2026-07-14 · Bonding
Humidity is the quiet cause of most delamination
Why primer flash-off, not adhesive quality, explains the seasonal shape of warranty claims — and what a bond agent does about it.
9 min read
Notes from the factory floor
Technical writing about footwear manufacturing autonomy — bonding chemistry, nesting mathematics, edge scheduling and the design of systems that are allowed to touch machines.
Recent writing
Long-form, method-first, written by the engineers who built the thing.
2026-07-02 · Cutting
Yield per hide, not yield per batch
Batch-level yield hides the variance that actually costs money. Measuring per hide changed what our design partners bought first.
7 min read
2026-06-21 · Autonomy
Why our bond agent asks permission
The engineering case for act-with-approval on validated parameters, and what happens when the approval SLA expires.
6 min read
Humidity is the quiet cause of most delamination
A warranty analysis that starts with adhesive suppliers usually ends in the wrong place. The seasonality of claims points somewhere else.
When you plot delamination claims against production date rather than sale date, the shape follows plant humidity, not adhesive lot. Higher relative humidity extends primer flash-off; at fixed line speed the upper reaches the press before the primer is fully flashed, and the bond forms with entrapped solvent. The defect is invisible at the QC gate and appears eight to fourteen months later at the edge of the toe.
- Claims correlate with production-date humidity, not adhesive supplier
- Flash-off extends roughly 18% between 58% and 71% relative humidity
- The failure is invisible to visual QC and appears months later
- Control means moving activation temperature — a validated parameter
- 10:31:07 bond.check_window(primer="PR-9", adhesive="AD-22", rh=0.71)
- 10:31:09 RH 71% pushes primer flash-off outside validated window
- 10:31:10 approval.request(param="activation_temp", from=64, to=66, reason="humidity drift")
- 10:34:14 approved by t.nguyen@[PLACEHOLDER] · change logged to audit trail A-88213
- 10:52:40 quality.inspect(scope="bond_line,fit,cosmetic", pairs=1200)
- 10:58:27 first-pass yield 98.2% · 21 rework (squeeze-out) · 0 bond-gap escapes
Every line is real, copyable text and is announced politely to screen readers.
On cutting, moulding and lasting
The physical steps, in detail.
2026-07-02 · Cutting
Yield per hide, not yield per batch
Batch-level yield hides the variance that actually costs money. Measuring per hide changed what our design partners bought first.
7 min read
2026-06-09 · Twin
Simulating a changeover before the line stops
How the shoe-and-line twin cut first-good-pair time from 90 minutes to 34 at a contract manufacturer running four brands.
11 min read
2026-04-30 · Materials
Recycled EVA moves the process window
Mono-material and recycled constructions are harder to bond. What that means for density control and activation temperature.
9 min read
On building autonomy safely
Design decisions behind an agent that is allowed to write setpoints.
2026-06-21 · Autonomy
Why our bond agent asks permission
The engineering case for act-with-approval on validated parameters, and what happens when the approval SLA expires.
6 min read
2026-05-28 · Edge
Twelve models on one Jetson, inside cycle time
Deterministic scheduling, memory budgeting and what we gave up to keep station decisions under 100 milliseconds.
10 min read
2026-03-27 · Engineering
Evaluation gates for models that cause scrap
Golden datasets, LLM-as-judge and regression gates when a false negative is not a bad answer but a ruined lot.
10 min read
Why our bond agent asks permission
An annotated version of the reasoning trace that appears in the article.
Premise Autonomy is a permission model, not a capability claim
An agent that can change a validated bonding parameter is not more advanced than one that asks; it is less deployable in a plant under a brand quality agreement.
Design Four levels, set per parameter
Observe, recommend, act-with-approval, act — resolved per parameter rather than per agent, so nesting can be fully autonomous while activation temperature is not.
Failure mode What happens when nobody approves
The agent holds the last known-good recipe and stops affected pairs rather than proceeding. A missed approval costs throughput, never quality.
Result 34 approvals in a quarter, median 4 minutes
Approval fatigue did not materialise because requests carry the evidence and the twin prediction, so the decision is usually obvious.
Why traces matter
Every Footeon decision that touches a cut, a mould, a last or a bond line is recorded as plan, thought, action and observation — with the material lot, sensor reading and spec clause that justified it. Quality engineers can replay any pair.
Traces are retained for the full product-liability window and exported to your MES and quality system.
On people and pairs
What the plant floor knows and how it gets encoded.
2026-05-15 · Quality
Sampling misses the pair that comes back
The statistics of footwear inspection, and why 100% bond-line vision changes warranty economics rather than just defect counts.
8 min read
2026-04-12 · Craft
Encoding what the lasting operator knows
Craft memory: capturing per-operator judgement as model behaviour before the knowledge retires out of the building.
12 min read
2026-07-14 · Bonding
Humidity is the quiet cause of most delamination
Why primer flash-off, not adhesive quality, explains the seasonal shape of warranty claims — and what a bond agent does about it.
9 min read
How we write
Method first, numbers reproducible, nothing hand-waved.
- 9 articles published
- 100% articles that state their method and baseline
- 2 articles per month, roughly
- 0 articles written by a model without an engineer author
What we write about
And what we do not.
We write about
- Bonding chemistry and process windows
- Nesting and yield mathematics
- Edge scheduling and control latency
- Evaluation for physically consequential models
- Craft capture and operator knowledge
We do not write about
- Model benchmark leaderboards
- Generic AI commentary
- Customer data or named plants without written consent
- Claims we cannot reproduce from a stated baseline
Who this is for
Process engineers, quality directors and plant IT.
“The nesting agent found yield our best pattern cutter could not — and then explained the layout hide by hide. We stopped arguing about scrap and started managing it.”
Leather yield 79.1% → 87.4%
“Bonding is where our warranty costs live. Having an agent watch humidity, primer flash-off and press pressure on every pair — and stop for a human when it wants to move a validated parameter — is the first thing that has actually moved delamination.”
Delamination returns −71%
“We run 40 model changeovers a month. Simulating the nest, the mould recipe and the bond window in the twin before the line starts took hours out of every launch.”
Changeover time −46%
Design-partner scenarios are illustrative and modelled on public footwear-manufacturing benchmarks — named references are [PLACEHOLDER] pending customer approval.
About the blog
-
Agents are bounded by guardrails and schema validation, they cite the datasheet or spec clause behind a decision, and they fail closed to the last known-good recipe. Continuous evaluation against golden datasets gates every model and prompt change in CI before it reaches a plant.
-
Yes — and they are exactly where bond control earns its keep. Recycled EVA, bio-TPU and mono-material uppers shift adhesion behaviour and process windows; the twin simulates the new window and the bond agent holds it as lot chemistry drifts.
-
It stops and asks. Every agent runs at a configured autonomy level: observe, recommend, act-with-approval, or act. Bonding, lasting and any parameter under a brand or safety validation defaults to act-with-approval, and every approval is written to an immutable audit log with the sensor evidence that triggered it.
-
A typical wedge deployment on one line runs six to ten weeks: two weeks of sensing and baseline capture, three to five weeks of shadow-mode model tuning against your material lots, then graduated autonomy. Whole-factory rollouts follow line by line.
Get new writing by email
Roughly two articles a month. No product announcements dressed up as engineering posts.