scoring: chain-truth resolutions overlay — refunds no longer count as wins

The cache's won (curPrice>=0.5 at pull) counts 50/50 refunds as wins for
BOTH sides — 521 of 2,128 chain-checked follow-set markets (24%) were
refunds, which is where the whales' 92-100% displayed win rates came from.

- live/payouts.py: resolutions table in cache.duckdb, filled from the CTF
  contract's payout vectors (batched+paced JSON-RPC, ~12 calls/s free tier;
  resolved rows immutable, unresolved recheck 6h, RPC failures never cached).
  truth(cond, asset) -> 1/0/0.5/None; refunds need no asset side.
- validate_timing: every displayed stat (conv/conv30/all-time/realized/copy
  replay) settles at truth; refunds count as neither W nor L, P&L is
  size*(wp-p)/p; new conv_ref/conv30_ref/all_ref feed fields.
- trust.conviction_record: optional truthfn — the selection gates
  (trust_wr/trust_roi) no longer select on refund inflation.
- portfolio.py: replay pays wp (refunds 0.5/share, was 1.0).
- conviction_scan: documented as the (refund-inflated) candidate layer;
  final selection re-judges against truth downstream.

Validation: 0x4bFb-whale conv 174-16 91.6% $1.61M -> 34-16 +140ref 68%
$214k, and truth-adjusted all-time P&L now sits within ~16% of lb-api's
PM P&L (was 7x apart); LSB1 (0 refunds) byte-identical, its P&L matches
PM P&L to 0.07%.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

live/conviction_scan.py

@@ -22,6 +22,14 @@ Profile gates (on TRAIN conviction bets):
   * median conviction stake >= MIN_MED_STAKE (dust wallets betting $2-$6 clips
     aren't followable and their fills aren't reproducible)
 Then validate forward and count how many keep the profile.
+
+NB (2026-07-06): this scan reads the cache's `won` marks, which count 50/50
+REFUNDS as wins for both sides (28% of resolved markets in the in-play niche)
+— so it OVER-generates candidates. That's acceptable: this is the candidate
+layer; final selection (validate_timing.py) re-judges every candidate against
+chain-truth payouts (payouts.py) and rejects refund-inflated records. Chain-
+checking all ~19M rows here would take days of RPC; the funnel does it only
+for the few dozen wallets that survive to the copy replay.
 """
 
 import math, os, time

live/payouts.py

@@ -0,0 +1,206 @@
+#!/usr/bin/env python3
+"""Chain-truth resolution overlay for cache.duckdb — the 2026-07-06 audit fix.
+
+The cache's `won` is `curPrice >= 0.5` at pull time, and that lies two ways
+(see the audit + [[polymarket-resolution-truth]] in project memory):
+
+  * 50/50 REFUNDS (walkovers/abandonments — 28% of the follow set's resolved
+    markets in the 30d sample!) leave curPrice at 0.5, so EVERY holder of
+    EITHER side gets won=True and (1-p)/p phantom profit. The whales' 92-100%
+    displayed win rates were largely this.
+  * operator-resolved in-play markets can leave a stale mark on the losing
+    side (one confirmed both-sides-won market), which resolved=TRUE certifies.
+
+The fix keeps the cache AS PULLED and overlays the on-chain truth: the CTF
+contract's payout vector per condition, plus the market's token order so a
+row's `asset` maps to its payout. Resolved payouts are immutable — fetched
+once, cached in a `resolutions` table forever. Unresolved conditions recheck
+after RECHECK_S.
+
+Usage:
+    import payouts
+    payouts.ensure(conds)              # batch-backfill (chain + CLOB, cached)
+    wp = payouts.truth(cond, asset)    # 1.0 / 0.0 / 0.5 / None (unknown)
+
+Scoring rule for consumers: truth 1/0 -> real win/loss; 0.5 -> REFUND, count
+as neither and P&L = size*(0.5-p)/p; None -> fall back to the cache's `won`
+(legacy NULL-asset rows, unresolved markets, RPC gaps).
+
+RPC: config.json `alchemy_key` (or ALCHEMY_RPC_URL env). Batched JSON-RPC
+(BATCH per POST); the CLOB market fetch supplies token order. Selection runs
+on the Mac, so this module is never needed by the Fly worker.
+"""
+import json
+import os
+import ssl
+import threading
+import time
+import urllib.request
+from concurrent.futures import ThreadPoolExecutor
+
+import cache
+
+_SSL = ssl._create_unverified_context()
+HERE = os.path.dirname(__file__)
+CTF = "0x4D97DCd97eC945f40cF65F87097ACe5EA0476045"
+SEL_DEN = "0xdd34de67"    # payoutDenominator(bytes32)
+SEL_NUM = "0x0504c814"    # payoutNumerators(bytes32,uint256)
+BATCH = 60                # JSON-RPC calls per POST
+PACE_S = float(os.environ.get("PAYOUTS_PACE_S", 5.0))   # sleep between batches:
+                          # free-tier Alchemy sustains ~12 eth_calls/s — an
+                          # unpaced backfill just trades 429 retries for gaps
+RECHECK_S = 6 * 3600      # re-ask the chain about unresolved conds after this
+
+cache.query("""CREATE TABLE IF NOT EXISTS resolutions(
+    cond TEXT PRIMARY KEY, p0 DOUBLE, p1 DOUBLE,
+    token0 TEXT, token1 TEXT, checked_at BIGINT)""")
+
+_mem = {}                 # cond -> row tuple, warm in-process copy
+_lock = threading.Lock()
+
+
+def _rpc_url():
+    url = os.environ.get("ALCHEMY_RPC_URL")
+    if url:
+        return url
+    try:
+        key = json.load(open(os.path.join(HERE, "..", "config.json")))["alchemy_key"]
+        return f"https://polygon-mainnet.g.alchemy.com/v2/{key}"
+    except Exception:
+        return None
+
+
+def _rpc_batch(calls):
+    """calls: [hex calldata] -> [int results or None], one batched POST."""
+    url = _rpc_url()
+    if not url:
+        return [None] * len(calls)
+    body = json.dumps([{"jsonrpc": "2.0", "id": i, "method": "eth_call",
+                        "params": [{"to": CTF, "data": d}, "latest"]}
+                       for i, d in enumerate(calls)]).encode()
+    # retry with backoff — Alchemy 429s a fast backfill after ~2 batches, and a
+    # failed batch must surface as None (skip-this-run), never "unresolved"
+    for attempt in range(4):
+        try:
+            req = urllib.request.Request(url, data=body,
+                                         headers={"Content-Type": "application/json"})
+            rs = json.loads(urllib.request.urlopen(req, timeout=30, context=_SSL).read())
+            if isinstance(rs, dict):                  # whole-batch error object
+                raise OSError(str(rs.get("error"))[:80])
+            by_id = {r.get("id"): r for r in rs}
+            out = []
+            for i in range(len(calls)):
+                v = by_id.get(i, {}).get("result")
+                try:
+                    # "0x" (revert/empty) must not kill the whole batch — one
+                    # bad item used to blank every cond in the chunk
+                    out.append(int(v, 16) if v and v != "0x" else None)
+                except (TypeError, ValueError):
+                    out.append(None)
+            return out
+        except Exception:
+            if attempt == 3:
+                return [None] * len(calls)
+            time.sleep(2 ** attempt)
+
+
+def _clob_tokens(cond):
+    """(token0, token1) in the CLOB market's outcome order — payout index order."""
+    try:
+        req = urllib.request.Request("https://clob.polymarket.com/markets/" + cond,
+                                     headers={"User-Agent": "Mozilla/5.0"})
+        m = json.loads(urllib.request.urlopen(req, timeout=15, context=_SSL).read())
+        toks = [str(t.get("token_id")) for t in (m.get("tokens") or [])[:2]]
+        return (toks + [None, None])[:2]
+    except Exception:
+        return None, None
+
+
+def _load(conds):
+    with _lock:
+        missing = [c for c in conds if c and c not in _mem]
+    if missing:
+        rows = cache.query(
+            "SELECT cond, p0, p1, token0, token1, checked_at FROM resolutions "
+            "WHERE cond IN (SELECT UNNEST(?::VARCHAR[]))", [missing])
+        with _lock:
+            for r in rows:
+                _mem[r[0]] = r
+
+
+def ensure(conds, workers=8):
+    """Make sure every cond has a resolutions row (fetching chain + CLOB for the
+    unknown ones). Resolved rows are permanent; unresolved recheck after
+    RECHECK_S. Safe to call with thousands of conds — everything is cached."""
+    conds = [c for c in {c for c in conds if c}]
+    _load(conds)
+    now = int(time.time())
+    with _lock:
+        todo = [c for c in conds
+                if c not in _mem
+                or (_mem[c][1] is None and now - (_mem[c][5] or 0) > RECHECK_S)]
+    if not todo:
+        return
+    # chain payouts, batched: den + num0 per cond (binary: num1 = den - num0)
+    payout = {}
+    for i in range(0, len(todo), BATCH // 2):
+        chunk = todo[i:i + BATCH // 2]
+        calls = []
+        for c in chunk:
+            h = c[2:].rjust(64, "0")
+            calls += [SEL_DEN + h,
+                      SEL_NUM + h + "0".rjust(64, "0")]
+        res = _rpc_batch(calls)
+        for j, c in enumerate(chunk):
+            den, n0 = res[2 * j], res[2 * j + 1]
+            if den is None:
+                payout[c] = "rpc-fail"      # do NOT cache — retry next run
+            elif den and n0 is not None:
+                payout[c] = (n0 / den, (den - n0) / den)
+            else:
+                payout[c] = (None, None)    # chain says: not resolved yet
+        if i + BATCH // 2 < len(todo):
+            time.sleep(PACE_S)
+    # token order for the RESOLVED ones (index -> asset mapping); cached forever
+    resolved = [c for c in todo
+                if isinstance(payout.get(c), tuple) and payout[c][0] is not None]
+    tokens = {}
+    with ThreadPoolExecutor(max_workers=workers) as ex:
+        for c, tk in zip(resolved, ex.map(_clob_tokens, resolved)):
+            tokens[c] = tk
+    with _lock:
+        for c in todo:
+            p = payout.get(c)
+            if p == "rpc-fail" or p is None:
+                continue                    # transient — leave uncached, retry next run
+            t0, t1 = tokens.get(c, (None, None))
+            row = (c, p[0], p[1], t0, t1, now)
+            _mem[c] = row
+            cache.query("INSERT OR REPLACE INTO resolutions VALUES (?,?,?,?,?,?)", list(row))
+
+
+def truth(cond, asset=None):
+    """Chain-truth payout of this position: 1.0 / 0.0 / 0.5 / None (unknown).
+    Refunds ([0.5,0.5]) need no asset; decided markets map asset -> index."""
+    _load([cond])
+    r = _mem.get(cond)
+    if not r or r[1] is None:
+        return None
+    p0, p1, t0, t1 = r[1], r[2], r[3], r[4]
+    if p0 == p1:                       # refund (or exotic even split)
+        return p0
+    if asset is not None:
+        if str(asset) == str(t0):
+            return p0
+        if str(asset) == str(t1):
+            return p1
+    return None                        # decided, but we can't map the side
+
+
+def stats():
+    n = cache.query("SELECT count(*), count(p0) FROM resolutions")[0]
+    return {"rows": n[0], "resolved": n[1]}
+
+
+if __name__ == "__main__":
+    print(stats())

live/portfolio.py

@@ -31,6 +31,7 @@ import urllib.request
 from concurrent.futures import ThreadPoolExecutor
 
 import cache
+import payouts
 import smart_money as sm
 import trust
 
@@ -179,24 +180,31 @@ def window_bets():
         rows = trust.trusted_wallet_rows(cache.query, w["wallet"], now)
         best = {}                            # one bet per market: largest-stake token
         for cond, asset, won, p, res_t, size in rows:
-            if cond not in best or size > best[cond][3]:
-                best[cond] = (won, p, res_t, size)
+            if cond not in best or size > best[cond][4]:
+                best[cond] = (asset, won, p, res_t, size)
         if w.get("class") == "whale":
             thr = 0.0
         else:
-            pre = [size for won, p, res_t, size in best.values() if res_t < START]
+            pre = [size for asset, won, p, res_t, size in best.values() if res_t < START]
             thr = cache.conv_cutoff(pre if pre else
                                     [size for *_, size in best.values()])
         _WALLET_THR[w["wallet"]] = thr
-        for cond, (won, p, res_t, size) in best.items():
+        for cond, (asset, won, p, res_t, size) in best.items():
             if size < thr or p > MAX_ENTRY:
                 continue
             et = ent.get(cond)
             if not et or et < START:                       # only in-window entries
                 continue
             out.append({"wallet": w["wallet"], "name": w["name"], "cond": cond,
-                        "cls": w.get("class", "volume"), "their": size,
-                        "entry_t": et, "p": p, "won": won, "res_t": res_t or 0})
+                        "asset": asset, "cls": w.get("class", "volume"),
+                        "their": size, "entry_t": et, "p": p, "won": won,
+                        "res_t": res_t or 0})
+    # chain-truth payouts for the replayed markets: refunds pay 0.5/share, and
+    # a cache `won` mark can be wrong on operator-resolved markets — the
+    # replay must settle at what a redeem actually pays (see payouts.py)
+    payouts.ensure({b["cond"] for b in out})
+    for b in out:
+        b["wp"] = payouts.truth(b["cond"], b.get("asset"))
     return out
 
 
@@ -301,7 +309,12 @@ def main():
         for ft, cost, payoff, rec in held:
             if ft and ft <= upto and rec["kind"] == "res":
                 cash += payoff; realized += payoff - cost; perW[rec["wallet"]]["realized"] += payoff - cost
-                perW[rec["wallet"]]["won" if rec["won"] else "lost"] += 1
+                wp = rec.get("wp")
+                won = rec["won"] if wp is None else wp > 0.5
+                perW[rec["wallet"]]["won" if won else "lost"] += 1
+                rec["won"] = won                      # truth-adjusted for the feed
+                if wp == 0.5:
+                    rec["refund"] = True
                 rec["pnl"] = payoff - cost
                 resolved.append(rec)
             else:
@@ -325,7 +338,11 @@ def main():
             cash -= cost; fees_paid += fee; perW[b["wallet"]]["bets"] += 1
             shares = stake / p_eff                        # lag-adjusted entry price
             if b["kind"] == "res":
-                payoff = shares * (1.0 if b["won"] else 0.0)   # redeem is fee-free
+                # chain-truth payout (1/0/0.5) when known, else the cache mark
+                wp = b.get("wp")
+                if wp is None:
+                    wp = 1.0 if b["won"] else 0.0
+                payoff = shares * wp                           # redeem is fee-free
                 held.append((b["res_t"] or now, cost, payoff, b))
             else:                                          # currently open -> mark to market, no free yet
                 held.append((None, cost, 0.0, b))
@@ -356,7 +373,10 @@ def main():
         stake = m.get("stake") or STAKE_MIN
         p_eff, fee, cost = entry_model(m["p"], stake)
         if "won" in m:
-            return (stake / p_eff) - cost if m["won"] else -cost
+            wp = m.get("wp")
+            if wp is None:
+                wp = 1.0 if m["won"] else 0.0
+            return (stake / p_eff) * wp - cost
         return stake * (m.get("cur", p_eff) / p_eff) - cost
 
     missed.sort(key=lambda m: m.get("res_t") or 0, reverse=True)

live/trust.py

@@ -103,17 +103,23 @@ def trusted_wallet_rows(runq, wallet, now=None):
         [wallet, now, now])
 
 
-def conviction_record(runq, wallet, days=90, pctile=0.80, now=None):
+def conviction_record(runq, wallet, days=90, pctile=0.80, now=None, truthfn=None):
     """Trailing trusted CONVICTION record for the held-edge gate:
     {n, wr, roi} over the wallet's top-(1-pctile) stake bets resolved in the
     last `days`. The conviction cutoff is that wallet's stake p80 over its FULL
     trusted history (matching cache.conv_cutoff semantics). roi is the flat-
     stake hold-to-resolution copy ROI per bet, fee/slip-free (gates compare it
-    to 0, and fees are already charged in copy_pnl, the other selection leg)."""
+    to 0, and fees are already charged in copy_pnl, the other selection leg).
+
+    truthfn(cond, asset) -> 1/0/0.5/None (payouts.truth): chain-truth payouts.
+    50/50 REFUNDS (28% of the follow-set niche's resolved markets!) then count
+    as NEITHER won nor lost, with roi credited at (0.5-p)/p — the cache alone
+    marks every refund won=True for both sides, which is exactly the inflation
+    these gates must not select on. None -> the cache's won (old behavior)."""
     now = int(now or time.time())
     rows = trusted_wallet_rows(runq, wallet, now)
     if not rows:
-        return dict(n=0, wr=0.0, roi=0.0)
+        return dict(n=0, wr=0.0, roi=0.0, refunds=0)
     sizes = sorted(r[5] for r in rows)
     k = (len(sizes) - 1) * pctile
     f = int(k)
@@ -123,11 +129,24 @@ def conviction_record(runq, wallet, days=90, pctile=0.80, now=None):
     best = {}
     for cond, asset, won, p, res_t, size in rows:
         if size >= thr and res_t >= cut:
-            if cond not in best or size > best[cond][2]:
-                best[cond] = (won, p, size)
+            if cond not in best or size > best[cond][3]:
+                best[cond] = (cond, asset, won, p, size)
     conv = list(best.values())
     if not conv:
-        return dict(n=0, wr=0.0, roi=0.0)
-    wins = sum(1 for won, _, _ in conv if won)
-    roi = sum(((1 - p) / p if won else -1.0) for won, p, _ in conv) / len(conv)
-    return dict(n=len(conv), wr=wins / len(conv), roi=roi)
+        return dict(n=0, wr=0.0, roi=0.0, refunds=0)
+    wins = losses = refunds = 0
+    roi_sum = 0.0
+    for cond, asset, won, p, size in conv:
+        wp = truthfn(cond, asset) if truthfn else None
+        if wp is None:
+            wp = 1.0 if won else 0.0
+        roi_sum += (wp - p) / p
+        if wp > 0.5:
+            wins += 1
+        elif wp < 0.5:
+            losses += 1
+        else:
+            refunds += 1
+    decided = wins + losses
+    return dict(n=decided, wr=(wins / decided) if decided else 0.0,
+                roi=roi_sum / len(conv), refunds=refunds)

live/validate_timing.py

@@ -37,6 +37,7 @@ import urllib.request
 from concurrent.futures import ThreadPoolExecutor
 
 import cache
+import payouts
 import smart_money as sm
 import trust
 
@@ -93,11 +94,25 @@ def _pm_profit(w):
         return None
 
 
-def _bet_pnl(b):
-    """Resolved (outcome) P&L of one cache bet: a $size stake at avg price p pays
-    size/p if won, else $0 — so P&L = size·(1−p)/p if won else −size."""
+def _wp(cond, asset, won):
+    """Chain-truth payout for a bet (1/0/0.5), falling back to the cache's
+    `won` mark when the chain can't say (unresolved, legacy NULL-asset rows on
+    decided markets, RPC gaps). The fallback keeps old behavior; the truth
+    path kills the two cache lies: 50/50 refunds counted as wins for BOTH
+    sides (28% of the follow set's resolved markets!) and stale both-sides-won
+    marks on operator-resolved markets."""
+    wp = payouts.truth(cond, asset)
+    return (1.0 if won else 0.0) if wp is None else wp
+
+
+def _bet_pnl(b, wp=None):
+    """Resolved P&L of one cache bet at payout wp: a $size stake at avg price p
+    returns size·(wp−p)/p — wp=1 win, 0 loss, 0.5 refund ($0.50/share, NOT
+    money-back: flat near coin-flip entries, ruinous for favorites)."""
     p = max(0.001, min(0.999, b["p"] or 0))
-    return b["size"] * ((1 - p) / p if b["won"] else -1)
+    if wp is None:
+        wp = _wp(b.get("cond"), b.get("asset"), b["won"])
+    return b["size"] * (wp - p) / p
 
 
 def display_stats(w):
@@ -121,35 +136,51 @@ def display_stats(w):
     now = time.time()
     bets = [b for b in cache.get_bets(w)
             if (b["size"] or 0) > 0 and (b["res_t"] or 0) <= now]
+    # chain-truth payouts for everything this wallet's stats touch (cached
+    # in the resolutions table — incremental after the first backfill)
+    trows = trust.trusted_wallet_rows(cache.query, w)
+    payouts.ensure({b["cond"] for b in bets} | {r[0] for r in trows})
     # ---- ALL-TIME stats over EVERY trusted bet (any size): the dashboard's
     # "of every bet placed" columns. Trusted rows only, deduped one-per-market,
-    # so scalper-poisoned cache marks can't inflate them. ----
-    trows = trust.trusted_wallet_rows(cache.query, w)
+    # truth-adjusted: refunds (wp=0.5) count as neither won nor lost. ----
     tbest = {}
     for cond, asset, won, p, res_t, size in trows:
-        if cond not in tbest or size > tbest[cond][2]:
-            tbest[cond] = (won, p, size)
-    all_won = sum(1 for won, _, _ in tbest.values() if won)
-    all_lost = len(tbest) - all_won
-    all_pnl = sum(size * ((1 - p) / p if won else -1.0)
-                  for won, p, size in tbest.values())
+        if cond not in tbest or size > tbest[cond][3]:
+            tbest[cond] = (cond, asset, won, p, size)
+    def tally(rows):
+        """(won, lost, refunds, pnl) over (cond, asset, won, p, size) rows."""
+        w_ = l_ = r_ = 0
+        pnl = 0.0
+        for cond, asset, won, p, size in rows:
+            wp = _wp(cond, asset, won)
+            pc = max(0.001, min(0.999, p or 0))
+            pnl += size * (wp - pc) / pc
+            if wp > 0.5:
+                w_ += 1
+            elif wp < 0.5:
+                l_ += 1
+            else:
+                r_ += 1
+        return w_, l_, r_, pnl
+    all_won, all_lost, all_ref, all_pnl = tally(tbest.values())
     thr = cache.conv_cutoff(b["size"] for b in bets)
     conv = [b for b in bets if b["size"] >= thr]
-    won = sum(1 for b in conv if b["won"])
     recent = sorted(bets, key=lambda b: b["res_t"] or 0, reverse=True)[:500]
     cut30 = time.time() - 30 * 86400
     conv30 = [b for b in conv if (b["res_t"] or 0) >= cut30]
-    won30 = sum(1 for b in conv30 if b["won"])
+    brow = lambda bs: [(b["cond"], b.get("asset"), b["won"], b["p"], b["size"]) for b in bs]
+    cw, cl, cr, cpnl = tally(brow(conv))
+    c3w, c3l, c3r, c3pnl = tally(brow(conv30))
     out = {
-        "conv_win": round(100 * won / len(conv), 1) if conv else None,
-        "conv_won": won, "conv_lost": len(conv) - won,
-        "conv_pnl": round(sum(_bet_pnl(b) for b in conv)),
-        "conv30_win": round(100 * won30 / len(conv30), 1) if conv30 else None,
-        "conv30_won": won30, "conv30_lost": len(conv30) - won30,
-        "conv30_pnl": round(sum(_bet_pnl(b) for b in conv30)),
-        "realized_pnl": round(sum(_bet_pnl(b) for b in recent)),
+        "conv_win": round(100 * cw / (cw + cl), 1) if (cw + cl) else None,
+        "conv_won": cw, "conv_lost": cl, "conv_ref": cr,
+        "conv_pnl": round(cpnl),
+        "conv30_win": round(100 * c3w / (c3w + c3l), 1) if (c3w + c3l) else None,
+        "conv30_won": c3w, "conv30_lost": c3l, "conv30_ref": c3r,
+        "conv30_pnl": round(c3pnl),
+        "realized_pnl": round(tally(brow(recent))[3]),
         "all_win": round(100 * all_won / (all_won + all_lost), 1) if (all_won + all_lost) else None,
-        "all_won": all_won, "all_lost": all_lost, "all_pnl": round(all_pnl),
+        "all_won": all_won, "all_lost": all_lost, "all_ref": all_ref, "all_pnl": round(all_pnl),
         "pm_pnl": _pm_profit(w),
         "avg_bet": round(sum(b["size"] for b in conv) / len(conv)) if conv else 0,
         "copy_pnl": 0, "held_pnl": 0, "held_won": 0, "held_lost": 0, "sold": 0,
@@ -206,13 +237,18 @@ def display_stats(w):
                 scalp += sh * pr - STAKE - openp[c]["fee"] - fee_out
                 sold += 1; del openp[c]
         for c, p in openp.items():                              # settle held bets at resolution
-            wv = resmap.get(p["a"])
+            wv = payouts.truth(c, p["a"])                       # chain first: refunds pay 0.5
+            if wv is None:
+                wv = resmap.get(p["a"])
             if wv is None:
                 wv = _clob_winner(c, p["a"])                    # clob fallback for out-of-pull markets
             if wv is None:
                 continue                                        # not resolved yet -> exclude
-            held += (p["sh"] if wv else 0) - STAKE - p["fee"]   # redeem itself is fee-free
-            hw += wv; hl += 1 - wv
+            held += p["sh"] * wv - STAKE - p["fee"]             # redeem itself is fee-free
+            if wv > 0.5:
+                hw += 1
+            elif wv < 0.5:
+                hl += 1
         out.update(copy_pnl=round(scalp + held), held_pnl=round(held),
                    held_won=hw, held_lost=hl, sold=sold)
     return out
@@ -281,8 +317,9 @@ def main():
         # holders are judged on their real resolved sample (the replay's own held leg
         # is mostly "unresolved" for them and only reported for display).
         tr = trust.conviction_record(cache.query, c["wallet"], days=TRUST_DAYS,
-                                     pctile=cache.CONV_PCTILE)
+                                     pctile=cache.CONV_PCTILE, truthfn=payouts.truth)
         c["trust_n"], c["trust_wr"], c["trust_roi"] = tr["n"], round(tr["wr"], 3), round(tr["roi"], 3)
+        c["trust_refunds"] = tr.get("refunds", 0)
         # SELECT a copyable sharp: active, copy-positive (fee-aware replay), and a
         # genuine hold-to-resolution edge — trailing trusted conviction record wins a
         # clear majority with positive flat-stake ROI on a real sample, so the edge