Composite study

This is a product-class write-up, not a named-client story and not a results press release. Counts below describe the pack we actually write (approach points, tools, layers) — not invented bug totals or conversion lifts.

4
Approach points
4
Tools in the pack
4
Layers covered
3
Challenges documented

Overview

The demo stream was live. When the link dropped, the last value stayed on screen and the model said “stable.”

Timestamp source mixed UTC and plant local time in prose.

Industry pattern this is grounded in

Industrial assistants fail on freshness and fail-safe, not on eloquence. A missing stale-data indicator is how people act on a dead picture.

Highlights

Typical launch bar: Block until pass rules hold on the money, identity, or safety path.

Challenges

01

Product called a 15-minute warehouse rollup “real time.”

02

The model interpolated missing points.

03

Mobile operators on bad networks saw even older caches.

How we tested

01

Every live claim includes age. Past a threshold, the assistant must refuse present-tense status.

02

Fail-safe UX when the link drops.

03

Alarm ack/escalate still works without the LLM.

04

Documented freshness SLO, measured.

Starting point vs what we required

Starting point versus the controls we required
Starting pointWhat we required
Product called a 15-minute warehouse rollup “real time.”Every live claim includes age. Past a threshold, the assistant must refuse present-tense status.
The model interpolated missing points.Fail-safe UX when the link drops.
Mobile operators on bad networks saw even older caches.Alarm ack/escalate still works without the LLM.

What we delivered

01

Journeys / UI

  • Playwright — Stale banner, dropped link, ack without AI
02

API / contract

  • Postman — Tag age headers
03

LLM traces / evals

  • LangSmith — Whether the model saw last-good vs live
04

Load / soak

  • k6 — Many points / high frequency UI lag

Outcomes

We forced a dropped-link fixture. Present-tense answers without a fresh tag failed. Interpolation was labeled estimate or forbidden.

Core alarm path was tested with the copilot disabled — the plant cannot depend on a model for fail-safe.

Stack

Testing stack used on this product class
LayerToolWhat we validated
Journeys / UIPlaywrightStale banner, dropped link, ack without AI
API / contractPostmanTag age headers
LLM traces / evalsLangSmithWhether the model saw last-good vs live
Load / soakk6Many points / high frequency UI lag

Same method as the rest of our work: severity-ranked findings, pass rules in writing, and a ship / wait / block call. See the evaluation method and tool comparison.