Notes from the plant floor
Long-form writing about the parts of tire autonomy that are genuinely hard: perception under real inspection conditions, cure-state estimation, twin fidelity, false-hold rates and the governance that makes any of it deployable.
Topics we write about
Recent posts
Why sampling cannot find a splice problem
Engineering · 9 min read
If a fault appears in one component out of four hundred, an audit sample is a coin flip. The arithmetic of per-component measurement, and what it costs to get there.
Curing to state instead of to a clock
Process · 11 min read
Fixed cure recipes carry a safety pad sized for the worst compound batch you have ever had. Estimating cure state per tire lets you spend that pad only when it is needed.
What an as-built tire twin has to model
Simulation · 13 min read
Builders, presses, molds, conveyors, robots and WIP buffers. What we learned about which fidelity actually changes a decision, and which is expensive theatre.
Fusing X-ray and shearography without doubling false holds
Perception · 10 min read
Two modalities disagreeing is information, not noise. How we combine confidences and why false-hold rate is our hardest gate.
The control envelope: making safety a reviewable file
Trust · 8 min read
Why we made the safety contract a versioned, signed artifact that a plant safety engineer can read in a pull request.
Synthetic splice-open defects that survive contact with reality
Data · 12 min read
Generating rare defect families is easy. Generating ones that improve a real detector is not. Our validation protocol.
Also worth reading
What we learned from 4,000 engineer overrides
Autonomy · 10 min read
Overrides are the most valuable data in the plant. A taxonomy of why engineers disagreed with the model, and what we changed.
Genealogy is a graph problem, not a reporting problem
Platform · 9 min read
Linking batch, component, builder, press and grade at write time turns recall scoping from a project into a query.
Latency budgets at the building drum
Edge · 7 min read
Where the milliseconds go between a camera frame and a hold decision, and which ones are worth fighting for.
Why sampling cannot find a splice problem
The arithmetic that ends the audit-sample argument
Suppose a drum drifts and produces a marginal splice in one component out of four hundred. An audit sampling one tire per hour has a vanishing probability of catching it before a full shift of production carries the same fault. Per-component measurement is not a nicety — it is the only sampling rate that matches the failure rate.
- Detection probability collapses as defect rate falls
- Detection lag, not detection rate, drives the cost
- Correction inside the build cycle changes the economics entirely
- False-hold rate is the constraint that makes it deployable
Ongoing series
Anatomy of a defect family
One post per family — splice-open, bubble, separation, bare spot, cord shift, bead mis-seat — covering signature, sensing, confusion cases and correction.
Gate by gate
What it actually took to move each agent from shadow to assist to bounded autonomy at a design-partner site.
Twin fidelity
Which parts of the line need physics, which need statistics and which need nothing at all.
The evidence chain
Building audit-grade logging that quality organizations trust and engineers do not route around.
What plant engineers say about the writing
“We had three people watching X-ray images and still shipped uniformity rejects. Treadum flagged the cord shift at the building drum, not four hours later at final inspection.”
“The cure agent found 40 seconds of margin on a construction we had run the same way for eleven years. It proved it in shadow mode before it touched a press.”
“Genealogy is the part I did not expect to care about. Every tire now has a linked record from batch to grade, and audit prep went from weeks to an afternoon.”
[PLACEHOLDER] Design-partner quotes are illustrative until pilot references are published.
How we publish
About the writing
Yes, with attribution and a link. Several partners circulate the defect-family series to their process teams, which we take as the highest compliment available.
From plant engineers and process specialists, yes. We do not run vendor guest posts.
Design-partner pilots and our own simulator. Anything not yet published with partner permission is marked [ASPIRATIONAL], including in posts.
Subscribe to the plant brief from the footer. It is roughly monthly, and it is writing rather than marketing.
Put one cell on autonomy in 90 days
Pick one wedge — splice inspection, X-ray defect detection, cure-state optimization or uniformity prediction. We baseline it, run shadow mode, then graduate to bounded autonomy under a signed control envelope.
Pilots start in shadow mode. No line changes until accuracy and safety gates pass.