Ship Safe — The Journey So Far

The Product what Ship Safe is

Ship Safe screens individual freight shipment lines against dangerous-goods rules and hands the screener a clear determination on each line — so a busy screener can see, at a glance, what is clear, what needs a closer look, and what must not move.

It was designed and built by a supply-chain and freight-forwarding professional, working with AI coding collaborators under his direction — domain judgment first, code second.

A written rules-language specification sits underneath it, so new screening rules are authored consistently against a defined standard, not improvised screen by screen.

Design Principles the rules behind the rules

• Layered rules. A public rules core, with client-specific rules layered on as additions — never replacements.
• Context adjusts, never silently. Richer context may move a result into a different band, but the adjustment is always visible — never applied quietly.
• Stated hazards are hard blocks. A shipper's own stated UN number or hazard class is a hard block that no AI output and no added context can soften.
• Deterministic rules are the authority. AI assists with research; it does not make determinations.

The Journey September – October 2026

• Sep 2026A 75-line synthetic test catalog is built — from plainly safe goods to very likely dangerous goods — to validate the screener's judgment before anything else is trusted.
• Oct 2026The rule set is re-architected into layers: a public rules core, with client-specific rules as additions on top.
• Oct 2026A design Charter (v1) is finalized, setting the product's principles down in one document.
• Oct 2026A completely rebuilt application shell for the desktop experience goes live, with an enterprise-style visual overhaul — light, clean, corporate.
• Oct 2026A knowledge ontology of the screening domain — the kinds of knowledge a screener actually uses — is drafted with worked examples and reviewed as part of an AI design review.
• Oct 2026The reference library ships with the app, putting the screening references a screener reaches for inside the tool itself.
• Oct 2026The ontology is published as its own page in the app, locked down so it is available only where it belongs.
• Oct 2026Stage 1 of the AI review goes live: a screener can ask Muse Spark to review an individual shipment line right in the review pane — observations and questions only; the determination never moves. Verified live on the published site against a drill line with spare lithium-ion battery packs, where Spark's questions were a screener's own: installed or loose, Watt-hour rating, damaged packs.
• Oct 2026The full automated test suite runs 719 of 720 green in Firefox — the single miss a font-download flake in the test rig, not the app.

AI Research-Assist advisory only

A research-assist mode uses Muse Spark, via the Meta Model API, to brief the human screener on an unfamiliar shipper — who they are, what they typically ship, and what is worth a second look before screening their freight.

The screener can also ask Spark to review an individual shipment line, right in the review pane. It reads the line and answers the way a second screener would — what is present, what is missing, and the questions worth asking before the freight moves — reaching the model through a small proxy the project controls, under a daily cap.

Both are advisory. The deterministic rule set stays the authority on every determination, and no AI output can soften a hard block.

Where It Stands Today

• Desktop first. Current focus is the desktop experience, with an enterprise-style visual language — light, clean, and corporate, inspired by modern warehouse-management systems.
• A mobile companion exists as a separate, purpose-built layout — not a shrunk-down copy of the desktop app.
• Screens run offline, with a reader for EDI shipment files included.
• 14 rule tables in the public core, with room reserved for client-specific additions.
• AI line review is live in the review pane — Stage 1 done, advisory only, with the determination untouched.
• Next: Stage 2 — feeding the ontology corpus into the AI review, so its questions draw on the project's own knowledge base.

The Scoreboard validation results on synthetic test data