I've written artifacts/selection.json with exactly 210 different ids from the data. Each has a reason under 120 characters that cites the wallet's numbers, and the method summary is 284 characters. The file is untracked, as required. I checked the count, uniqueness and length limits with a script; there is no ground truth to check the choices themselves against.

Data: I downloaded the file to /tmp/dl/ and its sha256 matches the expected value. It holds 1951 profiles. No profile has a unit field.

How I chose: each wallet gets a score, and the 210 highest scores are chosen. The score is 40% history (age, number of transactions, how recently it was active), 40% IMD (seats count most, then IMD balance), and 20% ETH. ETH is on a log scale, so big balances help but aren't required. There's no randomness, and the ranking never uses id order.

The score is then cut for wallets that look like farms:

• younger than 90 days
• fewer than 10 transactions
• less than 0.002 ETH
• unknown age
• more than 20 transactions a day
• part of a group of near-identical profiles: same IMD holdings, age within 7 days, transactions within 10%, ETH within 0.002

What came out:

• 47 of the 63 wallets that hold IMD seats were chosen.
• No wallet from a near-identical group of 3 or more was chosen. There are 301 such wallets, mostly in one block of about 122-day-old wallets with roughly 110 transactions and 0.006 ETH. For example, 22 wallets are exactly 110 transactions, 122 days, 0.006 ETH.
• No wallet younger than 90 days was chosen.
• 150 of the chosen wallets hold no IMD; they got in on long history plus ETH.

Limits:

• The weights and penalties are my judgement, not fitted to known real or fake wallets.

• The last wallet in scored 0.4433 and the first one out scored 0.4432, so small weight changes would swap wallets near the line.

• Treating the 122-day group as one person is an inference. The data has no addresses or transfers to confirm it.

• Some IMD holders were left out:

  • New wallets that hold seats, such as W1865 (78 days old, 5 seats, 23.9 ETH).
  • Old wallets with almost no ETH, such as W1015 (5020 transactions, 2 seats, 0.001 ETH).

  They still enter the raffle.

Open questions:

• The file's own filter says the minimum is 0.002 ETH, yet 10 wallets have only 0.001.
• What is the unit token, given no profile has it?
• How were IMD seats handed out? That decides how much a seat says about the holder.

README.md has the full write-up, separating facts from inferences, uncertainty and open questions. scripts/select.py re-creates the selection and refuses to run on a file with the wrong hash. I committed those two files on a new branch, swarm-selection, using your email as the author. The data file is not in the repo.