We publish what we read and what each instrument cannot see.
Most providers describe their sources in a way that cannot be checked, listing categories broad enough to include anything. We take the opposite position. This page sets out the eight classes of evidence behind every record, what each class reads, what each one cannot see, and the confidence band every data point carries, without naming a single source.
Every data point in a Zylabs record comes from one of eight classes of evidence and carries one of four confidence bands. We never read a company on its own: the suppliers, clients, subsidiaries and partners connected to it are read as well, because each of them publishes evidence about the company that it never publishes itself.
We never read a company on its own.
Every supplier, client, subsidiary and partner connected to a company is read as well, because each of them leaves evidence about the company that it never publishes itself. A supplier’s client list, a partner’s case study and a subsidiary’s filing all describe the account from the outside, and reading them together is what turns a thin profile into a complete picture.
Eight classes of evidence, each with a stated blind spot.
We publish the classes and we withhold the channels inside them, because the class tells you how much weight a line deserves while the channel list is the part a competitor would copy. A record states which classes were silent on the account you are looking at.
| Class of evidence | What we read | What it cannot see | |
|---|---|---|---|
| 01 | Compelled disclosure | Annual reports, statutory filings and investor materials. | Private companies and markets where disclosure is thin. |
| 02 | Additional filings required by governments | Government filings naming the employer, worksite, role and wage. | Countries where the filing is not mandated, and roles that do not require one. |
| 03 | What the company writes | Job specifications, engineering writing, conference talks and tender documents. | Roles filled without a public posting. |
| 04 | What its people write | Professional profiles and first-hand accounts of the work, from primary and secondary sources. | Built from a sample of the most relevant profiles, so it shows presence rather than volume. |
| 05 | Third-party compelled records | Enforcement actions, breach and layoff notices, court dockets and procurement awards. | Anything a regulator or court has not yet made public. |
| 06 | Corporate records | Registries, legal entity identifiers, group hierarchy and transactions, including for private companies. | Private firms below registration thresholds, with coverage that varies by region. |
| 07 | Technology and lifecycle | Technology fingerprints and vendor end-of-life calendars. | Anything that runs entirely behind the firewall and leaves no trace of its implementation. |
| 08 | Continuous web and news | Company announcements, reviews and archived page history. | Anything that was never published. |
Every data point carries one of four confidence bands.
The band says how a data point was produced and how much weight it can bear. A modeled figure is always a range, and when the methods behind it disagree by more than 30 percent its confidence is marked very low rather than averaged into a number that looks precise.
Most of what we read never reaches a record.
Volume is the least interesting number on this page, because reading is cheap and judgement is not. The figures below describe one account read in September 2026, and they are typical of a deep read rather than exceptional.
| Stage | One retailer, read 7 September 2026 |
|---|---|
| Public records screened for the supplier map | 12.6 million |
| Primary documents screened | 3,051 |
| Primary documents read in full | 813 |
| Open seats detected | 90 |
| Graded lines written to the record | 105 |
| Claims withdrawn since the first pass | 2 |
Much of what we read, the company wrote itself.
Classes three and four are primary evidence, meaning the company or the people who did its work produced the document for their own purposes rather than for a researcher. Classes one, two, five and six are compelled evidence, meaning a company or a third party had no choice about producing it. The two fail differently, which is why reading both matters more than reading a lot of either. A record built only on compelled evidence misses offshore delivery entirely, and a record built only on primary documents misses everything a company has never written about itself.
The most useful lines in a record are not in any source.
No channel reports that a supplier relationship is cooling, that a team is roughly forty people, that a program has moved from evaluation to build, or that a technology wallet sits between $863M and $2.6B. The model computes each of those from evidence read across classes, states the basis it used, and labels the result modeled wherever it appears.
| Derived line | Computed from | Stated as |
|---|---|---|
| Supplier presence and trend | Evidence density by year across classes one, two and four | Current, active, cooling or lapsed, with first and most recent evidence |
| Wallet band by function | Top-down revenue benchmark reconciled against bottom-up headcount and loaded cost | Modeled, with the benchmark named and where it misapplies |
| Team size | Record counts and document volume, which undercount | Modeled, as a range |
| Program phase | The language of the documents that name the program | Evaluating, building, migrating or established, with confidence |
| Says against does | Public commitments read beside hiring and supplier movement | The gap, named, with the evidence on both sides |
Each fact is re-read on its own clock.
| Fact type | Re-read |
|---|---|
| Disclosed financials | When the company next reports |
| Open seats and programs | Days |
| Supplier placements | Days to weeks |
| Stakeholder seats | Weeks, and a seat is current only while the role has no end date |
| Estate components | Weeks to months |
| Modeled wallet | When its inputs move |
Corrections are published, and so is what they take down with them.
When a line is found to be wrong it is corrected in the record with its date, and every claim that rested on it is withdrawn at the same time rather than left standing. Three seats shown as current in an earlier pass of the retailer record had left the company, so the record says so and withdraws the two claims that depended on them.
Common questions
Because a channel that cannot see a kind of company returns nothing for that company, and nothing reads exactly like a clean account. A reader who knows what an instrument misses can judge what a record is worth, and a reader who does not will eventually be misled by a gap they took for an absence.
The class tells you how much weight a line deserves, which is what you need to judge a record. The channel list is the part a competitor would copy, and it is most of what separates the Atlas from a database anyone can buy.
Because much of what matters about a company is published by someone else. A supplier names the account in its own case study, a subsidiary files what the parent never discloses, and a partner describes the program it was hired to deliver. Reading only the company itself leaves all of that out.
It is corrected in the record with its date, and every claim that rested on it is withdrawn at the same time rather than left standing. Corrections are published rather than quietly absorbed, because a provider that never publishes one is not more accurate than a provider that does.