Over the last several weeks, AXIOM has been moving through one of the largest architectural changes in the project so far.
What began as a monitoring application for a small print farm is becoming something broader: a human-supervised operations platform designed to coordinate printers, observe jobs, evaluate failures, and learn from the people actually running the machines.
Several major systems have now reached the point where they work together as one platform rather than as separate features.
Operator Certification System
One of the most important additions is AXIOM’s operator certification workflow.
When the computer-vision system identifies a possible failure, the operator can review the captured image and classify what actually happened.
The operator can confirm the failure type, reject the detection, or mark the image as unclear.
That distinction matters because AXIOM does not treat every operator interaction as automatic training truth.
The system records the evidence, the operator decision, the detector output, and the context surrounding the event.
This creates a traceable chain between what the AI saw and what the operator decided.
Operator Chooses. AI Qualifies It.
AXIOM now follows a simple principle:
Operator chooses. AI qualifies it.
The operator determines what happened.
The system determines whether that evidence is suitable for future training.
This prevents accidental or ambiguous classifications from immediately becoming training data.
Before evidence can be promoted, AXIOM checks things such as image identity, detector context, classification state, and evidence integrity.
The result is a training pipeline where human judgment remains authoritative while the system protects the quality of the dataset.
Training Data With Provenance
Every training example AXIOM accepts now carries provenance.
That means the system can answer questions like:
- Which image was reviewed?
- Which operator certified it?
- What did the detector originally predict?
- What region of the image was selected?
- Which model produced the detection?
- Was the evidence later promoted into the training dataset?
This is especially important as AXIOM moves toward continuous model improvement.
Instead of building a dataset from a folder full of images with uncertain origins, the system builds a record of why each example exists.
Region Editing
The operator can also correct the region associated with a detection.
If the AI identifies the correct failure but places the bounding box incorrectly, the operator can adjust it.
That corrected region becomes part of the certified evidence.
This allows AXIOM to improve not only classification but localization.
Model Qualification
AXIOM is also being built so that new AI models cannot simply replace the current detector because they appear better in a benchmark.
New models enter a challenger role.
They are evaluated against certified operator evidence and must demonstrate measurable improvement before becoming eligible for promotion.
The goal is to make model upgrades evidence-driven rather than assumption-driven.
Failure Detection Architecture
The computer-vision pipeline has also been redesigned around explicit decision stages.
AXIOM now records how each detection passes through the safety pipeline.
Stages include:
- motion checks
- projected toolhead position
- build-volume checks
- temporal validation
- burst confirmation
Each stage records whether it passed, blocked the detection, was not consulted, or was intentionally bypassed.
This gives AXIOM something extremely valuable: the ability to explain why a detection was allowed or withheld.
Instead of simply saying “the AI did not trigger,” the system can trace the decision path.
Job Queue
The Job Queue has also been rebuilt into more of an operations board.
Jobs are separated into:
- Queued
- Printing
- Unknown
- Done
The Unknown state is especially important.
If AXIOM cannot confidently determine what happened to a job, it does not pretend the job succeeded or failed.
It moves the job into a reconciliation state where the operator can investigate.
This keeps uncertainty visible instead of silently hiding it.
Printer Controls
Printer control actions have also been hardened.
Pause, resume, cancel, and emergency actions now pass through explicit operator controls and validation paths.
Critical actions require deliberate confirmation.
This is part of a larger design philosophy inside AXIOM: automation should reduce operator workload without removing operator authority.
Timelapse and Visual Evidence
Timelapse capture is becoming more than a visual feature.
AXIOM is moving toward consistent layer-based capture where the printer head moves to a predictable location before the image is taken.
That consistency opens the door to additional forms of visual analysis.
For example, future versions of AXIOM may be able to compare the expected location of a print with its actual position on the bed.
That could allow the system to detect bed detachment or part movement without requiring another machine-learning model.
Instead, the system could use geometry, printer position, slicer information, and camera alignment.
What AXIOM Is Becoming
The larger goal is becoming clearer.
AXIOM is not intended to replace the operator.
It is intended to give the operator leverage.
The system watches the printers, organizes the work, records evidence, evaluates failures, and helps surface the situations that actually require attention.
The operator remains responsible for judgment.
AXIOM handles the observation, coordination, and memory.
That relationship is becoming the foundation of the platform.
What’s Next
Development is continuing across several areas:
- bed-adhesion detection using geometric tracking
- camera pose normalization using fiducials
- improved layer-based timelapse capture
- expanded model qualification
- operator-certified training pipelines
- job scheduling and reconciliation
- fleet-level printer coordination
Each of these systems is being built around the same principle:
Automation should make the operator more capable, not less informed.
AXIOM is still under active development, but the pieces are beginning to connect.
And that is where the project starts becoming much more interesting.
0 comments