere

Core current (µA)

line = lowest passing so far
05010010203040candidate

What the architect saw and asked for

architect.jsonl
Architectround 0

The baseline design fails four constraints due to excessive Stage-1 transconductance, short channel lengths, and tight diff-pair tail headroom. First, UGF is 101 MHz (10x the 10 MHz spec) with Cc = 0.25 pF, pushing unity-gain well into the non-dominant poles and collapsing phase margin to 27.1 deg at ff/125. Second, DC gain is only 54.6 dB at fs/125 (spec >= 60 dB) because L_in = 0.3 um and L_load = L_drv = 0.7 um yield low output resistance at high bias current. Third, i_core hits 107.6 uA at ff/125 (spec <= 100 uA) because nominal I_tail (40 uA) + I_sink (60 uA) already equals 100 uA at TT.

→ Balanced First-Pass Spec Closure: Make measured adjustments to pass all specs comfortably around 55-70 uA core current. Lower Ibias_uA slightly to 15-18 uA and trim W_tail_um to 2.5-3.5 um and W_sink_um to 3.0-4.0 um.

→ Keep Frozen Bias at 20 uA, Scale Mirrors Only: Keep Ibias_uA near its design center (18-20 uA) so the frozen wide-swing cascode bias network maintains nominal gate-bias tracking for sink (currently 6 mV margin) and tail across corners, and cut core current solely by narrowing W_tail_um…

→ Asymmetric Stage-1/Stage-2 Split for High PM: Exploit Miller pole-splitting by starving Stage 1 relative to Stage 2: set Ibias_uA = 14-17 uA, W_tail_um = 1.0-1.8 um (very low gm_in and low I_tail), and W_sink_um = 2.8-3.8 um (high gm_drv to keep the output pole high).

→ Bold Low-Current Push Toward Objective: Aggressively target 25-40 uA worst-case core current right away while maintaining UGF >= 10 MHz across all 15 corners. Reduce Ibias_uA to 8-12 uA, with W_tail_um = 1.5-2.5 um and W_sink_um = 2.0-3.0 um;

The winning design at every PVT point

Corner-40 °C27 °C125 °C
tt93° · 72 dB · 15.4 MHz · 23 µA88° · 73 dB · 15.4 MHz · 26 µA85° · 73 dB · 13.7 MHz · 30 µA
ff89° · 73 dB · 22.0 MHz · 26 µA84° · 73 dB · 20.2 MHz · 30 µA83° · 71 dB · 16.7 MHz · 34 µA
ss92° · 72 dB · 11.5 MHz · 20 µA88° · 73 dB · 12.1 MHz · 24 µA85° · 73 dB · 11.4 MHz · 27 µA
fs89° · 73 dB · 12.3 MHz · 21 µA85° · 73 dB · 12.7 MHz · 25 µA83° · 68 dB · 12.0 MHz · 30 µA
sf95° · 72 dB · 20.0 MHz · 25 µA90° · 73 dB · 19.0 MHz · 28 µA87° · 73 dB · 15.8 MHz · 31 µA

guard_s0, candidate #35: phase margin · DC gain · unity-gain frequency · core current. Green = passes the guard-banded target at that point.

Corner-40 °C27 °C125 °C
tt76° · 74 dB · 14.5 MHz · 29 µA70° · 75 dB · 13.9 MHz · 34 µA67° · 74 dB · 12.4 MHz · 38 µA
ff74° · 75 dB · 16.9 MHz · 33 µA69° · 75 dB · 15.7 MHz · 37 µA66° · 72 dB · 13.7 MHz · 42 µA
ss76° · 74 dB · 12.4 MHz · 26 µA71° · 74 dB · 12.2 MHz · 30 µA68° · 74 dB · 11.2 MHz · 35 µA
fs71° · 75 dB · 12.8 MHz · 27 µA66° · 75 dB · 12.5 MHz · 32 µA65° · 68 dB · 11.3 MHz · 37 µA
sf80° · 74 dB · 16.2 MHz · 31 µA74° · 74 dB · 15.2 MHz · 35 µA70° · 74 dB · 13.3 MHz · 39 µA

guard_s1, candidate #37: phase margin · DC gain · unity-gain frequency · core current. Green = passes the guard-banded target at that point.

Corner-40 °C27 °C125 °C
tt80° · 72 dB · 19.7 MHz · 27 µA74° · 73 dB · 18.8 MHz · 32 µA71° · 72 dB · 16.7 MHz · 36 µA
ff78° · 73 dB · 23.9 MHz · 31 µA73° · 73 dB · 21.9 MHz · 35 µA69° · 71 dB · 18.9 MHz · 40 µA
ss80° · 72 dB · 16.3 MHz · 24 µA75° · 72 dB · 16.1 MHz · 28 µA71° · 72 dB · 14.8 MHz · 33 µA
fs74° · 72 dB · 16.8 MHz · 25 µA69° · 72 dB · 16.5 MHz · 30 µA67° · 70 dB · 15.3 MHz · 35 µA
sf85° · 72 dB · 23.0 MHz · 29 µA78° · 72 dB · 21.3 MHz · 33 µA74° · 72 dB · 18.3 MHz · 37 µA

guard_s2, candidate #39: phase margin · DC gain · unity-gain frequency · core current. Green = passes the guard-banded target at that point.

Views

What the crew looks at

Click any figure to enlarge it.

Schematic
SchematicDevices coloured by the role the intake model gave them.
Bode plot
Bode plotOne curve per PVT point for the passing design.
Saturation margins
Saturation marginsEvery device, every PVT point.
Search trajectory
Search trajectoryWorst-case metrics against candidate, guard_s0.
Search trajectory
Search trajectoryWorst-case metrics against candidate, first40.
Layout plan
Layout planThe crew's best floorplan (not routed, no DRC).
Naive layout
Naive layoutThe starting floorplan the plan is measured against.
More evidence

Beyond the one op-amp

AMS-SizingBench: five circuits we did not write

Budget 100 simulations at tt/27 °C, three seeds. Cell = runs that pass · simulations to the first pass in each.

CircuitroutrOptuna GP-BOOptuna TPERandom
Active zero-comp. op-amp azc3/3 · 2, 2, 33/3 · 3, 3, 53/3 · 3, 3, 53/3 · 3, 3, 5
Fully-diff. gain boost fdgb3/3 · 5, 2, 23/3 · 5, 13, 103/3 · 5, 14, 103/3 · 5, 11, 10
Indirect AC op-amp iac2/3 · 2, –, 531/3 · –, –, 181/3 · –, –, 152/3 · –, 11, 81
Nested Miller op-amp smc3/3 · 27, 29, 373/3 · 4, 12, 173/3 · 4, 26, 111/3 · 4, –, –
Telescopic OTA telescopic_ota3/3 · 27, 25, 203/3 · 11, 11, 13/3 · 12, 13, 13/3 · 16, 34, 1
all five14/1513/1513/1512/15

Reading the circuit: photo against netlist

Device roles right: 46/49 with the schematic photo, 43/49 from the netlist alone.

CircuitInputDevice roles rightCircuit identified asLatency
diffschematic photo + netlist9/9diff · 0.921.6 s
diffnetlist only9/9diff · 0.920.3 s
fcschematic photo + netlist16/19fc · 0.991.8 s
fcnetlist only15/19fc · 0.990.6 s
teleschematic photo + netlist21/21tele · 0.992.1 s
telenetlist only19/21tele · 0.990.5 s

From a request to a sized circuit

A sentence or a hand sketch in; a netlist, a sizing and an explanation out.

TestRequestCrew: result · best FoMRandom: result · best FoM
a_diffdiff_pair_mirror_load · from textpass · 18.3pass · 16.7
b_twostagetwo_stage_miller · from textpass · 23.2pass · 19.6
b_twostage_pmtwo_stage_miller · from textpass · 10.9pass · 10.5
c_imagediff_pair_mirror_load · from a sketchpass · 18.8pass · 17.1
hand sketch of a five-transistor OTA
input sketch

Layout planning

Finger counts, matching patterns, dummies and rows for the winning sizing; estimated parasitics fed back into the simulation.

MethodArea µm²Area cutOffset µVCostFeasibleFirst feasible
routr56239.3%0.000.60712/161
Optuna TPE67726.8%2.780.7594/1611
Random69025.5%92.461.6702/1612

Against a naive layout of 926 µm² and 23.3 µV offset. layout planning with estimated parasitics — not routed, no DRC/LVS

Best cost so far

lower is better · no feasible plan yet drawn at 2
0.511.5251015candidate
routrOptuna TPERandom
layout before and after
naive → crew plan
Checks

Who checks the judge?

Independent sign-off

15/15 points matchedverdicts agreemax gain diff 0.0001 dBmax PM diff 0.015°2 SPICE calls · 628 s
Managed Agent verdict, written on its own simulator
SIGN-OFF: The two-stage Miller-compensated op-amp passes both the required specification and all guard-banded design targets across all 15 SKY130 PVT corners (-40, 27, 125 C across tt, ff, ss, sf, fs) in unity-gain follower configuration (VCM=0.9 V, Cload=500 fF). Worst-case loop metrics remain comfortably within guard-banded targets: DC loop gain >= 67.79 dB (fs/125 C vs 63 dB target), UGF >= 11.40 MHz (ss/125 C vs 11.0 MHz target), phase margin >= 82.69 deg (fs/125 C vs 65 deg target) with gain margin >= 10.31 dB (ff/-40 C), and core current <= 33.52 uA (ff/125 C vs 90 uA target). All nine signal-path MOSFETs stay in saturation above guard-banded targets across PVT, with tightest margins on tail bottom device xnm3 (11.20 mV vs 10 mV target at fs/-40 C), sink bottom device xnm4 (15.72 mV vs 10 mV target at fs/-40 C), and active load xpm1 (86.20 mV vs 60 mV target at fs/125 C).
All 15 points, ours vs the agent
PointGain oursGain agentPM oursPM agentUGF ours MHzUGF agent MHz
tt/-4072.45972.45992.8292.8315.42815.422
tt/2773.17973.17988.0788.0815.38015.374
tt/12572.76872.76884.9884.9813.71113.708
ff/-4072.99572.99588.5688.5721.98721.985
ff/2773.40273.40284.3784.3820.23820.233
ff/12571.11871.11883.2483.2516.74316.740
ss/-4071.77571.77591.8991.8911.54211.541
ss/2772.75772.75788.2988.2912.08312.081
ss/12573.12173.12185.2585.2611.40011.398
sf/-4072.04772.04794.6694.6619.97719.976
sf/2772.91372.91489.7689.7618.98618.984
sf/12573.38473.38487.0987.0915.83315.833
fs/-4072.54872.54888.6188.6212.34812.344
fs/2772.90372.90384.6384.6412.73012.727
fs/12567.78967.78982.6982.6911.96211.961

Red team: an agent paid to break our judge

A Managed Agent was told to break our judge. 20 turns, 37 judge calls, 6/6 findings confirmed with a reproducer.

The judge is currently untrustworthy against automated optimization: we reproduced 15-corner guarded PASS verdicts on both a closed-loop unstable op-amp with negative gain margin (-0.095 dB) and a power-gamed design that cuts reported core current by 14.9% while inflating true supply current to 323.4 uA and driving the bias network into triode across 14 of 15 corners. Surrounding optimizer and parser bugs compound this by quantizing Cc to whole picofarads during Swarm deduplication (dropping passing candidates and risking infinite LLM loops), passing NaN into SPICE decks whose crashed runs outscore valid simulations in Swarm.score(), and granting 0 mV nominal saturation margin to output cascodes. The single highest-priority fix is hardening sizing/opamp2stg.py's AC stability and operating-point checks—measuring UGF at CROSS=LAST, enforcing gain_margin_db >= 6 dB (10 dB guarded), rejecting non-empty bias_triode, and constraining/scoring total supply current i_total_a.
Honest failures

Where it did not work

Two circuits held out from development. Neither the crew nor TPE found a passing design; the binding problem is keeping the PMOS cascode in saturation.

CircuitMethodRuns that passBest core current (µA)Candidates passing nominal spec
Folded cascodecrew v10/372, 85, 680
Folded cascodecrew v2 (re-plan + plot critic)0/367, 72, 870
Folded cascodeOptuna TPE0/3––
Telescopiccrew v10/345, 60, 550
Telescopiccrew v2 (re-plan + plot critic)0/378, 54, 760
TelescopicOptuna TPE0/3––
Architect, v2_fc_s0, final diagnosis
Candidate #1 fails on four coupled fronts: (1) Stage-1 PMOS headroom collapse at fs/125 (pcas -177 mV, psrc -52 mV) because W_drv=20 um (m=3) at L_drv=0.7 um makes Vsg_drv too small to accommodate both Vsd_psrc and Vsd_pcas, pushing V_out1 (gate of drv) too close to VDD and driving pcas and psrc into triode; (2) this triode collapse destroys stage-1 output resistance, dropping worst-case DC gain to 44.4 dB (spec >= 63 dB); (3) tail is -38 mV at fs/125 because W_in=5 um gives too large a Vgs_in, pulling the tail drain too low, while nsink (9 mV) and sink (6 mV) sit just under the 10 mV wide-swing margin at -40 C; (4) i_core is 151.2 uA (spec <= 90 uA) and PM is only 27.8 deg at fs/27 because UGF is over-driven to 45.5 MHz (spec >= 11 MHz) with Cc=0.4 pF and Rz=3 kohm while the trioded pcas degrades the fold pole.

One Gemini call, no loop

The same model asked once for a sizing, with no simulator feedback.

CircuitOne Gemini call passesFailures per try
diff1/35, 10, 0
fc0/214, 12
tele0/213, 14
Live

Run the crew yourself

routr live stage

Photo of a schematic in, spoken spec in, the crew sizes the circuit live, then layout, sign-off and red team. Replays work offline.

Open the stage
The stage is not running. Start it with web/serve.sh (or .venv/bin/python -m demo.server --port 8000) and reload this page.