ASL Product Vision — Source library
Public documentation snapshot: 2026-09-08
Stop shipping product copy that lies about the product.
Product copy grounded in what the photo shows, and no copy at all when it can't be.
Stop shipping product copy that lies about the product.
Built for: Shop owners with hundreds of products, one photo each, and description fields that are still empty · Anyone who tried an AI copy writer and found it confidently describing materials, sizes and features the photo never showed · Merchants in categories where a wrong claim is not a typo but a return, a chargeback, or a legal problem · Store owners who want blind and low-vision customers to actually be able to shop the catalog
ASL Product Vision exists because of this.
You have a few hundred products and one photograph of each. Writing a real description for every one is the job that never gets done. So the fields sit empty, or they hold the one line the vendor supplied, or a sentence you typed at eleven at night that says nothing. Search engines have nothing to index. Shoppers land, learn nothing, and leave. You know exactly what it is costing you and you still cannot find the week it would take to fix. So you try an AI writer. It is fast, it is fluent, and it invents. It calls painted resin solid brass. It calls printed vinyl hand-stitched leather. It adds dimensions nobody measured and a care instruction nobody wrote. You catch most of it. The ones you miss turn into a return, an angry message, a chargeback, and in some categories a claim you are not allowed to make at all. Once you are checking every generated paragraph against the photograph line by line, the tool has saved you nothing. You are back to writing it yourself, only now you are also proofreading a machine. There is a third cost that rarely gets counted. Someone using a screen reader arrives at your store, hears an empty alt attribute and a paragraph of adjectives, and still has no idea what the object looks like. They cannot judge it, so they do not buy it. Good product writing is not decoration. It is the only description of the object that some of your customers will ever get.
ASL Product Vision exists because of this.
Looking and writing are separated. The part that reads the picture never writes prose, and the part that writes prose never sees the picture. An invented detail has nowhere to enter unnoticed.
ASL Product Vision exists because of this.
Refusal is a first-class outcome. When evidence is thin the program produces nothing and says why, instead of producing a fluent paragraph that is wrong.
ASL Product Vision exists because of this.
Every fact carries its source, so a finished draft can be audited sentence by sentence rather than trusted as a whole.
ASL Product Vision exists because of this.
Failed approaches stay on the record with exact numbers. A study that misses its bar is frozen and kept as evidence rather than retuned until it passes.
ASL Product Vision exists because of this.
It is careful about its own evidence. Two readings from one model are called corroboration, not independence, and the difference is written down rather than blurred.
ASL Product Vision exists because of this.
Accessibility is a pass condition, not a feature request. A draft that does not describe the object usefully to a listener does not get released.
ASL Product Vision exists because of this.
The high-risk claims that cause real disputes, material, size, care, origin and certification, are blocked from being inferred at all rather than merely discouraged.
How ASL Product Vision works, start to finish.
ASL Product Vision is a research program with one narrow job. You give it a single product photograph and up to nine words of your own. It gives back one description of between 75 and 350 words in which every sentence can be traced to something visible in that picture. It never publishes anything and it never touches a store. The end of the road is a draft sitting on your own machine, marked as requiring a person's approval before it goes anywhere. The order of operations is the whole idea. Most tools ask a model to look at a picture and write marketing copy in a single breath, which is exactly what makes an invented detail impossible to catch. Product Vision separates looking from writing. The picture is read twice, with the two readings prompted differently and each one sealed off from the other's answer. A fact survives only if both readings report it and place it in the same part of the frame. The program is careful about what that does and does not prove: the two readings come from the same underlying model, so their agreement is corroboration, not statistical independence, and it says so in writing. Whatever survives goes into a ledger where every entry is labelled with where it came from: seen in the image, read off printed text on the item, supplied by you in the seed words, or taken from your own catalog record. The part that writes prose is only allowed to use that ledger. It never sees the photograph. Then the draft is checked back against the ledger, sentence by sentence. A color is bound to the part it was actually seen on, so a white backdrop cannot authorize calling the product white. A stated count has to match the number of things that were actually located. A whole class of claims is permanently off limits from a picture alone: material, dimensions, edition, care, origin, certification, contents, price, stock, shipping, warranty and performance can never be inferred from appearance. If a sentence cites a fact but attaches the wrong value to it, it fails. There is exactly one repair attempt, isolated from the first response, so the second try cannot simply echo the first one's mistakes. If the evidence still cannot support 75 non-repetitive words, the program refuses to write. Refusing is a correct result, not a breakdown. Accessibility is written into the contract rather than added afterward. At least two sentences have to describe the physical object or serve accessibility directly, in an order that is useful to someone listening rather than skimming. Alt text is produced at 12 to 40 words and is bound to the same facts as the body. Exactly one or two closing sentences are allowed to sell, and only using benefits that follow from facts already stated. That constraint is enforced by a check, not by a style guide someone might ignore. Naming a specific product is deliberately hard to earn, and today it is narrower than the rest of the machinery. An exact identity is released only when a checksum-revalidated barcode resolves to one record in your own catalog, or printed text on the item matches one catalog title strictly enough, or your seed words match one record exactly, and sources that point at different records fail closed. Even then, exact release is currently limited to physical books and Bibles; other object families stop at a supported family or type and cannot use record-level catalog prose. A guess about the exact item made from the picture alone is kept as private review information and is never printed in the description. And this is where the honest limit sits: reliably getting from an arbitrary product photograph to the right object family is the part that is not solved yet. The writing machinery, the evidence rules, the validators and the review step are built and covered by more than a thousand automated checks. The recognition step in front of them has not yet cleared the accuracy bar the program set for itself, and the program says so in writing rather than shipping around it.
Everything in the current release.
Each of these is built and working today. Nothing on this list is a roadmap item.
One photo, up to nine words, one description
The input is deliberately tiny: a single product image and one to nine words of seed text from you. Zero words, or ten and up, is rejected outright rather than quietly truncated. The output is one description between 75 and 350 words, or an explicit refusal. There is no setting that widens that contract, which is what makes results from different runs comparable.
The photo is read twice, and the readings never see each other
Two readings of the same image are taken over the same ordered inputs, each prompted differently and each sealed off from the other's answer. A detail survives only when both readings report it and place it in the same region of the picture. Disagreement, uncertainty, or an ambiguous match gets the fact thrown out rather than averaged. Because both readings come from the same underlying model, the program calls their agreement corroboration and explicitly declines to call it statistical independence.
A ledger that remembers where every fact came from
Facts are not pooled into one bucket. Each entry is kept with its source: seen in the image, read from printed text on the item, given by you in the seed, drawn from your approved catalog record, or flagged uncertain. The part that writes prose can only use approved ledger entries. Because the sources stay separate, you can look at a finished draft and see which sentence rests on which kind of evidence.
It refuses rather than pads
If the surviving evidence cannot support 75 words of non-repetitive, grounded prose, no description is produced. Nothing is invented to reach the floor and nothing is repeated in different words to bulk it out. In one measured run across eight fresh photographs of drinkware, every single case was set aside rather than described. That was recorded as a result, not hidden.
Claims it is structurally not allowed to make
Material, dimensions, edition, care instructions, origin, certification, included contents, price, stock, shipping, warranty and performance can never be inferred from appearance. They can appear only when you supply them. This is enforced by a checker that reads the finished draft, not by a hopeful instruction in a prompt.
Colors stay attached to the part they were seen on
A color observed on one component is not allowed to migrate to another component, and a color seen on a background or a display stand cannot become a claim about the product. A white backdrop does not make a white mug. The same discipline applies to counts: a stated number of anything must match the number of distinct places that thing was actually located in the frame.
Written to be listened to
At least two sentences must describe the physical object or serve accessibility directly, in an order useful to someone hearing the page rather than scanning it. Alt text is generated at 12 to 40 words from the same facts as the body, so the short version and the long version can never disagree. The draft has to pass an accessibility check as well as a factual one before it is released for review.
One or two selling sentences, at the end, and earned
Sales language is confined to exactly one or two closing sentences and must follow from benefits derived from facts already stated in the draft. It cannot open the description, it cannot be sprinkled through it, and it cannot introduce a claim of its own. Everything before it is description.
Naming the exact product has to be proved, and today only for books
An exact product identity is released only through a checksum-revalidated barcode that maps to one record, printed text that strictly matches one catalog title, or seed words that match one record exactly. The barcode checksum is recomputed rather than trusted, and sources that point at different records fail closed. That exact route is currently scoped to physical books and Bibles; other families stop at a supported family or type. A guess made from the picture alone stays private and never reaches the page.
Nothing leaves the machine it runs on
All analysis and writing happen on hardware you control. Inference is restricted to the local machine, and adding a remote or paid outside service is prohibited by the rules the project is built under rather than discouraged by a setting. Your product photographs, your seed words and your drafts stay where they were created.
Receipts that prove what happened without keeping what was said
Every run leaves a record of content fingerprints, pinned versions, fixed outcome codes, word counts and timing. It does not keep your image, your seed text, the generated description, or the raw model exchange. You can prove that a specific run happened, with which inputs and which result, without that audit trail becoming a second copy of your content.
Studies are frozen before they are run
Each measurement run declares its population, its thresholds and its pass conditions in advance, then becomes immutable once it completes. Failed candidates are kept as evidence with their exact numerators and denominators rather than quietly retuned until they look better; several are recorded verbatim as immutable and rejected without tuning or rerun. Forty distinct numbered study documents are on file, with the numbering running through V83, including the ones that did not work.
A workspace where a person accepts or rejects each draft
A sign-in protected review page is built and running, where a reviewer sees the image beside the words and rules on factual alignment, clarity, natural flow, repetition, usefulness to a screen-reader listener, the closing sales sentences, and any claim that looks unsupported or contradicted, with severity attached. Unsupported claims are logged by category rather than as free text. Nothing is treated as approved because a machine produced it.
It cannot reach your store
There is no publishing, syncing, inventory, pricing, cart or storefront administration anywhere in it, and that boundary is written into the rules the project is built under. The last thing it produces is a local draft flagged for review. Putting approved words on a live product page is a separate, deliberate act by a separate system.
Numbers we can stand behind.
Every figure below comes from the product's own release record or test suite, not from a marketing estimate.
Numbers we can stand behind.
It runs on hardware you control and has no route to an outside service, so photographs of unreleased products never leave the building.
Surfaces and status.
Status as of 2026-08-31. , 'Integrate Product Vision research through V83', dated 2026-08-31. The frozen study checkpoint of that date records the failing result that ended the merchandising-category approach and closes with a six-step list of what is still required, ending with the statement that production activation requires a separate staged acceptance and rollback checkpoint. The project's own running status note of 2026-08-29 describes the work as checkpointed local development, not promoted or deployed. Declared package version is 0.7.0. The writing, evidence, validation and review machinery is built and covered by 1,074 passing automated checks as of 2026-08-30, and that checkpoint states in its own words that it is integration evidence and not recognition-accuracy, prose-quality, production or deployment evidence. There is no public address and no released service; the only running surface is a sign-in protected internal review page. Forty distinct numbered study documents, running through V83, are frozen on the record, including the failures.
Worth more together.
Products on this platform share one sign-in, one support queue, and one engineering standard. These pair naturally with ASL Product Vision.
ShopFit
ShopFit is the storefront product this research is aimed at. The project states the division plainly: ShopFit would own catalog selection, the reviewer workflow and publication, while Product Vision would only ever hand over a local draft and a validation receipt.
ASL AI Hub
If what appeals here is AI that runs on hardware you control rather than sending your material to somebody else's service, AI Hub is the available product built on that same principle.
Viavitna
Viavitna is self-hosted: it runs on your own hardware, on your own network, and nothing leaves it. It applies the same discipline to answering questions instead of describing objects, keeping answers tied to your source material and declining rather than guessing.
Recent progress.
This product ships often. The most recent verified changes, newest first.
Recent progress.
2026-08-23 an object specialist reached 22 of 32 correct families on a frozen 32-image sample and the first automatically bridged draft was accepted at 82 words. On.
Recent progress.
2026-08-24 a leaner version produced an accepted 76-word draft, and the accessibility-and-identity architecture was defined. Through.
Recent progress.
2026-08-28 a run of identity studies tested more than twenty distinct methods on fresh, non-overlapping populations; none passed every condition and none was promoted, and several are recorded verbatim as immutable and rejected without tuning or rerun. On.
Recent progress.
2026-08-30 the recognition work reached 135 of 136 family precision on its own population, the description path was connected end to end with 1,074 automated checks passing and zero failures, and a 100-item review set was put in front of a reviewer to exercise the workflow, using already-verified catalog summaries rather than machine output. That same day a composed run over fresh photographs released none of eight supported cases, an honest failure to carry over to new pictures. On.
Recent progress.
2026-08-31 a store-aware study asked whether a photograph matched the department its catalog record already claimed; it reached 11 of 67 and was frozen as failed, establishing that merchandising departments are not visually coherent object classes. The next steps are written down: build a reviewed visual object-family list covering books and Bibles, candles, apparel, mugs and tumblers, jewelry, wall art, cards, figurines, bags and tableware and train against that instead; keep fit and evaluation lanes separated by product and vendor; and only after those gates pass, generate real drafts and put those machine drafts in front of human reviewers.,.
Recent progress.
Pricing for ASL Product Vision is quoted after a short conversation about your situation, because the right scope differs from one team to the next. There is no charge for that conversation.
Can I buy this or try it today?
No. Product Vision is an active research program, not a released product. The writing, evidence and validation machinery is built and covered by more than a thousand automated checks, but the step that recognizes what is in an arbitrary photograph has not yet met the accuracy bar the program set for itself. Rather than release it early and let it guess, the work stays on the record as a measurement history, including what failed.
What does it cost?
Nothing is published, because nothing is for sale yet. No plan, tier or per-description price exists. If you want to be told when that changes, ask through the contact page and you will get a straight answer about where the work actually stands rather than a waitlist position.
We already use an AI writing tool. Why would we need this?
If your current tool is producing copy you do not have to check against the photograph, keep it. The difference here is what happens when the picture does not support a sentence: most tools write it anyway, and this one refuses. That refusal is the point of the whole design. It is also why this is still research rather than a purchase, because refusing correctly turns out to be easier than recognizing correctly, and the recognition half is not finished.
Would my product photos or catalog leave my building?
No. Everything runs on hardware you control, and inference is restricted to the local machine, with remote and paid outside services prohibited by the rules the project is built under rather than by a setting someone could flip. Run records keep content fingerprints, pinned versions, outcome codes, word counts and timings, and deliberately do not keep your images, your seed words, your drafts or the raw model exchange.
Would it ever change something on my live store?
No. There is no publishing, syncing, pricing, inventory or storefront administration anywhere in it. The last thing it produces is a draft on local disk flagged as requiring review. Getting approved words onto a live product page is a separate deliberate act by a separate system, and that boundary is a written rule for the project rather than an oversight that could be patched around.
Can it identify exactly which product is in my photo?
Only when that can be proved, and today only for some things. An exact identity needs a checksum-revalidated barcode that maps to one record, printed text that strictly matches one catalog title, or seed words that match one record exactly, and conflicting sources fail closed. Even then the exact route is currently limited to physical books and Bibles; other object families stop at a family or type and cannot use record-level catalog prose. A guess from the picture alone stays private and never reaches the page.
What happens to my descriptions if the project stops or I stop paying?
The output is plain text on your own machine, produced from your own photographs and your own catalog records, and it is yours whatever happens next. There is no hosted store of your copy to lose access to and nothing that stops working when a subscription lapses, because the descriptions are files, not entries in somebody's service.
How would moving to this work if I already have hundreds of descriptions?
It does not replace what you already have. It is meant to fill the gaps, the products with an empty field or a one-line vendor blurb, one product at a time, with a person approving each draft before it goes anywhere. Nothing gets rewritten in bulk and nothing gets overwritten, because the program cannot reach your store to overwrite anything.
How do I know the numbers you publish are real?
Every measurement run declares its population and its pass conditions before it runs, then becomes immutable when it finishes, with exact numerators, denominators and per-category cuts kept intact. Failures stay on the record: the most recent study reported 11 of 67 correct and 5 of 11 conditions passed, and was stopped rather than adjusted until it looked better. Forty distinct numbered study documents are on file and most did not pass. This is the company's own measurement work, not an outside audit, and it is described that way.
What should I know before I rely on it?
We would rather you hear this from us than discover it later. As of 2026-08-31:
What should I know before I rely on it?
This is a research program, not a product you can buy or sign up for today. There is no public address, no plan, and no date being promised.
What should I know before I rely on it?
The open problem is recognition. Getting from an arbitrary product photograph to the correct object family is not solved. In the most recent frozen study, raw category top-1 was 11 of 67 and only 1 of 67 cases released, so that approach was recorded as failed and stopped rather than tuned.
What should I know before I rely on it?
Naming an exact product is narrower than the rest of the system. Exact record-level release is currently limited to physical books and Bibles; candles, stoles, candle wicks and miscellaneous objects stop at a supported family or type and cannot use record-level catalog prose at all.
What should I know before I rely on it?
There is no accepted real-photograph success rate for descriptions. Individual grounded drafts have been produced and accepted in single measured cases, most recently at 76 and 82 words, but no percentage across a real catalog has been earned.
What should I know before I rely on it?
The sign-in protected review page that is running today holds already-verified catalog summaries, not machine-generated drafts. The project says explicitly that this page is not the test of generated output, and that human review of actual machine drafts is a step still to be run.
What should I know before I rely on it?
Speed is not there yet. Measured end-to-end runs have taken roughly twenty to forty minutes for a single image. A usable version needs to run in seconds or low minutes, and that gap is stated as a requirement rather than glossed over.
What should I know before I rely on it?
The strongest retrieval numbers on record, including 141 of 145 top-1 recall, come from transformed copies of the same source photographs. They measure how well it copes with edits to one image, not recognition of a product photographed again with a different camera.
What should I know before I rely on it?
No independently photographed, product-labelled evaluation set exists yet. The program has written down exactly what one would need for its first category: approximately 150 distinct products, one clean reference image each, and three to five phone photographs per product across devices, lighting, perspective, backgrounds, hands and partial occlusion.
What should I know before I rely on it?
Some public image collections used in study work carry licence terms that have not been cleared for commercial use. The project states in its own words that neither of the two sources involved is declared commercially cleared by it.
What should I know before I rely on it?
The bar it has set for itself is deliberately high and not yet met: at least 95% top-1 recall and 99% top-5 on covered identities, at least 98% precision on visible facts with region provenance, unsupported factual claims under 1%, zero prohibited claims, and at least 80% of drafts accepted without substantive rewriting.
Stop shipping product copy that lies about the product.
Prefer email? contact@autosecurelogin.com