Auto Secure LoginLook first, write second — copy that refuses to invent what the image doesn't show.
ASL Product Vision is a shared machine-learning module: given one product image and a one-to-nine-word seed, it produces a single grounded 150–350-word product description. It analyzes the visible evidence first, then writes — and refuses to draft at all rather than hallucinate when the evidence is too sparse. It never publishes, syncs, or mutates a storefront; its only output is a local structured draft, marked for human review.
The problem
AI copy invents features the product doesn't have.
Auto-generated product descriptions are notorious for confidently describing a color, material, or capability that simply isn't in the photo. For a catalog, that's not a quirk — it's a false claim that can mislead a buyer and expose the merchant.
Product Vision makes anti-hallucination the entire product. It grounds the description in what the image actually shows, and would rather produce nothing than fabricate.
Structural, not hopeful
- Evidence-before-prose ordering
- Response isolation between reading and writing
- Dual-reading consensus before it commits
- Abstention over invention when evidence is thin
Capabilities
A narrow, honest copy engine.
Evidence-first analysis
It analyzes the visible evidence in the image before a single word of copy is written.
Grounded description
Produces one 150–350-word description from an image plus a 1–9 word seed — grounded in what's actually shown.
Abstention over invention
When evidence is too sparse, it refuses to draft rather than fill the gap with a plausible guess.
Dual-reading consensus
Two independent readings must agree before the module commits to a claim.
Receipt-bound inference
Deterministic, receipt-bound, local-only inference — no remote endpoint in the loop.
Never mutates a store
It never publishes, syncs, or mutates a storefront — the output is a local draft only.
Human-review gate
Every draft is marked LOCAL_DRAFT_REVIEW_REQUIRED — a person approves before anything goes anywhere.
No GPU required
15 unit tests run with no model or GPU needed — the discipline is testable in isolation.
A capability, not a runtime
Diagnostic-only and designed to sit behind ShopFit's catalog copy, not to run as a public service.
Copy grounded in the photo — or no copy at all.
An image-to-description engine built to abstain rather than fabricate, and never to touch a live store on its own.