INDEX / TECHNOLOGY

How the instruments are built.

Five engineering decisions, made in public. This page is the argument for each one, with the measurements that back it. If you disagree with a decision, the contact form is open and history shows we have changed our minds twice.

01 Component inspector

Open the chassis.

Five components carry the sound. Select one to read why it is built the way it is, and what tolerance it is held to.

02 Tolerance stacks

Four decimals or it does not ship.

A spec is only as good as the worst part in the chain. We budget tolerance like money: every component gets an allocation, and the stack has to close before the board is approved.

TOLERANCE STACK / VOICE ARCHITECTURE / REV C
Oscillator core±0.010% after compensation, per unit
CV scaler±0.006% matched resistor network
Pitch bus distribution±0.002% star ground, guarded trace
Calibration residual±0.001% EEPROM trim table
Stack total±0.019% worst case, RSS ±0.012%
Shipped spec±0.02% the number on the box, measured per batch
The stack closes at 0.019% worst case, which is why the printed spec is 0.02% and not 0.01%. We do not print a number the stack cannot guarantee. This annoys marketing and we consider that a feature.
03 Batch QC

Every batch argues with the spec sheet.

Production runs in batches of 120. Each batch is measured as a group, because instruments that ship together should agree with each other. The pass rate below is the trailing twelve months, and the failures are not hidden: they are reworked or scrapped, and the reason is logged.

0batches Shipped, trailing 12 months
0% First-pass QC rate
0% Reworked or scrapped, logged
0pts Measurements per unit before boxing

What the bench measures

  • Full-range pitch sweep, 20 points, both oscillator cores
  • Self-oscillation frequency and amplitude stability, cold and after 30 minutes warm
  • Signal to noise at full output and at unity, IHF-A weighted
  • Crosstalk between every adjacent voice pair
  • Jack contact resistance, sampled per batch, cycled to 12,000 on 2% of units
  • Power rail draw at idle, typical use, and all-voices-active
  • Thermal imaging of the drive stage under sustained load

What the certificate says

Your serial number carries its own measurements: pitch error at 20 points, SNR, crosstalk figure, warm-up drift over the first 30 minutes, and the firmware hash it was calibrated under. If the certificate disagrees with this website, the certificate wins, because it describes your unit and the website describes the batch worst case.

Certificates are PDF and CSV. The CSV exists so you can graph your own instrument against the published distribution, which at least one customer has done and published, and which we endorse.

04 Materials

Chosen for the fourth decade, not the fourth week.

PANEL

6061-T6 aluminum

Machined from a single billet, 2 mm, type-III hard anodize with laser-etched legend. No screen printing to wear off, no adhesive labels to curl. The etch is deeper than the anodize on purpose.

KNOBS

Machined, not molded

Aluminum knobs with a stainless set screw and a polyamide carrier. Indicator lines are engraved, filled once at the factory. Wobble measured at under 0.2 mm at the skirt, checked per batch.

HARNESS

Labeled, socketed, reversible

Every internal cable is labeled at both ends. Every IC is socketed. The power connector is keyed and reverse-polarity protected twice: once in hardware, once in firmware sense. Repairs take an afternoon, not a returns department.

05 Firmware

CERN-OHL-S. Read it. Change it. Ship it back.

Every line of firmware for every module is published under CERN-OHL-S, with the repository history intact. The hardware accepts signed community builds alongside factory firmware, and two features in the current release started as community patches: the ARRAY-16 skip engine's beat-following mode, and the TESSERA correlated-noise spatial walk.

The firmware is small on purpose. The full HELIOS-8 voice management system is 11,400 lines of C, commented like we meant it. You can read the whole thing in an evening. If you find a bug, the bounty table in the changelog pays in modules.

License
CERN-OHL-S, all modules
Language
C, no RTOS, bare metal
Current release
3.2.0, October 2026
Community builds
Signed keys supported, one toggle in settings
Bug bounty
Confirmed bugs pay in modules or credit
Update path
USB-C, offline, no account required
Firmware changelog

Disagree with a decision?

Good. Two of the five components on this page were argued into existence by customers. The bench reads every message, and the best arguments end up as lab notes with credit.

05 Calibration, end to end

Calibration is not a final polish. It is a 96-step process that starts before the board has a panel and ends with a signed certificate, and it is the reason a HELIOS-8 from batch 2026-10 and one from batch 2025-11 agree with each other to within 0.02%.

  • Step 0, before assembly: the oven-stable reference in the Berlin rig is itself calibrated monthly against a locked standard, and the comparison sheet is public in the repository.
  • Steps 1 to 40, raw board: every trim point measured cold, before any human warmth touches the panel.
  • Steps 41 to 78, thermal soak: the module runs at load for 90 minutes inside the chamber, and every trim is re-measured warm, where it actually lives.
  • Steps 79 to 92, listening verification: a human plays the batch. Not a formality: the 9-engineer panel in the drift note came out of this step.
  • Steps 93 to 96, certificate: the six numbers on your sheet are read off the rig and signed. Nothing on the certificate is estimated.

RIG RECORD / OCTOBER 2026

Reference checks
12 this year, all logged
Chamber hours
1,240
Modules calibrated
3,180 this year
Failed at step 79
7, all reworked and resold only after a second full pass

THE RIG IS CALIBRATED BY THE PEOPLE WHO DOUBT IT

06 The measurement rig

RIG 01 / BERLIN

The reference bench

A temperature-stabilized room, a calibrated audio analyzer, and the oven-stable voltage reference that every other measurement on this site ultimately compares against. The reference is recalibrated monthly against a locked standard, and the comparison sheets are public in the repository.

RIG 02 / BERLIN

The thermal chamber

Where every module spends ninety minutes at load before its trim points are read warm. 1,240 chamber hours this year. The freezing-point myth note in the lab came out of this room, and the room has opinions about marketing copy, all of them unfavorable.

RIG 03 / DETROIT

The verification bench

Identical measurement chain to Berlin, run independently, so every second-source component and every Detroit-assembled batch is verified against two benches that do not share an office, a language, or an incentive to agree too quickly.

Rig documentation, including the calibration log and the method notes behind every number on this site, ships with the hardware files under the open hardware license. If a measurement interests you more than a module does, that is a valid way to enjoy this company.