AI Research Loop

// Queue message flow:
// [2h cron, every 6h] → seshat-research-queue → [research worker]
//                                                       ↓
//                                              gather context (Bankr + Dexscreener)
//                                                       ↓
//                                               Claude single completion
//                                                       ↓
//                                           R2: reports/{address}/{timestamp}.json
//                                           D1: research_reports row → re-score token

// Research trigger criteria:
//   market_cap_usd > 1000
//   AND no research_report in last 24h
//   ORDER BY volume_24h_usd DESC LIMIT 20

// Manual trigger: POST /api/admin/trigger-research

const SYSTEM = `You are Seshat, an on-chain intelligence agent specializing in
crypto project authenticity analysis. Evaluate whether this project is
genuine or a LARP (fake project with no real technical substance).

Output ONLY valid JSON matching the ResearchReport schema. No prose.
Be precise and conservative. Missing data is not evidence of fraud,
but it does lower confidence. Do not invent GitHub repos or team members.`;

// Context passed per token:
const prompt = `
Token: ${symbol} / ${name}
Contract: ${address} (Base)
Website: ${websiteUrl ?? "none"}
Market cap: $${mcap?.toLocaleString() ?? "unknown"}
Volume 24h: $${vol24h?.toLocaleString() ?? "unknown"}
Age: ${ageDays} days since launch

Developer (feeRecipient wallet: ${feeRecipientWallet}):
  Category:        ${devCategory ?? "unknown"}
  On-chain score:  ${onchainScore ?? "not yet scored"}
  Background:      ${backgroundScore ?? "not yet scored"}
  Smart followers: ${smartFollowers ?? 0}
  KOL associations:${kolNames.join(", ") || "none"}
  Contract verified: ${contractVerified ? "yes" : "no"}
  Ownership renounced: ${ownershipRenounced ? "yes" : "no"}

Evaluate and output JSON.
`;

// research_reports schema — key LARP fields:
is_larp:          INTEGER    -- 1 if Claude determined it's a LARP
larp_probability: INTEGER    -- 0-100 (not 0.0-1.0)
confidence:       TEXT       -- 'low' | 'medium' | 'high' | 'very_high'

// How larp_probability affects composite score (in scoring.ts):
// larp_probability >= 75  → tech_research_score = max(2,  raw_score * 0.1)
// larp_probability >= 50  → tech_research_score = max(10, raw_score * 0.4)
// larp_probability < 50   → tech_research_score = raw_score (no penalty)

// Tokens with larp_probability > 50 are never sent as alerts
// regardless of composite score.

// Novelty scale:
// 80–100: Genuinely novel — new mechanism, new use case, no close comparables
// 60–79:  Some novelty — variation on existing theme with differentiation
// 40–59:  Low novelty — common concept (e.g. generic "AI agent" token)
// 0–39:   Clone / direct copy of existing project

// prior_art field: structured list of similar existing projects
prior_art: [{
  name:       string,   // e.g. "BarnBridge"
  url:        string,   // e.g. "https://barnbridge.com"
  similarity: number,   // 0-100
  launch_date: string,
}]

// R2: full report JSON (one per research run — multiple runs accumulate)
const r2Key = `reports/${tokenAddress}/${Date.now()}.json`;
await env.REPORTS.put(r2Key, JSON.stringify(fullReport), {
  httpMetadata: { contentType: "application/json" },
});

// D1: latest summary row (upserted on each run)
-- research_reports table
CREATE TABLE research_reports (
  id               TEXT PRIMARY KEY,
  token_address    TEXT NOT NULL,
  is_novel_tech    INTEGER,           -- 1 = Claude assessed genuine novelty
  novelty_score    INTEGER,           -- 0-100
  is_larp          INTEGER,           -- 1 = LARP detected
  larp_probability INTEGER,           -- 0-100
  is_leading       INTEGER,           -- 1 = ahead of existing solutions
  leadership_score INTEGER,           -- 0-100
  traction_score   INTEGER,           -- 0-100
  notability_score INTEGER,           -- 0-100
  prior_art        TEXT,              -- JSON: PriorArt[]
  github_signals   TEXT,              -- JSON: GitHubSignals
  traction_signals TEXT,              -- JSON: TractionSignals
  summary          TEXT,              -- 2-3 sentence summary
  key_findings     TEXT,              -- JSON: string[]
  confidence       TEXT,              -- 'low' | 'medium' | 'high' | 'very_high'
  model_used       TEXT,              -- e.g. 'claude-sonnet-4-6'
  tokens_used      INTEGER,
  r2_key           TEXT,              -- pointer to full report in R2
  generated_at     TEXT DEFAULT (datetime('now'))
);