From a story to a thread.

Reading the map

After you trace, the map works the answer out in front of you, one step at a time, the way you would on paper. Use the buttons under it to pause, step back, or jump to the answer.

• Indigo marks the question: the two people you chose, and a dashed arrow with a “?” between them.
• Gold arrows are facts from the story, numbered in the order the reasoning uses them. Each step quotes the sentence it came from.
• Red is the answer: the red thread between the two people that replaces the question.
• Shapes follow the pedigree charts used in biology: squares are male, circles female, and a diamond means the story never says.
• Rows are generations: parents above, children below, people of the same generation side by side.
• Arrows point from a person to their relative: Ashley → daughter → Lillian means Lillian is Ashley’s daughter.

Solve it like a student

Blood-relation questions are solved by combining two relations at a time. Each step on the map does exactly that, so you can check your own working against it. These combinations are worked out by the same rules the map uses:

Some combinations have more than one honest answer. A daughter’s uncle is usually your brother, but could be your brother-in-law; the map says “usually” and names the other possibility. When the steps the model wrote do not add up to its own answer, the map says so instead of hiding it.

Everything happens on your device

On your first visit the browser downloads the two models (61.4 MB) and keeps them. Each file is checked against a SHA-256 checksum before it is stored, and the decoded weights are checked again (CRC-32) every time they load. Later visits read the models from this device.

The models run in a background worker so the page never freezes. The arithmetic happens in a small engine written in C and compiled to WebAssembly, using SIMD instructions and several CPU threads when the browser allows. Your story is not sent to a server for inference.

Step 1: finding the people

The name model is a 5.12M-parameter GPT (4 layers) trained from scratch to decide, word by word, whether the word is a person’s name. A GPT only looks backwards, so it reads the text twice: while labelling the second copy, every word can see the whole sentence.

It learned from tens of thousands of stories with places, companies, months and holidays mixed in, so “Tokyo” or “Google” do not become people. It still makes mistakes; if it misses someone, use Missing someone? Add a name under the people.

Step 2: reading the story the way the model does

Before the relationship model sees the story, every name becomes an entity id: the first person is@0, the next @1, and so on. The model therefore learns relationships from the wording (“her son”, “his sister”), not from what particular names usually mean. The question comes first, then the story:

1. <bos>
2. <q>
3. @0
4. @2
5. <s>
6. @0
7. 's
8. daughter
9. ,
10. @1
11. ,
12. asked
13. her
14. mom
15. …
16. <c>

The relationship model is a 13.0M-parameter transformer with 6 layers, 6 attention heads and a 320-token context. Before answering, it writes the family facts along the path, a chain of thought: @0 daughter @1, then @1 brother @2, then the answer, son.

Step 3: searching for a path

The chain is found with a constrained beam search that keeps the 4 most likely partial paths. At each step the model may only pick a relation word and a person who appears in the story and is not already on the path, and the path must end at the person you asked about.

After every step the model also judges whether that fact is actually stated in the story. Paths built on facts it cannot verify lose weight. This step checker was trained on deliberately wrong chains, including convincing mistakes found by running the model’s own search.

Knowing when to say “I don’t know”

• No link in the text. If the two people are never mentioned together, or through people who are, Kinthread answers that the story doesn’t establish a relationship without running the model.
• No verified path. The model estimates how likely it is that the story gives no family link. At 0.50 or above, it says so instead of guessing.
• Disagreement. Agreement is the share of its weighted reasoning paths that end in the top answer. Below 0.70, the answer is shown as a tentative best guess.

Want the full story of how both models were built and trained? Read the build log. For what the models handle well and where they struggle, see limits & tips.