Rating evidence, downloadable instantly

The problem

Proving a batch meets its UN rating takes days of digging through paper.

A chemical or pharma customer asks for evidence that a specific batch meets its UN rating. Today someone spends days rebuilding what test was run, when, and with what result, from loose paperwork that breeds distrust whenever there are gaps.

• A chemical or pharma customer asks for evidence that a specific batch meets its UN rating. Often before awarding the next order, sometimes after an incident in their own plant.
• Someone then spends days rebuilding which test was run, when, and with what result. The work falls on quality, usually on top of everything else that week. And the customer waits, wondering why proof takes so long.
• The sources are loose: test sheets, photos on a phone, a signature in a notebook, an email from the lab. None of them were designed to be put together afterwards.
• Every gap in that reconstruction breeds distrust with exactly the customers a drum maker for dangerous goods can least afford to worry.

How it fits the IRIS system

Edge captures each test as it happens; Agents files it by batch and standard.

With Edge + Agents, every critical test — drop, stacking, leak-tightness — automatically generates pre/post photos, a timestamped decision log, organized by batch and standard. The customer or auditor logs into the web and downloads the batch's evidence pack.

The evidence is never rebuilt, because it is never scattered: each drop, stacking or leak-tightness test leaves its photos and log on the batch the moment it is run. A customer request becomes a download instead of a project. And quality stops spending its week reconstructing what it already did.

See the full IRIS architecture →

Before and after

Rebuilding UN evidence today versus the dossier built as tests run

Aspect	Today	With iLEAN Edge + Agents
Answering a customer's evidence request	Days of reconstruction	An instant download
Drop test photos	On someone's phone, if anywhere	Pre and post, filed by batch
Stacking and leak-tightness results	In separate sheets	In the same dossier
Timestamp and who signed	Inferred afterwards	Logged at the moment
Organized by standard	By hand, each time	Automatically
Auditor access	Emailed files	A scoped web view to download

Before: days spent rebuilding evidence for every customer request. After: instant download, no post-hoc human work.

Impact estimate

Impact estimate — to be validated with your numbers.

The block below is an estimate to be validated against your plant's actual data. We put it forward so the committee has an order of magnitude; we refine it during the assessment.

• Estimated payback 6-12 months, depending on how often customers request evidence.
• It comes from continuously protecting relationships with dangerous-goods customers, not from a one-off saving. Those relationships are where a drum maker's margin lives. Every request answered in minutes reinforces them.
• From days spent rebuilding evidence for every request to an instant download with no post-hoc human work. The same dossier serves the customer, the certification body and your own quality reviews.
• And a gap in the evidence is noticed when the test is run, not when the customer asks. Missing photos or an unsigned test raise a flag the same day. Quality fixes the gap while the drums are still in the plant.

Estimated payback 6-12 months depending on request frequency, from continuously protecting relationships with dangerous-goods customers. Estimate to be validated.

And the fair question from the production manager

“Will an auditor accept evidence filed by AI?” — the AI files, it does not judge: the result is the test rig's and the sign-off is the quality lead's. Matching each photo and log to its batch and standard is an anchored task, where the best models drop below 1.5% error [1], and the dossier keeps the original images so anyone can check them. The dossier shows the evidence; it does not replace it. An auditor sees exactly what the rig recorded and who signed it.

^[1] OpenAI paper "Why Language Models Hallucinate", 2025 — on the reliability of AI in anchored tasks.