The mechanism

English in. Evidence out.

A sentence becomes a versioned sequence, runs on your bench, and files its own evidence.

Book a demo See the sequence

573 instruments · 135 manufacturers · Apache-2.0

galois sequence — newDemo — synthetic data

compiling → hv_bringup.seq

✓ 01 psu.set ramp 0→1000 V · 50 V/s

✓ 02 hold 60 s · log I_leak @ 10 Hz

✓ 03 limit.check I_leak ≤ 10 µA

run: cloud · approval required: qa-lead

Fig. 1 — Sequence from English

01 — Sequence

One artifact, from English to evidence.

Sequences are versioned YAML — reviewed, locked, and run with every step traced.

Draft

Generated from the prompt.

Reviewed

Diffed like code.

Locked

Production revision, immutable.

Run

Executes on the bench.

Report

Evidence, assembled.

revision gatev1.4 ≠ approved v1.2

limit.checkI_leak 4.82 µA ≤ 10 µAPASS ✓

Fig. 2 — Sequence lifecycle

02 — Bench

Real instruments, one daemon.

Two static binaries connect the bench — no Python, no Docker, works offline.

galois-edge statusDemo — synthetic data

galois-edge active (running)

galois-edge-daemon active (running)

mesh tailnet-lab-01.ts.net

instruments 4 connected

Keithley 2400 GPIB::24

Keysight 34461A USB

Rigol DS1054Z LAN

SRS SR830 GPIB::8

grpc :50051 ready · websocket :8765 ready

Fig. 3 — Edge daemon

Vendor SDKs proxy through the daemon, so NI, Keysight, and Digilent stacks keep working.

Buses — GPIB · USB · LAN · Serial · Modbus · CAN · SPI · I²C · OPC-UA

SDKs — NI DAQmx · NI-VISA · Keysight IO · Digilent WaveForms · Zurich LabOne · Quantum Design PPMS

FPGA bitstream library — in build

Cloud, dedicated tenant, or air-gapped — see deployment.

Browse the instruments

03 — Evidence

Every run files its own evidence.

Per-step traces draft reports today; attested sign-off packages are in build.

Evidence report · EN 61010-1Demo — synthetic dataSigned

Step	Measured	Verdict
insulation_resist	1000 V · 4.82 µA	PASS
rail_5v0_limit	3.31 V ∈ [3.20, 3.40]	PASS
n77_evm_check	2.8 % ≤ 3.5 %	PASS

sha256 9f2c…e41a · qa-lead · run 1284 · seq pcba_eol v1.4

Fig. 4 — Evidence report

Every step records

measured valuelimitsraw command + responseinstrument IDoperatorDUT serialgit SHAsequence revisiontimestamp

Reports

Drafted from live run data.

Sign-off packages

RTM, findings, and attestation compiled from the same run record.

In build

Failure economics

About 50% of products fail EMC testing on the first attempt (Intertek). Every failed loop costs weeks of schedule and five figures in re-tests and rework — and a full certification campaign in a regulated market runs well into six figures.

Galois automates the loop — so the first attempt is the last.

Source: Intertek — EMC first-pass failure Source: Tektronix — EMI/EMC re-test costs Source: PCBSync — respin schedule cost Source: Quilter — board respin average Source: Medical Device Academy — 510(k) testing

04 — Design · Sim

Catch it before the fab run.

One ECAD ingest feeds nine board analyses and keeps the test plan tracking every revision.

Simulation

IR-drop to 2.5D EM — nine analyses before boards are ordered.

stokes — s-param correlationIn build

Fig. 5 — Sim vs measured

Design plugins

In build

The Altium extension snapshots the design; findings cross-probe back to the schematic.

Altium 26.6 · KiCad M0

Fig. 6 — Design to test

05 — Memory

Memory

The bench remembers.

Runs, annotations, and recurring failures ingest automatically; every answer comes back cited.

évariste — memoryDemo — synthetic data

? why did rail_5v0 start failing on line 4

V_rail drifted +0.12 V after the 06-12 fixture swap; the 3.60 V limit was first exceeded in run 1279.

run 1279run 1284pcba_eol v1.2→v1.4fixture log 06-12

Fig. 7 — Memory query

Memory, in depth

$ curl -fsSL https://get.galoislabs.ai | sh

Now run it on your hardware.

Apache-2.0 daemon, free forever. Enterprise is custom — cloud or on-prem.

Book a demo Install the daemon