#!/usr/bin/env python3 """T15 EXPLORATORY (2026-07-23) — SHARP HALF-LIFE: how long does a wallet stay sharp after we detect it? The live follow set has principled entry criteria and no exit criteria; this measures whether detection-day edge decays with age-in-set and how fast wallets churn out of the published screen. Detection dates are REAL as-of history, not reconstruction: the git commit series of live/watch_sharps.json (published daily by the pipeline since 2026-06-18) — a wallet's detection date is the first commit that contains it. Wallets already present in the FIRST commit are left-censored (true detection unknown) and excluded from the age curves. Performance is the wallet's OWN bets from the cache (live/cache.duckdb bets, read-only snapshot taken first so the payouts writer never contends), graded to CHAIN TRUTH per bet via payouts.ensure/truth — the cache's won/res_t columns are never trusted (the res_t=ts poison). EV is per $100 at the wallet's own entry price, feeless: this measures SIGNAL decay, not our execution (T3/T11 own execution). Outputs: EV/bet by age-in-set bucket (0-2d / 3-6d / 7-13d / 14d+), pooled AND per-wallet-day mean-of-means (concentration guard), plus the survival curve (fraction of detected wallets still published at age k). NOT pre-registered — informs a rotation policy for the follow set. VERDICT (2026-07-27, cached-only grading, COVERAGE 14%): NO EDGE DECAY. EV/bet by age-in-set: 0-2d +$13.75 (n=867) · 3-6d +$17.97 (n=839) · 7-13d +$22.24 (n=1,569) · 14d+ +$19.81 (n=2,406) — the FRESHEST bucket is the weakest and edge holds past two weeks. Set against the survival curve (49% of detections gone within 1 day, 11% at 28d), the churn is THRESHOLD JITTER around the z-cutoff, not wallets going stale. => Do NOT build an age-based rotation rule; age is not a decay signal. Caveats: 14% coverage (cached chain truth only — rises on its own as other nightlies fill the payouts cache; rerun later to confirm) and top-5 wallets carry 44% of the pooled total. METHOD NOTE: ensure() over every cond these wallets ever touched was a TREADMILL (84k conds, RPC-bound, count grows daily) — five runs across five days never finished it. Cached-only + honest coverage answered the same question in minutes. Pass --ensure to crawl the tail anyway.""" import collections import json import os import subprocess import sys import time HERE = os.path.dirname(os.path.abspath(__file__)) ROOT = os.path.dirname(HERE) sys.path.insert(0, os.path.join(ROOT, "live")) BAND = (0.05, 0.95) BUCKETS = [(0, 3, "age 0-2d"), (3, 7, "age 3-6d"), (7, 14, "age 7-13d"), (14, 10**6, "age 14d+")] def set_history(): """[(date_str, {wallets})] oldest-first from git history.""" log = subprocess.run( ["git", "log", "--reverse", "--format=%ad %H", "--date=short", "--", "live/watch_sharps.json"], cwd=ROOT, capture_output=True, text=True).stdout.split() pairs = list(zip(log[0::2], log[1::2])) out = [] for date, sha in pairs: try: blob = subprocess.run( ["git", "show", f"{sha}:live/watch_sharps.json"], cwd=ROOT, capture_output=True, text=True).stdout rows = json.loads(blob) ws = {r["wallet"].lower() for r in rows if r.get("wallet")} if ws: out.append((date, ws)) except Exception: continue # one snapshot per date (last commit of the day wins) byday = {} for date, ws in out: byday[date] = ws return sorted(byday.items()) def main(): hist = set_history() print(f"set snapshots: {len(hist)} days " f"({hist[0][0]} .. {hist[-1][0]})", flush=True) first_day, censored = {}, set() for i, (date, ws) in enumerate(hist): for w in ws: if w not in first_day: first_day[w] = date if i == 0: censored.add(w) pool = set(first_day) - censored print(f"wallets ever published: {len(first_day)} · " f"left-censored (first snapshot): {len(censored)} · " f"age-eligible: {len(pool)}", flush=True) day_n = {d: time.mktime(time.strptime(d, "%Y-%m-%d")) // 86400 for d, _ in hist} det_n = {w: day_n[first_day[w]] for w in pool} # survival: still-published at age k (right-censored by last snapshot) last_n = day_n[hist[-1][0]] print("\nSURVIVAL (still in the published set at age k):", flush=True) for k in (1, 3, 7, 14, 21, 28): elig = [w for w in pool if det_n[w] + k <= last_n] if not elig: continue alive = 0 for w in elig: # snapshot on-or-after detection+k (carry-forward between days) snap = None for d, ws in hist: if day_n[d] <= det_n[w] + k: snap = ws else: break alive += snap is not None and w in snap print(f" age {k:>2}d: {alive}/{len(elig)} = " f"{alive/len(elig):.0%}", flush=True) # bets snapshot (read-only, close before payouts writes the same file) import duckdb con = duckdb.connect(os.path.join(ROOT, "live", "cache.duckdb"), read_only=True) wl = ",".join(f"'{w}'" for w in pool) bets = con.execute(f""" SELECT lower(wallet) w, cond, asset, any_value(p::DOUBLE) p, min(ts) ts FROM bets WHERE lower(wallet) IN ({wl}) AND p BETWEEN {BAND[0]} AND {BAND[1]} AND resolved GROUP BY lower(wallet), cond, asset""").fetchall() con.close() print(f"\nresolved cache bets for age-eligible wallets: {len(bets)}", flush=True) import payouts # CACHED-ONLY grading (2026-07-27): ensure() over every cond these # wallets ever touched is a TREADMILL — 84k conds, RPC-bound, and the # count grows daily as they keep betting; five runs across five days # never finished it. The question ("does edge decay after detection") # does not need every historical bet: it needs an unbiased sample of # post-detection ones. So grade what chain truth is ALREADY cached and # report coverage; the cache fills from every other nightly anyway, so # coverage rises on its own without a dedicated crawl. conds = sorted({b[1] for b in bets if b[1]}) if "--ensure" in sys.argv: print(f"ensuring {len(conds)} conds against chain…", flush=True) payouts.ensure(conds) else: print(f"cached-only grading over {len(conds)} conds " f"(pass --ensure to crawl the uncached tail)", flush=True) per_bucket = {tag: [] for _, _, tag in BUCKETS} per_wd = {tag: collections.defaultdict(list) for _, _, tag in BUCKETS} graded = skipped = 0 for w, cond, asset, p, ts in bets: if ts is None or p is None: # cache rows with no entry time skipped += 1 continue if ts // 86400 < det_n[w]: continue # pre-detection bet pay = payouts.truth(cond, asset) if pay is None or pay == 0.5: skipped += 1 continue graded += 1 age = int(ts // 86400 - det_n[w]) ev = (100.0 / p) * (pay - p) for lo, hi, tag in BUCKETS: if lo <= age < hi: per_bucket[tag].append(ev) per_wd[tag][(w, int(ts // 86400))].append(ev) break cov = 100 * graded / max(1, graded + skipped) print(f"chain-graded post-detection bets: {graded} " f"(ungraded/refund skipped: {skipped}) · COVERAGE {cov:.0f}%" f"{' ⚠ low — treat buckets as indicative' if cov < 50 else ''}\n", flush=True) print("EV BY AGE-IN-SET (per $100 at the wallet's own entry, feeless):", flush=True) for _, _, tag in BUCKETS: evs = per_bucket[tag] if not evs: print(f" {tag:>10}: n=0") continue pooled = sum(evs) / len(evs) wd = [sum(v) / len(v) for v in per_wd[tag].values()] mom = sum(wd) / len(wd) nw = len({k[0] for k in per_wd[tag]}) print(f" {tag:>10}: n={len(evs):>5} · pooled EV/bet {pooled:+6.2f}" f" · wallet-day mean {mom:+6.2f} · {nw} wallets", flush=True) # concentration guard: top-5 wallet share of |pnl| overall by_w = collections.defaultdict(float) for _, _, tag in BUCKETS: for (w, _), v in per_wd[tag].items(): by_w[w] += sum(v) if by_w: tot = sum(by_w.values()) top = sorted(by_w.items(), key=lambda kv: -abs(kv[1]))[:5] print(f"\nconcentration: total {tot:+.0f} · top-5 wallets " f"{[(w[:8], round(v)) for w, v in top]}", flush=True) if __name__ == "__main__": main()