Fantasy Smart Money League + Due-Diligence Copilot

Nansen Meridian Buildathon entry. A Telegram game where you draft proven smart-money wallets into weekly rounds settled on live Nansen PnL — plus a copilot that turns any token's smart-money flow into a graded 0–100 dossier in ~6 seconds.

• Play: @righttofight_bot on Telegram
• Watch: smart-money-league.fly.dev (convictions, leaderboard, track record, credit economy — refreshes every 20s)
• Stack: Nansen data → Jev (TypeSafe System One) triage + verdicts → Telegram game. LLMs are reserved for deep dives only.

Why

Token due-diligence bots are either slow LLM essays or raw dashboards. Nansen already is the dashboard, so another one can't win. What wins is a loop (weekly draft rounds), a memorable demo moment (the 6-second dossier), and data only Nansen has (smart-money labels, PnL leaderboards, fund flows).

The distinctive idea: the bot rations its own API budget. Every Nansen call burns credits, so a calibrated Jev judgment decides how deep to look before a single credit is spent — 1cr glance vs 10cr dossier — and every verdict is later graded against actual price. The system proves its own edge on the dashboard instead of claiming it.

How it works

flowchart LR
    U[Telegram user] -->|/deep mint| B[bot.py on Fly]
    B -->|guard: glance or dossier?| J1[Jev triage<br/>Score + Noul, 1 call]
    J1 -->|dossier| F[Nansen fan-out<br/>6 endpoints in parallel]
    F --> N[(Nansen API<br/>flows, buyers, sellers,<br/>flow-intel, OHLCV, holders)]
    N --> F -->|trimmed evidence| J2[Jev dossier<br/>5 questions, 1 call]
    J2 -->|scores + action| C[Code: 0-100 conviction<br/>weights + gates]
    C -->|stored| DB[(league.db<br/>dossiers, picks, rounds)]
    C -->U
    DB --> D[dashboard.py<br/>0-credit polling]
    D -->|/record grading| N

Loading

/check is the same pipeline at low cost (guard → 1–6cr → single verdict). /record replays stored dossiers against subsequent OHLCV closes into a hit-rate table — the validation loop anyone can re-run.

Commands

Command	Cost	What happens
/check <mint>	1–6cr	Jev guard → flows (1cr), +PnL leaderboard (5cr) if convicted → verdict + confidence
/deep <mint>	10cr	6-endpoint parallel fan-out → one 5-question Jev batch → 0–100 conviction, stored for grading
/record	1cr/dossier	Grades past dossiers vs subsequent price → hit-rate table
/board	5cr (cached 10m)	Top smart-money wallets with tap-to-copy /draft commands
/draft <wallet>	0 + 2cr scout	Drafts into the weekly round, then profiler bundle → Jev S/A/B/C scout grade
/picks, /settle	0 / 5cr	Your picks; anyone can settle the round on live PnL
/budget	0	Today's burn vs cap, per-endpoint breakdown

Nansen APIs used

Endpoint	Cost	Used for
smart-money-pnl-leaderboard	5cr	League settlement + /board — realized/unrealized PnL, win rate, trades
token flows	1cr	7d flow history (first-glance evidence)
token who-bought-sold (BUY + SELL)	1cr each	Labeled-buyer conviction: who accumulates, who distributes
token flow-intelligence	1cr	Per-segment net flow (smart-money slice)
token holders	5cr	Concentration / whale-risk evidence
token pnl	5cr	Per-token PnL leaderboard (deep /check)
token ohlcv	1cr	Dossier entry price + /record grading closes
token info	1cr	Sanity metadata
token screener	1cr	Discovering seed tokens by live volume
profiler pnl-summary	1cr	Draft scout: top tokens, realized PnL/ROI
profiler balance	1cr	Draft scout: current book
search general	0cr	Mint discovery (free)

Solana only — the CLI's Ethereum token paths hang, so the bot defaults to Solana and says so. Coverage varies by token (majors are rich, long-tail is thin); thin evidence flows into low Jev confidence, which is itself the signal.

Jev judgments (all typed, all batched)

One batched POST /v1/systemone (jev-latest) per decision — independent questions run in parallel, code owns thresholds and composition:

• Triage (triage_token): bullishness Score + worth_deep_dive Noul → plan is deep (guard ≥ 0.65, conf ≥ 0.5), glance (guard ≤ 0.35), else escalate to LLM teardown.
• Dossier (dossier): bullishness + accumulation + whale_risk Scores, dumping Noul, action Choice (accumulate/watch/avoid) → conviction = 40·bull + 30·acc + 20·safety + 10·action-confidence, gated: avoid/dump>0.7 caps at 25, watch caps at 55, avg-conf < 0.5 escalates.
• Scout (scout_wallet): consistency + degen Scores → S/A/B/C grade composed in code.
• API notes: Score/Choice take criteria (Score: ordered level array; Choice: {option: description} map) — levels 422s. Choice answers carry .choice + full distribution + confidence. judge() retries transient network blips with backoff.

Typed output guarantees the interface, not truth — hence /record.

Credit economy

Per-action caps (glance 1, deep 6, dossier 10, scout 2, record 4, settle 5), 100cr/day hard stop, 10-minute file cache so re-renders and dashboard polls burn nothing, every burn appended to .spend_log.jsonl. Observed: full dossier ≈ 10cr in ~5s; R&D to date ≈ 100cr of a five-figure balance. Seeded track record: STONK 25/avoid, PUMP 25/avoid, ETH 26/watch, HYPE 38/watch, CBBTC 40/watch (Sep 2026 — graded live via /record).

Project layout

• bot.py — Telegram handlers; every failure reports back, nothing hangs
• jev.py — TypeSafe client: triage, dossier, scout + conviction composer
• nansen_client.py — CLI wrapper: budget guard, TTL cache, parallel fan()
• league.py — sqlite: users, picks, rounds, dossiers
• dashboard.py + static/ — stdlib read-only web UI (LeagueDesk)
• seed_dossiers.py — batch dossier seeding Friday ritual
• Dockerfile, start.sh, fly.toml, DEPLOY.md — one-machine Fly deploy

Run locally / deploy

pip install -r requirements.txt
cp .env.example .env   # NANSEN_API_KEY, TELEGRAM_BOT_TOKEN
export TYPESAFE_API_KEY=$(cat ~/.config/typesafe/api_key)
python bot.py          # Telegram polling
python dashboard.py    # http://127.0.0.1:8765

Production is one Fly machine (bot + dashboard, sqlite on a volume) — see DEPLOY.md. Never run the local bot alongside Fly: two pollers cause Telegram getUpdates 409 Conflict.

Vision & roadmap

• Instant verdicts + credit guard + escalation ladder
• /deep dossier engine (6 endpoints → 5 judgments → conviction)
• Self-grading /record + live dashboard + Fly deploy
• Calibration report: 20+ graded dossiers → publish hit-rate by conviction bucket (the chart that wins arguments)
• Copy-trader detection: profiler related-wallets + counterparty overlap → Jev same-operator Noul → warn on correlated drafts
• EVM support once CLI paths stabilize (currently Solana-only)
• Bear-case mode: the copilot that talks you out of your bag, as a first-class command
• Real stakes: entry fees in USDC covering query costs onchain, payouts to round winners — the economy this points at

Honest limits

• Jev scores read cautious in current conditions (see seeds) — bias or market truth is undecided until /record accumulates grades.
• Thin-token evidence → low confidence by design; the bot says "watch", not "trust me".
• Free-tier Nansen + ngrok-grade links were the dev path; judging runs on Fly with paid credits behind it.