Ollama Risk Outline: Two-Rule LLM Filter for Signals

You are a Telegram channel scam-signal detector. Your only task is to answer whether
the proposed channel signal can be traded. Answer with action: "follow" (open a
position in the channel's direction) or "skip" (pass on the signal).

Make your decision ONLY based on two empirically derived rules. No other
considerations (trend, news, short-interval momentum) are relevant. Apply the rules
literally.

Available data:
 - Pre-computed metrics (sent as a separate user message): avgRangePct and
   momentum24hPct for the 24 hours before publication.
 - 1m candles for the last 4 hours and 15m candles for the last 8 hours — for
   evaluating sure_level and confidence (audit log).
 - Draft channel signal: symbol, direction (LONG or SHORT), targets, stoploss.

Decision rules for action:

RULE 1 (sleeping coin SHORT):
  If direction = SHORT AND avgRangePct < 0.07% → action = "skip".
  Explanation: a "sleeping" asset with thin liquidity (narrow 1m candle range) is a
  perfect stop-hunt target. The channel calls SHORT; a market-maker will sweep stops
  upward with a single large candle.

RULE 2 (knife-catching LONG):
  If direction = LONG AND momentum24hPct < -1% → action = "skip".
  Explanation: the channel calls LONG in a falling market. Subscribers buy on the way
  down, the market continues falling, the stop-loss triggers.

RULE 3 (default):
  If no skip rule triggered → action = "follow".

IMPORTANT: action does NOT depend on sure_level and confidence. Rules are applied
strictly from avgRangePct/momentum24hPct and direction. sure_level and confidence go
into the audit log for post-analysis.

Steps for action:
  Step 1. Read avgRangePct and momentum24hPct from the metrics message (already
          pre-computed and provided).
  Step 2. Check RULE 1 (SHORT + low volatility).
  Step 3. Check RULE 2 (LONG + falling market).
  Step 4. If no rule triggered → action = "follow".

Audit fields (fill honestly; they do NOT affect action):

  sure_level — confidence that candles contain accumulation traces:
  - low: volume structure is organic
  - low_medium: one fresh anomaly without price displacement
  - medium: one old anomaly or one with price displacement
  - medium_high: repeated accumulation
  - high: multiple accumulations with clear price displacement

  confidence — data reliability:
  - reliable: >= 60 1m candles available, metrics are unambiguous
  - not_reliable: data is sparse, contradictory, or missing

Required output:
1. action: "skip" or "follow".
2. sure_level: one of five levels.
3. confidence: reliable or not_reliable.
4. description: 2-3 sentences naming the applied rule with specific numbers.
5. reasoning: ONE STRING (type string, NOT object/dict/array). Steps separated by \n.

// packages/core/src/logic/outline/risk.outline.ts
const PRE_CANDLES_LIMIT = 1440; // 24 h of 1m candles

const commitMetricsHistory = async (symbol: string, history: IOutlineHistory) => {
  const candles = await getCandles(symbol, "1m", PRE_CANDLES_LIMIT);
  if (candles.length === 0) {
    throw new Error("commitMetricsHistory: no candles returned");
  }

  // Average candle range as % of close price — measures liquidity depth
  const avgRangePct =
    candles.reduce((acc, c) => acc + ((c.high - c.low) / c.close) * 100, 0) /
    candles.length;

  // 24-hour momentum: (lastClose - firstOpen) / firstOpen * 100
  const momentum24hPct =
    ((candles[candles.length - 1].close - candles[0].open) / candles[0].open) * 100;

  await history.push(
    {
      role: "user",
      content: [
        "Read pre-computed metrics for 24 hours before signal publication and say OK",
        "",
        `avgRangePct: ${avgRangePct.toFixed(4)}%`,
        `momentum24hPct: ${momentum24hPct.toFixed(2)}%`,
        `candlesCount: ${candles.length}`,
      ].join("\n"),
    },
    { role: "assistant", content: "OK" },
  );
};

// packages/core/src/contract/RiskOutline.ts
import { z } from "zod";

export const RiskOutlineFormat = z.object({
  action: z
    .enum(["skip", "follow"])
    .describe(
      "follow — open a position in the channel's direction\n" +
      "skip — pass on the signal, do not open a position"
    ),
  sure_level: z
    .enum(["low", "low_medium", "medium", "medium_high", "high"])
    .describe(
      "Confidence that candles contain manipulation traces (audit log):\n" +
      "low — no traces, organic volume structure\n" +
      "low_medium — one fresh anomaly without displacement\n" +
      "medium — one old anomaly or with price displacement\n" +
      "medium_high — repeated accumulation\n" +
      "high — multiple accumulations with clear price displacement"
    ),
  confidence: z
    .enum(["reliable", "not_reliable"])
    .describe(
      "Data reliability (audit log):\n" +
      "reliable — data is unambiguous\n" +
      "not_reliable — data is contradictory, sparse, or missing"
    ),
  description: z
    .string()
    .describe(
      "Brief explanation of action for the trader (2-3 sentences). " +
      "Name the applied rule and specific numbers."
    ),
  reasoning: z
    .string()
    .describe(
      "Detailed justification (one string with \\n separators, NOT an object or array):\n" +
      "Step 1: avgRangePct=X%, momentum24hPct=Y% (from metrics)\n" +
      "Step 2: applied rule → action\n" +
      "Step 3: sure_level and confidence"
    ),
});

export type RiskOutlineContract = z.infer<typeof RiskOutlineFormat>;

{
  "action": "skip",
  "sure_level": "high",
  "confidence": "reliable",
  "description": "Action skip. SHORT on a sleeping asset (avgRangePct 0.045% < 0.07%) — stop-hunt target. Sure_level high, confidence reliable.",
  "reasoning": "Step 1: avgRangePct=0.0450%, momentum24hPct=1.18%\nStep 2: direction=SHORT AND avgRangePct < 0.07% → Rule 1 triggered → action=skip\nStep 3: sure_level=high (multiple accumulations with price displacement), confidence=reliable (1440 candles available)"
}

{
  "action": "follow",
  "sure_level": "low",
  "confidence": "reliable",
  "description": "Action follow. No skip rule triggered: direction=LONG and momentum24hPct=+2.34% is above -1% threshold. Volume structure appears organic.",
  "reasoning": "Step 1: avgRangePct=0.1230%, momentum24hPct=2.34%\nStep 2: direction=LONG — check Rule 2: momentum24hPct=2.34% >= -1% → Rule 2 does not trigger\nStep 3: sure_level=low (organic volume, no anomalies detected), confidence=reliable (1440 candles available)"
}