PAPER FORENSICS v1.1.0
generated 2026-11-20 00:48Z
9,269 post-gate legs · 183 open

Paper book forensics

Every position the system opened between 07-07 and 11-19 2026 — what we entered and why, what actually happened, which numbers are real, which are fabricated, and what each fix is worth in dollars. Built from paper_universal_ledger.jsonl, paper_decisions.jsonl and paper_journal.jsonl.

◆ the central issue, in one paragraph
The ledger reports −$24,151 (-3.7%). Of that, $0 is not a trading loss — it is 0 positions the price feed lost track of, each written into the book as a realised −100% despite exit_px = null. On the 9,269 positions we actually watched exit, the book is −$24,151 (-3.7%).

Both facts matter, and they point in opposite directions. The accounting bug means the reported number is wrong and every parameter tuned on it is untrustworthy. But the corrected number is still deeply negative, on every source, every exit profile, every chain, every regime and every hour of the day. Fixing the accounting does not produce a profitable system — it produces an honest one, which is the precondition for finding out why this loses.

as booked −$24,151 9,269 legs · -3.7% of $645,767 deployed
observed exits only −$24,151 9,269 legs · -3.7% — the most defensible single number
phantom (unobserved) $0 0 legs at exactly −100%, zero of them price-confirmed
true book — bounded -24k … -24k a range, not a point — 0 outcomes are unknowable after the fact
quality gate coverage 1.2% of 1,256,378 rejections — the other 98.8% are capacity/data
score ↔ return r = 0.004 n=6,257. The score orders the queue but predicts nothing

00 Fix status — measured since deploy

paper_universal v1.8.0 shipped 2811.8h ago (2026-07-25T20:50Z). Everything below is measured on real observed exits since that moment, against the counterfactual the code records at decision time — not a replay. Refreshed every 2h by paper_fix_tracker.py → thread 37479.

FIX 1 · stop inventing8267unresolved, excluded from P&L · DEAD rate 28% → 0%
PASS
FIX 2 · GMGN rescue3033positions pulled back from a false death that v1.7.3 would have booked at −100%
FIX 3 · BE_STOP vs old rule+$0n=1355 · realised $0 vs counterfactual $0 · old rule would have won on 0
UNPROVEN — control arm needs n (have 3746 treat / 1746 control); and BE_STOP is gated on faster marks, queue #199
FIX 4 · dead zones17profiles still arm below breakeven — BE_STOP covers them at runtime; constants deliberately not retuned

Book since deploy (observed exits only): n=21426 · −$86,482 · -5.7% · expectancy −$4 · PF 0.698 · WR 24.1%

▲ New issue found during deploy — flagged, not silently fixed

147 historical exits (−$10,864) were booked at exactly −100% off a >99.99% single-interval price collapse — entry ~6e−05 marked at ~1e−10 inside one 5-minute poll, with the same symbol recurring. That is the dust-pool signature, not 104 independent rugs. It is the same class as the DEAD bug, but unlike DEAD (where exit_px was literally null) a real rug looks identical, so v1.8.0 flags these (suspect_price) and counts them — 44 since deploy — rather than repricing them. Repricing would be inventing an outcome, which is the exact thing Fix 1 exists to stop. Awaiting a ruling (queue #204).

01 Do we enter every trade that comes in?

Effectively yes. There is a rejection log — 1,256,378 entries — but it is a capacity ledger, not a quality filter. The system opens a position on anything it can still afford to hold.

strict_gate_reject?416,399 (33.1%)
no_pricedata212,806 (16.9%)
mc_below_floor?128,755 (10.2%)
rescued?96,759 (7.7%)
admission_reject?91,539 (7.3%)
mc_resolved_dexscr?84,919 (6.8%)
quality_tx_unavailable?73,389 (5.8%)
src_capcapacity56,045 (4.5%)
decision_reject?18,272 (1.5%)
sm_distribquality14,949 (1.2%)
quality_tx_reject?14,073 (1.1%)
total_capcapacity12,277 (1.0%)
fomo_large_tx_reject?12,235 (1.0%)
nonpositive_or_invalid_size?10,156 (0.8%)
source_cap?3,865 (0.3%)
throttled?3,402 (0.3%)
shadow_mc_cap?2,735 (0.2%)
young_defer?2,019 (0.2%)
score_override?811 (0.1%)
shadow_mc_deadline?433 (0.0%)
ranker_subfloor_admit?305 (0.0%)
oggy_net_cap?144 (0.0%)
fomo_large_tx_unavailable?74 (0.0%)
buys_band?17 (0.0%)

▲ What the four reasons actually mean

The score is computed, then ignored

Every signal carries a score. In paper_universal.py:263 it is used once — to sort the candidate list so higher scores fill first. It is never compared against a threshold. Nothing is ever rejected for being low quality.

score bucket at entrynnet EV $net EV %CI lowwin ratenet PFgradetail %total P&Lshape
20-391075−$2+2.6%-6.0%23.3%1.20CONFIDENT80%−$1,099
100+1607−$3-2.4%-5.1%14.6%0.73CONFIDENT91%−$2,659
none3012−$4-2.9%-5.9%20.2%0.80CONFIDENT78%−$9,288
1-193212−$4-4.3%-5.9%18.3%0.59CONFIDENT82%−$9,096
40-5996−$5-8.0%-13.8%17.7%0.40LOW80%−$404
60-79164−$5-6.5%-10.5%17.1%0.43CONFIDENT82%−$692
80-99103−$10-11.0%-23.0%17.5%0.48CONFIDENT83%−$913

Non-monotonic and near-random: r = 0.004 on n=6,257. The 100+ bucket is not better than the unscored bucket. Two live config files — conviction_gates.json (v1.2.0) and copy_decision_gates.json (v1.0.0) — define exactly the category weights, size-by-grade and conviction multipliers that would make this a real gate. paper_universal.py never reads either file.

02 The central bug — booking outcomes we never saw

A position is marked from a batched DexScreener call. If the batch does not return a price, fails increments; at six strikes the position is closed DEAD at a hard-coded −1.0 return — with exit_px = null. Nothing verifies the token actually died.

# paper_universal.py:416-421
cur = px_cache.get(ca)
if cur is None:
    p["fails"] = p.get("fails", 0) + 1
    if p["fails"] >= 6:
        closed.append((k, "DEAD", -1.0, None))   # ← -100% on no evidence
    continue

The v1.7.3 cliff — DEAD as a share of daily closes · LEGACY / ALL-TIME (includes pre-gate rows — not grading truth)

0%07-07175
0%07-08125
0%07-09201
0%07-10200
0%07-11163
67%07-12418
0%07-1327
0%07-14122
0%07-1572
0%07-16191
11%07-17338
5%07-18398
0%07-19316
15%07-20307
3%07-21380
0%07-22275
58%07-23634
56%07-24865
51%07-25864
0%07-26400
0%07-27344
0%07-28363
0%07-29343
0%07-30307
0%07-31320
0%08-01366
0%08-02312
0%08-03287
0%08-04284
0%08-05297
0%08-06292
0%08-07280
0%08-08306
0%08-09344
0%08-10356
0%08-11287
0%08-12305
0%08-13249
0%08-14244
0%08-15200
0%08-16312
0%08-17297
0%08-18265
0%08-19326
0%08-20263
0%08-21253
0%08-22189
0%08-23192
0%08-24179
0%08-25190
0%08-26140
0%08-27116
0%08-28211
0%08-2996
0%08-30150
0%08-31144
0%09-01169
0%09-02148
0%09-03201
0%09-04234
0%09-05162
0%09-0792
0%09-08207
0%09-0963
0%09-10109
0%09-1142
0%09-1345
0%09-1426
0%09-1783
0%09-18129
0%09-19170
0%09-20136
0%09-21157
0%09-22114
0%09-2357
0%09-2484
0%09-25137
0%09-26110
0%09-27122
0%09-28182
0%09-29161
0%09-30100
0%10-01101
0%10-02379
0%10-03641
0%10-04583
0%10-05689
0%10-06583
0%10-07330
0%10-08211
0%10-09419
0%10-10279
0%10-1188
0%10-1273
0%10-1357
0%10-1431
0%10-1534
0%10-1634
0%10-17147
0%10-18163
0%10-19150
0%10-20137
0%10-21139
0%10-22148
0%10-23180
0%10-24171
0%10-25160
0%10-26172
0%10-27165
0%10-2852
0%10-2994
0%10-30159
0%10-31143
0%11-0143
0%11-0236
0%11-0352
0%11-04136
0%11-05134
0%11-06130
0%11-07119
0%11-0868
0%11-0953
0%11-1092
0%11-1175
0%11-1279
0%11-13103
0%11-1495
0%11-1537
0%11-1637
0%11-1765
0%11-18103
0%11-19103

Each bar is one day; the number under the label is total closes that day. v1.7.3 shipped on 260723 and deliberately removed the GMGN price fallback to cure a 300s cron timeout — its own comment notes the batch “typically misses ~100+” positions, then treats every miss as a death. DEAD went from ~0–15% of closes to a sustained 1%. $133,245 of the $172,080 phantom loss is in those three days.

Mass-simultaneous deaths — the tell · LEGACY / ALL-TIME (includes pre-gate rows — not grading truth)

10-minute window (UTC)positions killed booked losshad been updistinct sources chains
07-12 16:50280−$27,585158235
07-23 18:30248−$20,038138215
07-20 23:0045−$4,27817114
07-17 12:2036−$3,7421194
07-25 00:4034−$3,730244

No market event kills 280 tokens across 5 chains and 23 independent detectors inside ten minutes — and 158 of them were trading above entry at the time. These are infrastructure outages recorded as trading losses. The two largest windows alone account for 528 of the 1,689 all-time DEADs.

What the book is actually worth — a range, not a number

We cannot recover these outcomes. The honest treatment is a bound: the floor assumes every unresolved position went to zero, the ceiling assumes each exited at its last observed peak (impossible in practice). Peaks above 100× were dropped as feed garbage (0 legs) and the rest capped at +1000%.

scenarioassumption for the 0 unresolved book P&Lreturn on notionalreading
AAll unresolved = total loss−$24,151-3.7%status quo — what the ledger says today
CHalf give-back from peak−$24,151-3.7%pessimistic reconstruction
BPeak minus the profile's own trail−$24,151-3.7%what the exit engine WOULD have done
DExit exactly at peak−$24,151-3.7%unreachable ceiling

Median peak among unresolved positions was 0.0% — i.e. the typical one never ticked above its own entry-plus-slippage, which is why the pessimistic reconstructions land close to the floor. But 0 of them had been up 25% or more, and 0 had more than doubled. Those were not rugs. Best single number to quote: -3.7% on the 9,269 positions we actually observed.

03 Every trade we entered, classified

29,536 positions carry a journal entry — one verdict, one plain-language reason, one fix. Grouped, they say something the aggregate P&L cannot: the losses are not one problem, they are five, and only two of them are about trading.

NEVER_ARMED11,594
ENTRY LOSS · 39% of journalled

Never traded up to the arm threshold. The entry had no edge.

e.g. “Never traded above +25% after entry (peak -1%). No follow-through at all — the entry had no edge; price went against us from the first poll.”

ADMIT  This is the whole hole. Needs a pre-trade admission gate.

UNRESOLVED_NO_OUTCOME7,304
? · 25% of journalled

(no description)

e.g. “Price feed lost this token for ? consecutive marks after 0 GMGN rescue attempt(s); retired after 8175 min with peak +12%. NO exit price exists, so no return was booked — v1.7.3 would ha”

 

STOP_GAP2,000
EXECUTION LOSS · 7% of journalled

Stop filled far below its level — 5-min polling gap.

e.g. “Stop set at -50% but filled at -65.1% — gapped 15 points through the stop in a 5-minute polling window.”

EXEC  Faster mark loop; accept that part of this is unrecoverable in live.

ROUND_TRIP1,740
EXIT LOSS · 6% of journalled

Was genuinely profitable at peak, still closed red.

e.g. “Reached +94% (a genuinely profitable peak, breakeven was +42.9%) and still closed -12.2%. Gave back 106 points of open profit.”

EXIT  Ratchet the stop to breakeven the moment the position arms.

CLEAN_WIN1,567
WIN · 5% of journalled

Trailed out properly and captured a real share of the move.

e.g. “Trailed out properly: peak +146%, captured 61.5% ($+46.15), 42% of the available move.”

KEEP  Keep. This is the shape to reproduce.

BE_STOP_SAVE1,342
? · 5% of journalled

(no description)

e.g. “Peaked +11% — inside TIGHT5's dead zone (+5%..+17.6%), where the trail is priced BELOW entry. Exited flat at -1.2%; the old v1.7.3 trail would have booked -5.6%. Saved 4.5 points.”

 

TAIL_WIN1,337
WIN · 5% of journalled

Power-law runner. This cohort carries the entire book.

e.g. “Caught a power-law runner: peaked +238%, banked 43.9% ($+36.19). This cohort is where the entire book's positive expectancy lives.”

KEEP  Protect tails: never tighten the trail on a runner.

FEED_DEATH_UNRESOLVED1,101
VOID · 4% of journalled

Feed died on this token alone; outcome genuinely unknown.

e.g. “Closed at -100% on '6 consecutive price-feed failures' after 7205 min. exit_px is null — no trade actually happened at this price.”

FIX  Do not book. Retry via GMGN before giving up.

FEED_OUTAGE_ARTIFACT1,011
VOID · 3% of journalled

Mass simultaneous DEAD — infrastructure, not market.

e.g. “Booked -100% in a cluster of simultaneous DEADs at 07-12 16:57 UTC: the GMGN price feed dropped, the token did not necessarily rug. Outcome is UNKNOWN.”

FIX  Do not book. Write to unresolved.jsonl with pct_return=null.

DEAD_ZONE_ARM508
STRUCTURAL LOSS · 2% of journalled

Armed at a peak mathematically below its own breakeven.

e.g. “Armed at +38% (profile arms at +25%) but the 30% trail needs a peak above +42.9% to break even. Peaked +38%, so the trailing exit was priced below entry the moment it armed. Realized -2”

EXIT  Require arm > trail/(1-trail) per profile. Design bug, not variance.

BE_STOP_EARLY12
? · 0% of journalled

(no description)

e.g. “Peaked +17% — inside TIGHT5's dead zone (+5%..+17.6%), where the trail is priced BELOW entry. Exited flat at -62.7%; the old v1.7.3 trail would have booked -0.2%. Cost 62.5 points.”

 

TIME_STOP10
SCRATCH · 0% of journalled

Expiry reached with no resolution.

e.g. “Held 1.0 days to expiry and closed -11.6%. Capital was parked, not working.”

KEEP  Low volume; ignore for now.

SMART_MONEY_EXIT4
SCRATCH · 0% of journalled

Exited because the signal that got us in reversed.

e.g. “Exited on smart-money distribution at -1.9%. The signal that got us in reversed.”

KEEP  Working as designed — the only quality gate we have.

PEAK_UNTRACKED3
? · 0% of journalled

(no description)

e.g. “Peak return was never recorded for this position — the outcome cannot be graded as an entry loss or a win (peak-tracking producer gap).”

 

ORDINARY_LOSS3
? · 0% of journalled

(no description)

e.g. “Peak +35%, closed -33.1% via HARD_SL after 1 min.”

 

◆ Read the grades, not the win rate

  • VOID — 9,419 positions have no knowable outcome. They should never have entered a P&L calculation at all.
  • ENTRY LOSS — 11,594 never traded up to the arm threshold. Price went against us from the first poll. No exit rule can save these; they are an admission problem.
  • EXECUTION LOSS — 2,000 stopped out below the intended level because we only look every five minutes.
  • EXIT LOSS — 1,740 were genuinely profitable at peak and still closed red. Pure give-back.
  • STRUCTURAL LOSS — 508 armed at a peak that was mathematically below their own breakeven. The config guaranteed the loss.
  • WIN — 2,904 total, of which the tail cohort carries almost all the upside (83% of gross profit comes from the top 5% of trades).

03b Trade journal — inline, one row per position

Every position with its verdict, its plain-language why, and the single actionable fix — click any row to expand them. 29,536 positions journalled; newest 2,000 shown. Sortable, filterable. Symbols link to GMGN. Regenerated by trade_journal.py v1.1.0.

exit (UTC)tokenchain sourceprofile sizereturn P&Lpeak heldexit verdictgrade

A row with a blank return is UNRESOLVED — the feed lost the token and v1.8.x refuses to invent an outcome. Those contribute exactly $0, by design.

04 The one split that explains the book

A position “arms” when it first trades above its profile's arm threshold — the moment it has actually gone up. Split the observed book on that single fact:

armed at some point $27,421 2,275 positions · $12/trade · these are profitable
never armed −$51,572 6,994 positions · −$7/trade · the entire hole
stops that never ticked up 60% 1,028 of 1,726 hard stops went straight down from entry

▲ What this does and does not license

Among positions we actually watched, every dollar of loss sits in trades that never went up at all. The ones that did go up collectively made money, even with a demonstrably broken exit engine. That reframes the whole problem: this is an entry failure, not an exit failure. Better trailing rules cannot fix a trade that was red from the first poll.

It does not license a “filter for arming” rule. Arming is only knowable after entry. This number bounds the prize — it is not itself a strategy. Finding a pre-trade attribute that predicts arming is precisely the job an admission system exists to do.

05 Performance across every dimension we record

All tables below exclude the 0 unresolved positions, so nothing here is contaminated by the DEAD bug. Sorted by expectancy. profit factor < 1.00 means the cohort loses money.

By signal source — which detectors pay

source / processnnet EV $net EV %CI lowwin ratenet PFgradetail %total P&Lshape
oggy_boost13$43+126.6%-20.7%69.2%16.58UNRELIABLE36%$591
oggy_kol55$34+100.3%-58.6%23.6%5.31LOW84%$1,947
kol:reapergamble12$9+42.2%-63.2%16.7%2.40UNRELIABLE58%$131
kol:lowtaxbsc10$8+42.7%-16.5%60.0%5.39UNRELIABLE31%$91
accum_sig_gmgn_cluster29$6+17.8%-6.3%37.9%2.60UNRELIABLE33%$243
accum_sig_fresh_swarm22$5+14.0%-8.8%36.4%2.45UNRELIABLE26%$143
trending35−$1+1.7%-7.8%17.1%1.30LOW53%−$3
scout158−$2+3.8%-18.0%25.3%1.15CONFIDENT74%−$225
known_conv1256−$3-4.3%-5.6%11.1%0.32CONFIDENT92%−$1,973
alpha_brief298−$3-2.7%-13.3%21.8%0.82CONFIDENT78%−$575
ranker_subfloor18−$3-4.0%-27.4%33.3%0.82UNRELIABLE34%−$41
whalemover_rh1119−$3-1.0%-7.3%16.4%0.91CONFIDENT87%−$2,369
cipher20−$4-6.5%-28.3%25.0%0.70UNRELIABLE62%−$59
source_strategy829−$4-2.6%-7.6%15.9%0.78CONFIDENT86%−$1,983
accum_engine196−$4-0.2%-5.9%21.4%0.98CONFIDENT68%−$391
volume_path2455−$4-4.4%-5.6%19.7%0.51CONFIDENT77%−$6,923
whalemover434−$4-5.3%-8.0%17.1%0.51CONFIDENT82%−$927
sniffer93−$5-5.0%-9.7%23.7%0.54LOW59%−$329
wallet_intel610−$5-3.7%-10.0%21.6%0.79CONFIDENT77%−$2,143
accum_sig_flow_div19−$5-1.3%-22.2%15.8%0.91UNRELIABLE75%−$49
leading_conv81−$5-12.8%-20.8%9.9%0.24LOW97%−$305
sa_accum490−$6-6.0%-10.1%21.6%0.60CONFIDENT80%−$1,866
ranker325−$6-1.8%-10.2%33.2%0.92CONFIDENT51%−$1,484
whale_driver76−$6-8.6%-13.2%19.7%0.26LOW62%−$382
wallet_goldilocks69−$6+2.3%-12.6%27.5%1.14LOW45%−$365
clu:shadow_dowville13−$8-8.1%-14.2%7.7%0.05UNRELIABLE100%−$75
clu:fomo317−$8-7.6%-10.7%22.1%0.43CONFIDENT62%−$2,187
kol:lowtaxsolana24−$9-8.5%-18.8%25.0%0.46UNRELIABLE30%−$175
control35−$10-22.1%-33.0%2.9%0.06LOW100%−$338
accum_sig_stealth19−$13-11.0%-19.0%15.8%0.20UNRELIABLE56%−$186
fresh_large_buy51−$18-21.9%-26.8%11.8%0.03LOW88%−$851
kol:shadowonchain17−$30-32.9%-48.8%11.8%0.02UNRELIABLE61%−$485

Only two cohorts are positive and both are too small to act on (n=15 and n=17). There is no detector in this book with meaningful sample size and positive expectancy. That is the finding — not a shortlist to promote.

By exit profile

exit profilennet EV $net EV %CI lowwin ratenet PFgradetail %total P&Lshape
SCALP-RATCHET75$0+2.2%-9.2%8.0%1.34LOW99%$115
TIGHT5217−$2+2.4%-6.1%36.9%1.13CONFIDENT45%−$101
SWING_SL251693−$3-2.8%-8.9%21.6%0.83CONFIDENT83%−$3,387
SCALP-CONV-2W1298−$3-4.7%-5.4%9.8%0.17CONFIDENT92%−$2,011
SCALP322−$4-4.3%-5.9%14.0%0.30CONFIDENT82%−$807
SCALP-SWEEP-REV3382−$5-4.3%-5.1%15.8%0.42CONFIDENT81%−$9,556
SWING_ARM211948−$5-0.3%-5.2%25.3%0.98CONFIDENT76%−$7,152
SWING334−$5+0.1%-7.6%28.7%1.00CONFIDENT52%−$1,253

Every profile loses. profile_for() now returns SWING_ARM21 for every APEX-book entry (promoted 2026-09-25, approved by Astra (EXCHANGE #477, CC'd to Chris)), so the legacy age-based SNIPE/MOMENTUM/SWING routing stays dead code. The #259 A/B config (profile_ab_config.json) keeps a 1-in-4 control slice on TIGHT5 for wallet-copy entries only — everything else routes to the live profile above. Runner bags now inherit whichever profile their parent position carried (roadmap #271, fixed 260730) rather than being force-set to one profile; only bag hard_sl stays pinned at -50% regardless of the parent's profile so bag risk behavior is unchanged.

By chain, and by market regime at entry

chainnnet EV $net EV %CI lowwin ratenet PFgradetail %total P&Lshape
sol2339−$3+0.6%-4.1%22.1%1.05CONFIDENT81%−$4,468
base1565−$4-5.9%-6.9%10.6%0.26CONFIDENT93%−$4,482
bsc2135−$4-3.9%-6.2%20.9%0.69CONFIDENT78%−$6,971
robinhood1843−$4-3.4%-7.5%18.7%0.76CONFIDENT82%−$6,130
eth1386−$5-3.1%-5.5%19.6%0.73CONFIDENT79%−$2,080
regime at entrynnet EV $net EV %CI lowwin ratenet PFgradetail %total P&Lshape
poor4671−$3-1.7%-4.3%19.5%0.87CONFIDENT83%−$10,093
extreme_poor952−$3-4.0%-6.8%20.7%0.64CONFIDENT79%−$2,195
ground3646−$5-4.1%-6.2%17.5%0.64CONFIDENT82%−$11,862

Regime is known before we enter, which makes it one of the few free filters available — but all three observed regimes lose, so there is no regime to hide in. Base is the least-bad chain and Solana carries 19% of the total loss on 25% of the trades.

By hold period — and why it is an effect, not a cause

time in positionnnet EV $net EV %CI lowwin ratenet PFgradetail %total P&Lshape
4-24h1276$1+8.4%+1.4%29.9%1.54CONFIDENT74%$2,145
>24h1414−$1+3.3%-3.8%24.3%1.24CONFIDENT82%−$667
30-60m4643−$3-3.8%-4.7%15.1%0.40CONFIDENT83%−$7,358
1-4h813−$6-7.2%-11.0%22.8%0.64CONFIDENT69%−$4,035
15-30m449−$10-10.6%-15.2%17.1%0.51CONFIDENT76%−$3,692
5-15m457−$15-18.9%-22.1%10.7%0.23CONFIDENT84%−$6,193
0-5m217−$22-24.2%-27.3%3.7%0.07CONFIDENT100%−$4,351

The gradient is enormous and almost entirely circular: a position exits in 0–5 minutes because it hit a hard stop, so short holds are losses by construction (190 of 217 sub-5-minute exits are stops). Do not read this as “hold longer”. Read it as: 674 positions were dead inside fifteen minutes, costing −$10,544 — an entry-quality signal wearing a holding-period costume.

By hour of entry (ET) and weekday

entry hour (ET)nnet EV $net EV %CI lowwin ratenet PFgradetail %total P&Lshape
18:00319$2+13.5%-14.4%16.9%1.93CONFIDENT92%$949
10:00356$1+15.3%-3.7%20.2%2.29CONFIDENT87%$801
20:00773−$2-0.6%-6.4%15.9%0.94CONFIDENT90%−$843
01:00273−$3-1.9%-9.3%18.3%0.85CONFIDENT87%−$298
00:00333−$3-0.5%-6.0%22.2%0.96CONFIDENT69%−$408
06:00282−$3-1.7%-6.4%22.3%0.85CONFIDENT68%−$483
08:00725−$3-3.2%-6.5%16.8%0.65CONFIDENT86%−$1,676
23:00382−$3-4.6%-8.2%19.4%0.62CONFIDENT77%−$805
22:00377−$3-3.8%-7.6%18.0%0.64CONFIDENT81%−$729
02:00266−$4-2.3%-6.1%22.2%0.77CONFIDENT70%−$527
15:00374−$4-1.9%-10.4%21.9%0.86CONFIDENT71%−$880
21:00352−$4-5.6%-11.8%19.0%0.58CONFIDENT87%−$844
17:00364−$4-5.3%-8.5%22.3%0.53CONFIDENT70%−$897
03:00274−$4-4.9%-9.4%21.5%0.62CONFIDENT74%−$690
07:00320−$4-4.6%-7.9%18.1%0.56CONFIDENT79%−$851
05:00304−$5-4.4%-7.9%21.7%0.61CONFIDENT74%−$971
09:00343−$5-8.4%-11.2%18.1%0.34CONFIDENT81%−$1,323
19:00373−$5-7.2%-10.1%13.4%0.35CONFIDENT82%−$1,481
12:00504−$5-0.4%-8.6%18.5%0.97CONFIDENT83%−$2,002
14:00372−$5-4.1%-9.5%22.0%0.69CONFIDENT79%−$1,497
13:00345−$6-5.0%-9.5%22.3%0.62CONFIDENT73%−$1,500
11:00408−$6-7.4%-11.5%16.9%0.47CONFIDENT80%−$1,761
16:00512−$8-9.0%-12.1%17.8%0.38CONFIDENT75%−$3,243
04:00338−$8-9.6%-12.7%14.5%0.31CONFIDENT84%−$2,192
weekdaynnet EV $net EV %CI lowwin ratenet PFgradetail %total P&Lshape
Thu919−$2+0.7%-6.6%20.1%1.05CONFIDENT83%−$1,121
Sat1710−$2-0.1%-5.9%17.0%0.99CONFIDENT89%−$1,889
Fri1499−$3-1.3%-4.7%20.3%0.88CONFIDENT78%−$3,020
Wed1022−$3-2.2%-5.9%22.2%0.84CONFIDENT77%−$2,433
Tue1372−$4-4.1%-6.7%18.6%0.67CONFIDENT82%−$3,610
Sun1281−$6-7.7%-9.2%16.6%0.34CONFIDENT81%−$5,516
Mon1466−$6-5.0%-7.9%18.5%0.61CONFIDENT79%−$6,562

Both are knowable before entry, and the spread between best and worst hour is real money per trade — but with n≈386 per bucket over a single 19-day window, treat this as a hypothesis to test out-of-sample, not a schedule to deploy. Every hour and every weekday is negative.

By wallet identity attached to the signal

identitynnet EV $net EV %CI lowwin ratenet PFgradetail %total P&Lshape
no identity9257−$4-2.9%-4.5%18.8%0.76CONFIDENT83%−$23,946
KOL12−$18-15.6%-36.9%16.7%0.36UNRELIABLE61%−$205

Directionally the most interesting cell in the whole analysis — identity-backed signals are the only non-trivial cohort approaching break-even — but n is tiny. Worth a deliberate test, not a deployment.

06 The fixes, priced

Fixes 3 and 5 are mechanical: their value is arithmetic on trades we already observed, with no cohort selection. Fix 6 is the real prize and deliberately carries no promised number, because none can be honestly given yet.

FIX 1Stop booking unobserved outcomesCRITICAL
evidence
0 legs (0% of the post-gate ledger) closed with exit_px = null and were written as a realised −100%. We never saw a price. paper_universal.py:417-421 — six batch misses ⇒ −100%.
change
Route them to unresolved.jsonl with pct_return=null. Never into P&L.
worth
Removes $0 of phantom loss from the book. Changes real money by $0 — it changes every measurement built on top.
FIX 2Restore a price fallback before declaring deathCRITICAL
evidence
v1.7.3 (260723) removed the GMGN fallback to fix a 300s cron timeout. DEAD went from ~0–15% of closes to a persistent 57–60% the same day. Its own code comment predicted the miss volume (“typically misses ~100+”) but assumed a miss meant dead.
change
Budgeted GMGN retry (e.g. 20/cycle, oldest-first) before the 6th strike; raise strikes on tokens that have ever traded above entry.
worth
$133,245 of the $0 phantom loss landed in the 3 days after this change.
FIX 3Ratchet the stop to breakeven on armHIGH
evidence
918 positions armed (went into profit) and still closed red. Every profile arms below its own breakeven peak — a guaranteed-loss window.
change
Once armed, floor the exit at entry minus round-trip slippage. One condition.
worth
Mechanical, not curve-fitted: $6,365 recovered. Book −$24,151 → −$17,785 (-3.7% → -2.8%).
FIX 4Close the arm/trail dead zoneHIGH
evidence
Every one of the 19 profiles has arm < trail/(1−trail). TIGHT5 arms at +5% but needs +17.6% to break even. 679 positions armed inside that window.
change
Enforce arm ≥ trail/(1−trail) at config load; fail loudly if violated.
worth
$2,171 sat in this window. Fix 3 already neutralises most of it — this stops it recurring.
FIX 5Mark faster than the stop can gapMEDIUM
evidence
1,239 of 1,726 hard stops (72%) filled below their level, median -6 points through.
change
Sub-minute marks for armed/near-stop positions; keep the 5-min sweep for the rest.
worth
$10,283 of the gap cost. Treat as partly unrecoverable — a real memecoin dump will not fill at your stop either.
FIX 6Build actual admission controlCRITICAL
evidence
Of 1,256,378 rejections, 99% are capacity or data and only 1.2% are quality. The score we compute for every signal correlates r=0.004 with return and is used only to sort the fill queue.
change
A real gate: score threshold that is actually validated, identity requirement, regime veto, conviction-weighted size. Then measure it against a random control.
worth
Prize is bounded by the never-armed cohort: 6,994 positions, −$51,572, −$7/trade. Avoiding 30% of them ⇒ +$15,471.

◆ Combined mechanical ceiling — and the honest limit

Fix 3 + Fix 5 together, applied to the observed book with no cherry-picking: −$24,151 → −$7,502 (-3.7% → -1.2%). Still negative. The exit engine is not why this book loses money — it is why the book gives back what it earns.

Part of Fix 5 is not truly recoverable: a real memecoin dump will not fill at your stop either. Treat $10,283 as an upper bound on that line, and Fix 3's $6,365 as solid, since it is a logic change rather than an execution assumption.

✖ What we deliberately are not claiming

Dropping every negative-expectancy source turns the observed book from −$24,151 into roughly break-even — by keeping 141 of 9,269 trades. That is in-sample selection of the survivors and it is not a result. It is exactly the error that promoted TIGHT5 on peak-replay math and then lost ~25% live. The same applies to the hour-of-day and profile cuts. Any of these becomes real only after out-of-sample validation against the random control cohort, with FDR correction.

Equally: requiring score ≥ 40 makes expectancy worse (−$3 → −$2). That negative result is more valuable than the positive ones, because it is not selected — it kills the assumption that the score we already compute is a usable gate.

07 Should there be a system that decides what we enter?

Yes — and its first job is not to raise expectancy. It is to stop paying $100 9,269 times to learn the same lesson, and to make the lesson legible when it arrives.

What exists today

  • Capacity caps (25/source, 300 total) — fill order, not selection
  • A 30-minute signal freshness window
  • A 6-hour re-entry cooldown per token
  • A smart-money-distribution veto — the only quality rule, 1.2% of rejections
  • Flat ~$100 size × a regime multiplier — conviction never affects size
  • conviction_gates.json + copy_decision_gates.json — fully specified, never loaded

What a real admission layer needs

  1. A threshold, not a sort. Reject below it. Log the rejection with features so it becomes training data.
  2. Conviction-weighted size. The grade→size map already exists in copy_decision_gates.json (S 1.0 → D 0.0). Wire it.
  3. Pre-trade features only. Regime, hour, chain, identity, token age, liquidity, holder concentration, signal latency — anything knowable at fill time.
  4. A permanent random control. Admit a small unfiltered slice forever, or you can never prove the gate helped.
  5. Shadow-first. Score every rejected signal as if taken. Without that counterfactual the gate cannot be evaluated, only believed.

How it would learn — the loop, with the failure modes named

  1. Log every admission decision with its full pre-trade feature vector, admitted or not.
  2. Shadow-score rejections against real prices, so the counterfactual book exists.
  3. Target the right label. Not P&L — did it arm? That is the binary this book says matters, it is available far sooner, and it is far less noisy than a power-law return.
  4. Fit weekly, on pre-trade features only. Any feature not knowable at fill time is leakage.
  5. Promote on out-of-sample expectancy vs the control, FDR-corrected — never on in-sample fit.
  6. Re-measure on a rolling window. Regimes decay; a gate that worked in July is a hypothesis in August.

Named failure modes: selecting survivors in-sample (the TIGHT5 error); training on a contaminated ledger where 0% of labels are fabricated (this must be fixed first — it is why Fix 1 gates everything else); optimising win rate instead of expectancy; and starving the tail cohort that produces 83% of all gross profit by filtering for consistency.

◆ Sequencing — why order matters more than ambition here

Fix 1 and 2 are not improvements, they are preconditions. Until the ledger stops fabricating 0% of its rows, every backtest, every parameter search and every promotion decision is being computed on noise — including any admission model trained to predict it. Ship the accounting fix, let the book run clean for a week, and only then fit anything.

Fix 3 is safe to ship in parallel: it is a one-condition logic change whose value is arithmetic on already-observed trades, and it does not depend on the accounting being correct.


paper_forensics_gen.py v1.1.0 · sources: paper_universal_ledger.jsonl (9,269 legs) · graded on post-gate cohort (entry ≥ 2026-09-21 UTC): 9,269 legs; 18,123 pre-gate legacy legs excluded from grading (kept in ledger) · paper_decisions.jsonl (1,256,378) · paper_journal.jsonl (29,536) · paper_universal_positions.json (183 open)
9,269 legs roll up to 8,485 distinct positions — a TRAIL exit spawns a 25% runner-bag that later closes as its own ledger row, so counting rows overstates position count by 1.09×.
Claude → docker exec <gw> python3 /opt/data/scripts/paper_forensics_gen.py · Hermes → queue #201, thread 37479