The model was right 99% of the time. The other 1% was an allergen.
Accuracy is a single number, and single numbers hide the thing that matters. Getting a colour name wrong costs you a return. Getting an ingredient wrong costs you a recall. No catalog should govern both with the same threshold.
This page is about what happens to a value between a supplier’s PDF and your storefront — where it came from, who approved it, and what stops it when the evidence is thin.
The number we actually optimise for.
Attributes populated across governed classes.
R1–R2. Review concentrates where risk does.
Any past state of the catalog is reconstructible.
Ninety-nine percent of what, exactly?
Take a modest catalog: forty thousand SKUs, eighty-four attributes each. That is 3.36 million published values. At 99% accuracy — a figure most teams would be delighted with — thirty-three thousand of them are wrong.
Almost all of those are harmless. A slightly wrong shade name, a lifestyle tag that does not quite fit, a category synonym nobody searches. You could ship those errors for a decade and never notice.
A small number of them are not harmless at all. A pressure rating on a valve. A voltage on an industrial drive. A CE mark that was never actually issued. An allergen matrix that was read from the wrong column.
Those live in the same 1%, and a single accuracy target treats them identically. That is the governance failure — not the error rate itself, but the decision to measure a catalog with one number when its attributes carry wildly different consequences.
Descriptive
Colour name, marketing copy, lifestyle tags, category synonyms
A return, occasionally a bad review.
Dimensional
Weight, pack size, carton dimensions, volumetric class
Shipping surcharges, mis-picks, carrier rejections at scale.
Compatibility
Fitment, thread spec, voltage, socket type, ACES/PIES linkage
The wrong part gets fitted. Returns, labour claims, liability.
Regulatory
UL, CE, RoHS, FCC, energy rating, pressure rating
An unsubstantiated compliance claim. Fines and delisting.
Safety & health
Allergens, ingredients, dosage, age grading, hazard class
Recall. Injury. In several jurisdictions, criminal liability.
These are defaults, not rules handed down. Every class threshold is yours to set, and most teams tighten R3 in their first quarter once they see which attributes their own returns data actually turns on.
Every value knows where it came from.
Not “the AI extracted it.” The document, the page, the table, the raw string before normalisation, and what happened next. Pick a value and follow it.
Industrial 4-Port PoE+ Gigabit Switch
120 W
unitCode: WTT
Vendor datasheet — MFR-9921-DS-rev4.pdf
Page 14, Table 2, row "System power budget"
Layout-aware document parse · table cell
"120W (30W max per port, IEEE 802.3at)"
120 → WTT · per-port 30 WTT recorded separately
Agrees with manufacturer product page, retrieved 12 Aug
99.8%
Auto-published · class R2 dimensional
Append-only · corrections create new entries · prior states remain queryable
“The model said so” has never once been accepted as evidence by a regulator, a court, or a customer.
Review is a routing decision, not a bottleneck.
Nothing reaches this queue because the model was unsure. Items arrive because a policy sent them, and every item states which one.
Industrial 4-Port PoE+ Gigabit Switch
Routed: R3 compatibility · below autonomous threshold for fitment data
Not populated
120 W total · 30 W per port · IEEE 802.3at
Hydraulic High-Pressure Ball Valve 1/2in
Routed: R3 compatibility · thread spec drives fitment
1/2" NPT (vendor label text)
1/2" NPT female × 1/2" NPT female · 6000 PSI WOG
Organic Cold-Pressed Matcha Powder 250g
Routed: R4 regulatory · every certification claim is reviewed
Natural green tea
USDA Organic · Non-GMO · Gluten-free · Vegan
The failure mode is a gap, not a guess.
This is the part that costs us in demos. Our catalogs have holes, and we will not fill them with plausible values. Six reasons a value is refused:
No source
The value cannot be traced to any supplied document. It is never inferred from a similar SKU.
Unit unresolvable
A number arrives with no unit and no context sufficient to infer one. 0.285 stays empty rather than becoming grams.
Source conflict
Two documents disagree and neither is authoritative. Both are surfaced to a human; neither is published.
Vocabulary miss
The extracted value has no valid member in the controlled vocabulary for that attribute class.
Schema failure
The emitted record fails channel or Schema.org validation. The SKU publishes without the attribute, never with a broken one.
Stale evidence
The source document is older than the attribute's configured evidence window. Re-extraction is queued.
A missing attribute appears on your gap report with the exact document you need to request, from which supplier, to close it. That is a worse demo and a considerably better catalog.
Your catalog was correct when you built it.
Governance projects tend to treat data quality as a state you reach. It is not. A supplier revises a spec sheet, changes a material, moves production to a second facility, or quietly drops a certification — and tells nobody, because from their side nothing happened worth an email.
Your record stays exactly as accurate as it was on the day it was built, which is to say it silently becomes wrong. Nothing in a conventional PIM will ever detect this, because the value never changed. The world did.
Median share of attributes out of date at any moment.
Cadence set per attribute risk class.
Regulatory and safety re-verify most aggressively.
A changed value raises a drift event, never a quiet edit.
The questions you’re about to ask.
You keep saying zero hallucinations. Every vendor says that.
So judge it structurally rather than on the claim. Extraction is not generative — a value is written only if it resolves to a span in a supplied document, and the span is stored with it. There is no code path that produces an attribute from model priors alone. That design has a real cost, which is the honest tell: our catalogs have gaps. An attribute nobody documented stays empty and appears on your missing-data report. A system that never returns empty is guessing.
Our compliance team will want to audit this. What can we actually show a regulator?
For any published value, at any point in its history: the source document, the page and table it came from, the raw string before normalisation, the transformation applied, the confidence, the policy class that governed it, whether a human approved it, which human, and when. The record is append-only — corrections are new entries, so the state of the catalog on any past date is reconstructible. That is usually the question a regulator actually asks.
Doesn't human review destroy the throughput advantage?
It would if everything were reviewed, which is why the risk classes exist. In a typical catalog, roughly 88% of attribute values fall into R1 and R2 and publish autonomously. Review load concentrates in R3 through R5, which is a few percent of values but nearly all of the actual exposure. The point of the taxonomy is to spend human attention where being wrong is expensive, rather than spreading it evenly and thinly across everything.
What happens when a supplier changes a specification and doesn't tell us?
This is the failure mode nobody plans for, because the catalog was correct when it was built. Every source is re-checked on a schedule set per attribute class, and a changed value raises a drift event rather than silently overwriting. Regulatory and safety attributes are re-verified most aggressively. In our audits, the median enterprise catalog carries between 4% and 9% silently-stale attributes at any moment, concentrated in exactly the classes that matter.
Who is accountable when an approved value turns out to be wrong?
You are, and we will not pretend otherwise — it is your catalog and your regulatory exposure. What the system changes is that accountability becomes traceable instead of diffuse. Today a wrong claim usually cannot be attributed to a decision at all. With a lineage record you can establish whether the source document was wrong, the extraction was wrong, or the approval was wrong, and each of those has a different and actionable remedy.
Find out what you are currently claiming.
Give us one category — ideally a regulated one. We will trace every published compliance and safety attribute back to its source document and hand you the list of claims you are making that nothing in your own files substantiates.
One category · Read-only · Findings in ten working days