The source was never the hard part. The layer above it was.

Composing the mark for the part that is actually under the head

Nothing is pre-rendered and nothing is stored twice. Every mark is built for the individual part arriving at the head, out of the plant own control and business systems, while the part is moving.

1. Step 1

   Trigger

   The Control Layer reads the PLC tag that identifies the physical part at the head. The posture is read only: the agreed tags and nothing else, no ladder logic or function blocks touched, and a write only where a handshake has been specifically scoped and signed off. OPC UA runs with certificate-based security where the controller carries it, and Modbus TCP, which has no authentication of its own, is scoped to a network segment agreed with your control engineer.

2. Step 2

   Job pull

   Job, part, batch, serial and order data are taken from the system that already owns them, through the connection pattern your IT team sanctions: REST or SOAP, a read-only database view, a watched file drop, a message queue, an OPC UA or Modbus TCP tag read, or middleware your organisation already runs. Each pattern carries its own authentication model and its own documented failure edge, and the pattern is chosen against what your environment runs today rather than against what would be convenient to build.

3. Step 3

   Composition

   A fresh mark is composed against that part grade and dimension: fixed logo, standards text, the variable date, time, shift, run and pack fields, and a machine readable code. Data Matrix is generated to ISO/IEC 16022, graded to ISO/IEC 15415 and linear symbologies to ISO/IEC 15416, with the accept threshold set against the scanner class your operation actually carries.

4. Step 4

   Velocity compensation

   An encoder tracks the part and the mark geometry is compensated for line velocity, so a shape drawn onto a surface that is moving under the beam comes out true rather than sheared. The compensation runs against the encoder the line already has, not against an assumed constant speed.

5. Step 5

   Fire

   The scanning mirror pair draws the whole mark inside the window the part spends in the head field of view. That window is the real engineering budget: a typical timber brand holds across 40 to 120 m/min on one head, and a full height dense brand, around 40 mm high and up to roughly 600 mm long, tops out around 75 to 85 m/min, because the ceiling is how much stroke length one mirror pair can draw in the time available, so above that the specification is a second head sharing the mark on the same encoder.

6. Step 6

   Write back

   The completion is written back as an audit record against the job, with timestamp, operator, machine and station. The plant keeps one source of truth, and on a multi-site programme that constraint is set once for the group rather than negotiated site by site.

A record-keeping problem is never allowed to become a production problem

Every brochure in the category describes a system that works. This is what the Control Layer does when the data underneath it is missing, contradictory or moving, and it is published as architecture because it is where a marking programme is actually decided.

• The station fills a pull-ahead cache rather than calling per part, so an ERP maintenance window never stops a physical line.
• Completion records queue locally through an outage and replay in order carrying the original mark timestamp rather than the write time. That is the difference between an audit trail and a fiction.
• The station never invents an identifier. No default, no placeholder, nothing assumed, and it will never increment an identifier itself. If the cache runs dry it stops issuing marks rather than composing from stale data.
• If the PLC signal is missing, or two signals conflict, the station refuses to match and logs the exact signal state, so your control engineer debugs from a record rather than from a description.
• If the schema or the credentials move underneath it, the station fails the pull rather than composing a mark out of fields it no longer understands.
• Where an upstream identifier could collide, either the uniqueness rule is fixed in the source system or the mark carries a composite of site, line, date and run, agreed in writing before build. Hold versus divert behaviour is chosen by you and written down before build.
• The security posture was a design input rather than a retrofit. Integrating a marking line requires no inbound path into the plant, and your ERP stays unreachable from the public internet, which is what lets the same architecture run on an isolated OT network without a variant build.

Optics, mechanical, control software and enterprise systems, engineered as one stack

A control layer is only useful if something physical presents the part to the beam, something optical puts energy into it in a controlled way, and something enterprise-grade owns the record afterwards. All four disciplines run in the same engineering programme, which is why the boundaries between them are specified rather than negotiated after install.

The constants the core is built on, published raw

One core was written in Nelson and then made to carry both lanes, the benchtop application and the industrial marking layer, which forced every number below to be pinned in the engine rather than tuned per install. The figures are published raw rather than summarised, because a reader who recognises them knows immediately whether the people who wrote the page also wrote the code.

• One internal, dialect-neutral motion program, emitted into GRBL, Marlin, Smoothieware and Generic dialects, each with its own laser-on command, travel strategy, power scale, precision and streaming discipline. The neutral program is the thing that was actually invented here: geometry is decided once, and every controller-specific quantisation happens at the last possible moment inside the emitter, so a controller nobody has shipped yet is an adapter rather than a rewrite.
• Emitted machine output is pinned by a golden fixture suite and checked by a conformance validator on every build, which is the unglamorous half of the invention and the half that took longest to get right. A change to the geometry engine cannot silently change what a machine does, which is the answer to the question an industrial buyer rarely gets to ask: what happens to my line when the software updates.
• Vector optimisation is a fixed pipeline rather than a set of switches: duplicate edge removal, short segment removal, connected path building, collinear merging by cross product, curve simplification, closed loop detection, ordering by descending signed area so outer contours run first, rotation to the nearest approach point, and travel insertion. Every stage carries a published tolerance, from a hundredth of a millimetre on the dedup threshold to two tenths on the path snap, and publishing them is the point. A pipeline whose tolerances live in someone head behaves differently on the day that person is away.
• Ten stage raster processing built on colorimetric physics rather than a brightness slider: bilinear resize to target DPI, sRGB to linear luminance on the Rec.709 coefficients, subject statistics with percentile clipping, auto levels, exposure, contrast around the midpoint, gamma, optional invert, mode-specific conversion, and emission with serpentine rows, 2.5 mm overscan for velocity stabilisation and inline power quantised into 50 steps. We treat the image as a light source, not as a picture.
• Three dithering error models, because on a laser this is a materials decision rather than an aesthetic one: Floyd-Steinberg diffusing error to four neighbours at 7/16, 3/16, 5/16 and 1/16 with serpentine bidirectional scanning, Atkinson diffusing one eighth of the error to six neighbours and deliberately discarding the rest to lift contrast, and Bayer using a 4 by 4 ordered threshold matrix. Which model is correct is a property of the substrate and the spot size, and every one of those calls was settled on burnt coupons rather than on a screen.
• Power is emitted inline on every move and computed against actual velocity rather than commanded velocity, because energy per millimetre is what the material sees. The software reads the machine own $30 and $31 registers at connect time and queries $110 and $111 to derive and validate the effective maximum feed rate. Every controller on the bench answered that question differently, which is why it is asked at connect time rather than stored in a profile.
• Character-counting streaming tracks bytes in flight against the controller 128 byte receive buffer and keeps it full rather than waiting for an acknowledgement per line, with a send and wait fallback on a 3, 8 then 12 second retry ladder. Recovery after a stop is a defined sequence rather than a reconnect, because the coordinate frame has to survive the interruption. An operator who has to re-home the machine after every stop stops using the stop button, and that is a safety outcome rather than a convenience one.
• Exactly one package may import the desktop framework runtime. Code generation, serial communication, controller adapters and power mapping, project and material persistence, licensing, image handling and export all sit below that line as pure library code, with the boundary enforced in continuous integration. That line is why the industrial lane inherits the core instead of forking it, and why a fix made on a benchtop machine reaches a production station.
• Every licence state transition is available fully offline. The application mints a Device Setup Code carrying the deviceId, an Ed25519 public key, device name and platform, and verifies the returned RS256 activation and lease signatures, issuer, deviceId and public key hash entirely locally. Secrets live in the OS credential store, macOS Keychain, Windows Credential Manager or freedesktop Secret Service, and updates are checked against a published SHA-256 checksum before installation. Air gapped operation on an isolated OT network is the design point, not a mode.

The laws the platform is engineered to

These are positions about the category rather than opinions about the company. Each one is testable, and each one shows up as a specific decision in the architecture above.

• The source was never the hard part. Sources, heads, enclosures and interlocks are catalogue equipment from named manufacturers through their New Zealand channel. The intelligence above them is where the engineering sits, and where the value moves.
• Beam time is path length, not area. Mark height, character density, total mark length, line speed and available power are the five variables that move the ceiling, which is why a published line speed with no mark stated beside it is decoration. Ask any party quoting one: at what mark height, at what character density, over what mark length, at what power.
• Constant power is not a constant result. Energy per millimetre is the controlled quantity, so power is modulated against actual velocity and the machine is asked for its own scale before a single move is emitted.
• Identity belongs in the material. A code printed on the packaging was always a temporary answer, and every part is getting a name. If a part goes through wash-down, weather, sterilisation or years of handling, the mark has to survive all of it, not just look right on day one.
• The record already exists. The mark carries the code your ERP or MES generated, against the job that system already opened, so the plant keeps one source of truth and nothing is maintained in parallel.
• A system is characterised before it is accepted. Acceptance criteria agreed after commissioning are a negotiation. Written before the build, they are engineering.
• Continuity is engineered rather than asserted. Every source, head, controller and interlock is standard industrial equipment from named manufacturers, every marking parameter, machine profile and interface specification is documented and resident on your site, and any competent integrator can pick the system up from the handover pack.