Methodology

Companies are resolved against the SEC's own ticker registry, which maps a ticker to a CIK — the identifier the SEC uses for a filer. That CIK is what every filing on a company page is fetched by, so a relationship is always attached to a specific legal filer rather than to a name that happens to match.

Funds are resolved from their N-PORT filings, which carry the fund's CIK and, where one exists, its EDGAR series identifier. Single-fund trusts such as SPY and DIA file under a registrant CIK directly and have no series id.

Holdings inside a fund filing are matched to a US ticker either by the identifier the filing itself supplies or by normalising the holding's name against the SEC ticker registry. A holding that matches neither keeps its name and is shown without a link rather than being dropped or guessed at. Multiple share classes of one issuer are summed into a single position.

For each company we take the most recent annual report — a 10-K, or a 20-F for foreign private issuers — and slice out the sections where commercial relationships are actually disclosed: Business, Risk Factors (where customer and supplier concentration is normally stated), and Management's Discussion and Analysis. Filings that do not follow the standard item layout are read in overlapping full-document chunks instead, so an unusual format loses coverage rather than silently losing the company.

A language model reads those sections and proposes edges. It is not trusted. Every proposed edge must clear all of the following before it is stored, and anything that fails is discarded rather than downgraded:

• The citation must be verbatim. The sentence the model returns has to appear literally in the filed text. This is a string containment check against the filing itself, and it is the single gate that makes a fabricated quote impossible to publish.
• Confidence floor. The model reports its own confidence in each extraction and anything below 0.8 is dropped.
• One party, not a category. Groups — “customers in Europe”, “our top ten customers”, “original equipment manufacturers” — describe who a company sells to in general, not a disclosed relationship with a counterparty, and are rejected.
• Tickers are validated, never inferred. A counterparty ticker is checked against the live SEC registry; if it does not resolve, the ticker is removed and the counterparty keeps only its disclosed name.
• Anonymised counterparties need materiality. Filings often say “one customer accounted for X% of revenue” without naming it. Those are kept only when the disclosed share is at least 25%, so an unnamed party has to be genuinely significant to earn a place on the map.

Direction is stored from the perspective of the filing. The subject of an edge is the company whose filing disclosed it, and the relationship type describes the counterparty's role relative to that subject. So when Apple's 10-K names a supplier, the edge is stored on Apple with the counterparty marked as a supplier.

The reverse view — seeing that same relationship on the supplier's own page — is derived when the page is read, and it is labelled as disclosed by the counterparty. Any revenue share attached to the original disclosure is deliberately dropped in that direction: a percentage disclosed in one company's filing describes that company's books, not the other party's, and carrying it across would silently invent a number nobody disclosed.

Filing

Directly supported by a cited public filing. The edge carries the verbatim sentence, a link to the filing document on SEC.gov, and the filing's date. Every relationship published today is filing-sourced.

Call · Q&A

Reserved for relationships disclosed in dated earnings-call transcripts. Varamir does not ingest transcripts today, so no published edge carries this tier.

News

Reserved for relationships reported by a named publisher with a source link and date. No published edge carries this tier today.

The last two tiers exist in the data model and in the interface's vocabulary, and they are how the graph will grow. They are listed here so that the labels you see on the product mean what they say: a badge reading FILING is not one option among three in use, it is the only kind of evidence currently behind any edge.

Confidence

The extraction model's own reported confidence in having read the sentence correctly, floored at 0.8 by the gate above. It measures reading, not truth: it is not a probability that the commercial relationship still exists, and it is not a statistical estimate.

Weight

A disclosed share of revenue or cost of goods sold, shown only when the filing states one. Most relationships are disclosed without a figure, and those carry no weight rather than an inferred one.

ETF exposure

A fund's position in a stock as a fraction of net asset value, taken from that fund's N-PORT filing, shown with the report period it belongs to.

Every relationship shows the date of the filing that disclosed it. That date is the honest answer to “how current is this?” — annual reports are annual, so a relationship disclosed in a 10-K reflects the company's position as of that filing, not as of today. Varamir does not re-verify edges independently between filings, and does not claim to.

Fund holdings lag further. Regulatory portfolio reports are quarterly and are published roughly sixty days after the period they describe, so the holdings shown are the most recent ones a fund has publicly reported — not what it holds right now. Every holdings figure is labelled with its report period for that reason.

Earnings dates carry their own distinction: a date a company announced itself is marked confirmed, and a date estimated by a market-data provider is marked estimated. The two are never presented alike, and where they disagree the company's own announcement wins.

• The relationship vocabulary is deliberately narrow: customer and supplier relationships disclosed in filings, plus fund holdings. There are no competitor, partner, subsidiary or joint-venture edges.
• A relationship only exists here if a company disclosed it in writing. Filings disclose what regulation and materiality require, so absence of an edge is not evidence that no commercial relationship exists.
• Counterparties that are private companies, foreign issuers without a US listing, or anonymised in the filing have no Varamir page to link to.
• Nothing on Varamir is investment advice, and no figure here is a prediction. Ranges derived from options are described as market-implied, which is a statement about what option prices currently imply, not about what will happen.

Quote pages and Earnings Rooms are public. A visitor with no account can read 5 ticker pages, or browse for 5 minutes, whichever comes first, with 2 of those pages being Earnings Rooms; after that Varamir asks you to sign in. Nothing on a page is hidden or truncated while you are within that allowance — it is a limit on how many companies a stranger can read, not a partial view of any one of them. Search-engine and AI crawlers are exempt from it entirely.

A free account raises that to 50 ticker views a month outside 5 starred symbols. Full terms are on the pricing page.

Entity resolution and extraction both make mistakes, and a wrong edge on a public page is worth fixing quickly. The process for issuers, funds and readers is on the corrections page. Where each field comes from is listed on sources.