WENGOLD

Verification that compounds, in the open.

WENGOLD keeps every verification result that still holds after a change, and reruns only what the change could have touched.

Open-core pitch · October 2026MIT OR Apache-2.0
The problem

Verification is the largest cost in a chip program

Building a chip takes hundreds of engineers and years of work, and most of that effort goes into proving the design is right. On a leading-edge program that share is now the single biggest item, and it is heading higher.

40–50% of a leading-edge chip budget, today
60–70% of engineering hours, on the current trajectory
0 open tools today that decide what still needs re-checking
Source: industry practice for leading-edge nodes.WENGOLD · 02
The iteration tax

Every change costs the whole regression

01
Rerun the regression
Correct, and the cost of a change equals the cost of the whole suite.
02
Hand-pick the tests
Fast, and unsound — the tests you skip are the ones that would have caught it.
03
Freeze the design late
Safe, and it is why verification drags to the end of a program.
WENGOLD · 03
The system

One graph, one kernel, one ledger

01
wengold-graph
One dependency graph of the design, the verification environment and the runs.
graph.wengold
02
wengold-kernel
Classifies every prior result against a change and returns the minimum new evidence.
valid · invalidated · unknown
03
wengold-ledger
An append-only, hash-chained record of every run, revision and verdict.
ledger.tof
Agents propose; the kernel and the tools decide.WENGOLD · 04
The kernel

Valid, invalidated, or unknown — never assumed

Verdict When What happens
valid No certain dependency on the change Keep it, with its evidence
invalidated Depends on, or is reached by, the change Rerun it
unknown Reachable only across an uncertain dependency Re-establish; never assume

Where the graph cannot resolve a dependency, the kernel returns unknown rather than a false pass — the safe direction.

WENGOLD · 05
Worked example

One changed line reruns three checks, not the regression

r1 to r2 · the change
rtl/sec_firewall.sv
- wire is_secure = ~prot[1];
+ wire is_secure = prot[1];
The protection bit is read backwards. A non-secure write returns OKAY and the DMA path to the key store opens.
sec_acl_test invalidated
dma_key_test invalidated
soc_smoke invalidated
a_secure_bit invalidated
npu_conv_test valid
4 results to re-establish · 1 kept with its evidence
Source: WENGOLD bundled SoC demo.WENGOLD · 06
Tool coverage

It runs on the free toolchain, and the ones you license

01
Simulation, debug and testbench
VCS · Xcelium · Questa · Verilator · Verdi · SimVision · Verisium · GTKWave · cocotb · UVM
02
Formal, emulation and FPGA
JasperGold · VC Formal · OneSpin · ZeBu · Palladium · Veloce · Vivado · Quartus · Yosys · vManager
03
Around the flow
Git · GitHub · GitLab · Jira · Slack · Tcl · Python · Make
Green = open source, first-class, no licence required.WENGOLD · 07
The model

The kernel stays open; scale and trust are the business

Open core · MIT OR Apache-2.0
Dependency graph, validity kernel and ledger · agent roles and the MCP server · open tool adapters and the commercial adapter contract · CLI, Python API and CI.
Free forever. No metering of the kernel.
WENGOLD Enterprise
Hosted graph at fleet scale · curated agent packs · air-gapped and on-prem deployment · SSO, RBAC, audit export · regression-farm orchestration and support SLA.
Annual, per design site.
We never gate a capability that decides whether a design is correct.WENGOLD · 08
The field

No open tool answers what still holds

Player What they do The gap
Synopsys · Cadence · Siemens EDA Full EDA suites Sell speed, not fewer runs
Closed AI-verification startups Autonomous agents, closed source Closed, and on commercial tools
Open agent projects CoreSmith, UCAgent, VeriPower, xverif Do a task, not the program layer
Open EDA toolchain Verilator, Yosys, SymbiYosys, cocotb Complementary — no validity kernel

The moat is the kernel and the graph behind it. A competitor can wrap a model around a tool in a weekend; a graph of the whole design that survives every revision is the compounding part.

WENGOLD · 09
The market

A large market, growing faster than the R&D it serves

Annual growth rate
7%
13%
Semiconductor R&D EDA
EDA market, 2025 $16–20B
Growth, CAGR 9–13%
Held by the Big-3 85%+
Verification share of budget 40–50%
Source: ESD Alliance / SEMI (Q4 2025), IDC (2025), SemiAnalysis (2026); scope varies by estimate.WENGOLD · 10
Why now

Two shifts made this the moment

Open silicon matured
Verilator, Yosys, SymbiYosys, cocotb and OpenROAD now cover the front of the flow, so a small team can design without a licence.
Agents arrived
LLM agents can drive tools and write tests. What they lack is a durable record of which results still hold across revisions.
The buyer widened
Hyperscalers and system companies now design their own silicon, adding buyers who never bought a full EDA suite.
A program-level verification layer needs all three at once. No open tool has them together today.
WENGOLD · 11
Roadmap

From v0.1 today to spec-to-signoff on Sky130

v0.1 Graph, validity kernel, ledger, tool registry, CLI, bundled SoC demo. Shipping today.
v0.2 LLM agents (plan, testgen, debug, repair), MCP server, cocotb runner.
v0.3 Register-map and CDC front-ends (SystemRDL, IP-XACT); Surelog and Verible backends.
v0.4 Formal loop (SymbiYosys + Z3) with counterexample-guided repair.
v0.5 Regression farm, coverage closure, waiver register.
v1.0 Spec-to-signoff on the open toolchain, reproducible on Sky130.
Not a foundry-signoff claim. Missing tools and unpinned PDKs are reported as missing, never as passes.WENGOLD · 12
WENGOLD

Open the layer that removes verification iteration.

The kernel stays open. Scale, air-gap and support are the business.

github.com/wengold/wengoldMIT OR Apache-2.0 · 2026