Bot: choose a way forward on screens nothing covers (instead of waiting for a human) #35

Open
opened 2026-11-19 13:20:30 +00:00 by kathleen · 1 comment
Owner

Problem. When the bot reaches a screen where none of its sources has a pick, it stops and waits for a manual choice through ntfy. Its sources are:

  • the Atlas route model;
  • the PDF guide rules in src/strategy.js;
  • learned rules.

A run can sit idle for hours, and an account's attempt is held up.

Where this happens (DragonKind/DragonKind bot._resolveSceneDecision):

  • In atlas_policy mode, the order is: burn protection → Account Lock Guard → Atlas advisor → PDF guide. When none of them covers the screen, nothing picks.
  • auto_learn already ends in learned rules → guide → Ollama AI → ntfy ping.
  • manual_notify pings on every new fork by design.

Goal

A decision step for uncovered screens that:

  • always picks something, so the bot never hangs. A human answer through ntfy can still override within a time limit;
  • never risks bonding a non-target dragon: the Account Lock Guard still has the final say;
  • logs how it chose (decisionMode, rule, propensity), so the Atlas can learn from these steps without bias (BOT_DATA_PROTOCOL.md).

Options under discussion

  1. Jev decision model (Kathleen's proposal): feed it the screen's offered choices and the run's context, and let it select. To spec: what Jev is, where it runs, and what it returns.
  2. Match to known answers. Fuzzy-match the offered texts against what the Atlas already knows:
    • the guide's encounter data (data/public/pdf_encounters.json, recommended answers);
    • the bot's screen tree (data/public/tree.json: answers that got through, answers that died);
    • the option-text crosswalk (graph.js).
  3. The existing LLM advisor (src/ollama.js), with a constrained prompt: the choices plus the matches from option 2. It must return one index and a confidence.
  4. Explore with a known probability. On research accounts, pick uniformly among the safe options and log propensity = 1/n. These are the most useful steps for the Atlas, because they're randomised.
  5. Timeout. Ping ntfy as today, but take the fallback after N minutes without an answer.

Most likely the answer is a chain of these. The order needs deciding: goal accounts and research accounts may differ.

Atlas side

  • Publish a small "screen advice" file the bot can query by option text: options 2 and 3 in one place, built from pdf_encounters.json and tree.json.
  • Learn from these steps: they arrive with decisionMode and propensity like every other step.
**Problem.** When the bot reaches a screen where none of its sources has a pick, it stops and waits for a manual choice through ntfy. Its sources are: - the Atlas route model; - the PDF guide rules in `src/strategy.js`; - learned rules. A run can sit idle for hours, and an account's attempt is held up. **Where this happens** (DragonKind/DragonKind `bot._resolveSceneDecision`): - In `atlas_policy` mode, the order is: burn protection → Account Lock Guard → Atlas advisor → PDF guide. When none of them covers the screen, nothing picks. - `auto_learn` already ends in learned rules → guide → Ollama AI → ntfy ping. - `manual_notify` pings on every new fork by design. ## Goal A decision step for uncovered screens that: - always picks something, so the bot never hangs. A human answer through ntfy can still override within a time limit; - never risks bonding a non-target dragon: the Account Lock Guard still has the final say; - logs how it chose (`decisionMode`, `rule`, `propensity`), so the Atlas can learn from these steps without bias (BOT_DATA_PROTOCOL.md). ## Options under discussion 1. **Jev decision model** (Kathleen's proposal): feed it the screen's offered choices and the run's context, and let it select. *To spec: what Jev is, where it runs, and what it returns.* 2. **Match to known answers.** Fuzzy-match the offered texts against what the Atlas already knows: - the guide's encounter data (`data/public/pdf_encounters.json`, recommended answers); - the bot's screen tree (`data/public/tree.json`: answers that got through, answers that died); - the option-text crosswalk (`graph.js`). 3. **The existing LLM advisor** (`src/ollama.js`), with a constrained prompt: the choices plus the matches from option 2. It must return one index and a confidence. 4. **Explore with a known probability.** On research accounts, pick uniformly among the safe options and log `propensity = 1/n`. These are the most useful steps for the Atlas, because they're randomised. 5. **Timeout.** Ping ntfy as today, but take the fallback after N minutes without an answer. Most likely the answer is a chain of these. The order needs deciding: goal accounts and research accounts may differ. ## Atlas side - Publish a small "screen advice" file the bot can query by option text: options 2 and 3 in one place, built from `pdf_encounters.json` and `tree.json`. - Learn from these steps: they arrive with `decisionMode` and `propensity` like every other step.
Author
Owner

Decided (Kathleen, 2026-11-19):

  • The model is Clef on Ollama Cloud (clef:cloud), instead of "Jev". It's a decision model: given a state (text, JSON, even a screenshot) and a typed choice question, it returns a probability for each option. It's called through /v1/systemone; cloud input is capped at 64K tokens, and clef-flash is a faster variant.
  • Goal accounts always take the safest choice, or the one most likely to lead to their target colour. Clef supplies the judgement, including what the books suggest earns a bond.
  • Research accounts may explore.

Proposed chain for an uncovered screen:

  1. Known answers. Fuzzy-match the offered texts against the guide's encounter data (pdf_encounters.json) and the bot's screen tree (tree.json). A confident match is used directly (rule: KNOWN_ANSWER, propensity 1).

  2. Clef. Send one choice question: the options are the on-screen choices, and the state includes:

    • the scene text;
    • the target colour, and the route so far;
    • the known answers and their outcomes, as evidence;
    • a short brief on what tends to earn a bond in the books (respect without submission, courage, honesty, not fleeing, and so on);
    • the instruction: survive, and stay on course for the target colour.

    What each account then does:

    • Goal accounts take Clef's highest-probability option (rule: CLEF, propensity 1).
    • Research accounts sample from Clef's probabilities mixed with uniform, for example 70/30, and log the exact probability of the option taken as propensity. This explores while staying sensible, and keeps the data unbiased for the Atlas.
  3. Account Lock Guard always has the final say. If it blocks Clef's choice, take the next most probable safe option.

  4. ntfy as an override, not a blocker: ping with Clef's pick and its probability. A human answer within N minutes wins; otherwise the bot goes ahead.

  5. Log Clef's full probability list on the step (private log only), so its calls can later be compared with outcomes.

Where it lives: mostly in DragonKind/DragonKind:

  • a src/clef.js advisor beside src/ollama.js;
  • a step in bot._resolveSceneDecision for every mode;
  • settings for the endpoint, key, mix rate and timeout.

The Atlas side publishes a "screen advice" file the bot can use for step 1.

To confirm: Ollama Cloud plan and key for clef:cloud (a paid plan, billed per token). Ken already has an Ollama Cloud key in his bot's integrations.

**Decided (Kathleen, 2026-11-19):** - **The model is [Clef](https://ollama.com/library/clef)** on Ollama Cloud (`clef:cloud`), instead of "Jev". It's a decision model: given a state (text, JSON, even a screenshot) and a typed `choice` question, it returns a probability for each option. It's called through `/v1/systemone`; cloud input is capped at 64K tokens, and `clef-flash` is a faster variant. - **Goal accounts** always take the safest choice, or the one most likely to lead to their target colour. Clef supplies the judgement, including what the books suggest earns a bond. - **Research accounts** may explore. **Proposed chain for an uncovered screen:** 1. **Known answers.** Fuzzy-match the offered texts against the guide's encounter data (`pdf_encounters.json`) and the bot's screen tree (`tree.json`). A confident match is used directly (`rule: KNOWN_ANSWER`, `propensity 1`). 2. **Clef.** Send one `choice` question: the options are the on-screen choices, and the state includes: - the scene text; - the target colour, and the route so far; - the known answers and their outcomes, as evidence; - a short brief on what tends to earn a bond in the books (respect without submission, courage, honesty, not fleeing, and so on); - the instruction: survive, and stay on course for the target colour. What each account then does: - **Goal accounts** take Clef's highest-probability option (`rule: CLEF`, `propensity 1`). - **Research accounts** sample from Clef's probabilities mixed with uniform, for example 70/30, and log the exact probability of the option taken as `propensity`. This explores while staying sensible, and keeps the data unbiased for the Atlas. 3. **Account Lock Guard** always has the final say. If it blocks Clef's choice, take the next most probable safe option. 4. **ntfy** as an override, not a blocker: ping with Clef's pick and its probability. A human answer within N minutes wins; otherwise the bot goes ahead. 5. **Log** Clef's full probability list on the step (private log only), so its calls can later be compared with outcomes. **Where it lives:** mostly in DragonKind/DragonKind: - a `src/clef.js` advisor beside `src/ollama.js`; - a step in `bot._resolveSceneDecision` for every mode; - settings for the endpoint, key, mix rate and timeout. The Atlas side publishes a "screen advice" file the bot can use for step 1. **To confirm:** Ollama Cloud plan and key for `clef:cloud` (a paid plan, billed per token). Ken already has an Ollama Cloud key in his bot's integrations.
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: DragonKind/threshing-atlas#35