AI Research Loop

// Queue message flow:
// [composite scorer] → seshat-research-queue → [research worker]
//                                                      ↓
//                                               Claude agentic loop
//                                                      ↓
//                                              R2: full report JSON
//                                              D1: research_reports row

// Research is skipped if the token was already researched in the last 24h
const RESEARCH_TTL_HOURS = 24;

const SYSTEM = `You are Seshat, an on-chain intelligence agent specializing in
crypto project authenticity analysis. Your task is to evaluate whether
a project is genuine or a LARP (Live Action Role Play — fake project
with no real tech).

Output ONLY valid JSON matching the ResearchReport schema. No prose.
Be precise. If uncertain, say so — do not hallucinate facts.`;

const schema = `{
  "larp_probability": number,          // 0.0–1.0
  "larp_reasoning":   string,          // 1–2 sentences
  "novelty_score":    number,          // 0–100
  "novelty_reasoning":string,
  "github_authentic": boolean | null,
  "github_notes":     string | null,
  "prior_art":        string[],        // similar existing projects
  "tech_summary":     string,          // 2–3 sentences on the tech
  "red_flags":        string[],
  "green_flags":      string[]
}`;

const prompt = `
Token: ${symbol} (${name})
Description: ${description}
Website: ${website}
GitHub: ${githubUrl ?? "none"}
Creator wallet: ${creatorAddress}
Creator FrontrunPro score: ${devScore}
Market cap: $${mcap.toLocaleString()}
Volume 24h: $${vol24h.toLocaleString()}
Token age: ${ageDays} days

${githubData ? `GitHub commit history (last 20):\n${githubData}` : ""}

Similar tokens found by name:\n${similarTokens.join("\n")}

Evaluate and output JSON.
`;

// Final LARP probability = weighted combination
function finalLarpProbability(
  aiScore: number,    // Claude's 0.0–1.0 estimate
  rules: RuleSignals
): number {
  const ruleScore = (
    (rules.noGithub       ? 0.15 : 0) +
    (rules.noWebsite      ? 0.10 : 0) +
    (rules.genericName    ? 0.10 : 0) +
    (rules.clonedCode     ? 0.35 : 0) +
    (rules.lowDevScore    ? 0.15 : 0) +
    (rules.zeroDayLaunch  ? 0.10 : 0)  // volume spike on day 1, then nothing
  );

  // 60% AI + 40% rules
  return Math.min(1.0, aiScore * 0.6 + ruleScore * 0.4);
}

// Tokens with larp_probability > 0.5 are soft-filtered from alerts
// Tokens with larp_probability > 0.75 are hard-filtered and flagged

// Novelty factors considered by Claude:
// - Prior art: does a token with identical concept already exist?
// - GitHub: original code vs. forked repo with no changes?
// - Description: template-looking or unique narrative?
// - Technical depth: does the description reference real tech?

// Scale:
// 80–100: Genuinely novel — new mechanism, new use case
// 60–79:  Some novelty — variation on existing theme
// 40–59:  Low novelty — common concept, some differentiation
// 0–39:   Clone / direct copy

// R2: full report (preserves all Claude output)
await env.R2.put(
  `research/${tokenAddress}.json`,
  JSON.stringify(fullReport),
  { httpMetadata: { contentType: "application/json" } }
);

// D1: summary row (queryable)
await db.prepare(`
  INSERT INTO research_reports (
    id, token_id, larp_probability, novelty_score,
    github_authentic, tech_summary, red_flags, green_flags,
    r2_key, model_version, researched_at
  ) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
`).bind(
  id, tokenAddress,
  report.larp_probability, report.novelty_score,
  report.github_authentic ? 1 : 0,
  report.tech_summary,
  JSON.stringify(report.red_flags),
  JSON.stringify(report.green_flags),
  `research/${tokenAddress}.json`,
  "claude-sonnet-4-6",
  Math.floor(Date.now() / 1000)
).run();