> ## Documentation Index
> Fetch the complete documentation index at: https://documentation.3cubed.vc/llms.txt
> Use this file to discover all available pages before exploring further.
# Connections
> Links you draw, connections Tendril infers from shared threads and meetings, and how duplicates are found.
Tendril has two kinds of relationship, and it keeps them separate:
| | Drawn links | Inferred connections |
| ---------------- | ----------------------------------------------------------------- | ----------------------------------------------------------- |
| **Comes from** | You, using **Link to someone…** on a contact | Mail threads, meetings, group chats and shared work domains |
| **Has a kind** | `knows` · `worked with` · `introduced by` · `family` · `investor` | No. It has a score, a confidence level and a channel. |
| **Stored** | In `tendril.json` | Recomputed on every index rebuild, never saved |
| **In the graph** | Solid lines | Dashed lines |
Inferred connections are never turned into links automatically. The kind of a link is your judgement, and guesses don't belong in data you curated. You make a suggestion a link yourself, and choose its kind when you do.
## How a connection is inferred
Every source produces *co-occurrences*: occasions when several people other than you showed up together. Tendril looks at each pair of contacts in a co-occurrence and adds weight to the connection between them.
| Co-occurrence | One per | Convener |
| ----------------------------- | ---------------------------------------------------------------- | -------------------------------- |
| Mail thread | Thread per month, with 2 to 12 other people including the sender | The sender, unless it was you |
| Calendar event | Occurrence, with 2 to 12 other attendees | The organizer, unless it was you |
| iMessage or Beeper group chat | Month the chat had messages, with 2 to 12 other members | Nobody |
| Shared work domain | Domain, with 2 to 12 addresses on your cards | Nobody |
### The weight
Each pair in a co-occurrence gets:
```text theme={"dark"}
weight = 1 / (n − 1) × channel weight × 0.5 ^ (months ago / 15)
```
Here `n` is the number of other people present. This is the weighting used for co-authorship networks. A three-way thread is good evidence that two people know each other, and a twelve-way thread is very weak evidence.
`n` includes people who don't have a card. A twelve-person meeting is still weak evidence even if only two of the attendees are in your address book.
| Channel | Weight | Why |
| ------------------ | ------ | --------------------------------------------------------------------------- |
| Calendar | 1.0 | A meeting means they spoke. |
| Shared work domain | 0.7 | Strong at a small company. The group-size factor reduces it at larger ones. |
| Email | 0.6 | Being cc'd together only means they were both nearby. |
| Group chat | 0.5 | |
Weight halves every **15 months**. Without decay, a project that ended in 2019 would outrank your current work forever. Shared work domains are dated at rebuild time and don't decay, because working somewhere is an ongoing fact, not a single event.
For a sense of scale: a recent meeting with two other people scores **1.0** for that pair. A twelve-way cc thread scores about **0.05**. Both halve every 15 months.
## Confidence
A raw score only means something next to other scores, and showing "0.83" on a contact card would suggest more precision than the evidence has. So Tendril shows three levels instead:
| Level | Score | Means |
| ------------ | ----------- | ------------------------------------------------------------ |
| **Strong** | 1.5 or more | Together repeatedly, in small groups or recently |
| **Likely** | 0.4 or more | In the same small group more than once |
| **Possible** | Below 0.4 | Seen together, but not much: big threads, or a long time ago |
A pair seen together on **only one occasion** is hidden. Two people cc'd on one thread haven't been shown to know each other.
Shared work domains are the exception, since working somewhere doesn't happen twice. A shared domain alone is enough if its score reaches **0.15**, which in practice only a small employer does. Five addresses on a domain give 0.175, which passes. Six give 0.14, which doesn't. Larger shared domains, up to 12 addresses, still add weight to pairs that have other evidence. Being cc'd once *and* working at the same company counts for more than either one alone.
Personal email providers and ISPs never count as a shared domain. See [Sources](/tendril/sources#shared-work-domains).
## Introduced by
If two people's **earliest** shared occasion was started by a third person, Tendril shows that person as having introduced them (`via …`).
* Only mail threads and meetings have someone who started them. Group chats and shared domains never produce an introduction.
* Threads you sent and meetings you organized don't count. You connected those people, not a third party.
* Converting a suggestion into an `introduced by` link doesn't keep who made the introduction yet. See [Limits](/tendril#limits).
## Where connections appear
### On a contact
The **Connections** card shows counts first, so you can see how many of your contacts someone knows without opening anything: **linked**, then **strong**, **likely** and **possible**. Click a count to see the people in it. Each inferred row shows the other person, who introduced them if known, and how they are connected: `met`, `cc`, `same company` or `group`.
### The Connections screen
A 3D force-directed graph of everyone involved in a link or suggestion. Stronger connections pull people closer together. Selecting a contact highlights them and their neighbours and shows their faces.
* The graph draws up to **160** suggestions, with stronger ones drawn more boldly. The layout supports up to 400 contacts.
* The **Suggested** list below the graph shows the top 12 suggestions that aren't links yet, each with a menu to make it a link.
* Uncheck **Suggested** to see only the links you drew.
### Checking the weighting
Your own links are the only ground truth available, so Tendril checks against them after every rebuild. The Connections screen shows a line like:
```text theme={"dark"}
14/17 manual links found, 9 in the top 10%, median rank 38 of 2,412
```
If your own links don't rank near the top, the weighting is wrong. Use this line when tuning the numbers. They're all defined together in `EdgeScore` in `Interactions.swift`.
## Duplicates
The **Duplicates** screen groups cards that share any of these:
| Key | Normalized as |
| --------- | ---------------------------------------------------------------------------------- |
| **Name** | First and last name with case and accents ignored, keeping only letters and digits |
| **Email** | Lowercased and trimmed |
| **Phone** | Last 10 digits. Numbers with fewer than 10 digits are ignored. |
Cards are joined transitively (union-find). If A and B share an email and B and C share a name, all three are in one cluster. A cluster that shares an email or phone is **high confidence** and listed first. A shared name alone isn't. Turn on **Only high-confidence** to hide name-only clusters.
Merging happens in Contacts.app. Tendril only finds the duplicates.