## Labels (15 scores, in this order)

| #     | Label                                                           | Fact         | Kind                 |
|-------|-----------------------------------------------------------------|--------------|----------------------|
| 0-4   | `kind:question` `kind:discussion` `kind:support` `kind:troll` `kind:spam` | `kind`  | one of them (softmax) |
| 5-7   | `toxicity:0` `toxicity:1` `toxicity:2`                          | `toxicity`   | a level (softmax)    |
| 8     | `scam`                                                          | `scam`       | yes/no (sigmoid)     |
| 9     | `misleading`                                                    | `misleading` | yes/no (sigmoid)     |
| 10    | `respectful`                                                    | `respectful` | yes/no (sigmoid)     |
| 11-14 | `action:none` `action:warn` `action:delete` `action:block`      | `action`     | one of them (softmax) |

Definitions for labelers:

- **kind**
  - `question`: a genuine question about the project, its products or the ecosystem.
  - `discussion`: governance, roadmap, ecosystem or market talk in good faith, and
    praise, criticism or suggestions for the team, meant constructively.
  - `support`: a user problem: a wallet, stake, transaction, canister or login that does
    not work.
  - `troll`: bait, provocation, mockery, derailing, FUD without substance, "wen moon"
    noise: posted to provoke, not to inform or ask.
  - `spam`: promotion, shilling another project, off-topic ads, copypasta, repeated
    posting.
- **toxicity**: `0` civil; `1` rude (dismissive, mocking, crude); `2` abusive (insults,
  harassment, threats, slurs).
- **scam**: a fake airdrop or giveaway, "DM me", a request for a seed phrase or keys,
  impersonating the team, a phishing or wallet-drainer link, a pump group or a
  guaranteed-returns scheme ("guaranteed 3% daily").
- **misleading**: presents something false or unsupported as fact ("the devs rugged",
  "the canister was hacked", made-up numbers, wrong technical claims). A question or a
  clearly marked opinion is not misleading.
- **respectful**: treats the team and other members with respect, also in disagreement.
  Civil but dismissive or mocking is not respectful.
- **action**: what a moderator should do with this one message, in OpenChat terms:
  `none`, `warn` the user, `delete` the message, `block` the user. Block is for scams,
  abuse, and repeated trolling or spam; one off-color message gets a warn or a delete.

### Policy: how the facts combine

Labelers apply these rules, so that the same message always gets the same labels:

1. **A scam** is always `kind:spam`, `misleading` yes, `respectful` no, `action:block`
   (whatever its tone: a polite "DM me, I'm from support" too).
2. **Other spam** (ads, shilling, copypasta) is `respectful` no. A single post gets
   `action:delete`; repeated posting or copypasta gets `action:block`. `misleading` only
   when it makes false claims ("the fastest DEX on the IC, audited by everyone").
3. **Toxicity 2** (abuse aimed at people, threats, harassment) gets `action:block`.
4. **Toxicity 1:**
   - in a troll message, `action:warn`;
   - in a question, support request or discussion with a real point (an angry
     support request, sharp criticism), `action:none`.
5. **Misleading** (not a scam, not toxicity 2):
   - a troll's FUD or false accusation gets `action:delete`;
   - a good-faith mistake in a question, discussion or support message gets
     `action:warn`.
6. **Trolls without the above:**
   - harmless noise ("wen", "wen moon", "gm") gets `action:none`, `respectful` yes;
   - bait, mockery or derailing gets `action:warn`, `respectful` no;
   - repeated trolling gets `action:block`.
7. **respectful** is no for every message with toxicity 1 or 2, every spam and scam, and
   every mocking, sarcastic or baiting message. Sharp but fair criticism at toxicity 0 is
   respectful.
8. Everything else (genuine questions, support, discussion, criticism, warnings about a
   scam) gets `action:none`.

When two rules apply, the one that leads to the stronger action wins
(`block` > `delete` > `warn` > `none`).