Token Selection — Feeds, Filters & Flow

🚧 Bottleneck map — where every candidate went (24h)

13,568 candidates · each colour is a stage that stopped them · green = entered paper · per source: slots.html · GMGN detail: gmgn.html · per filter: filter_dashboard.html

Other refusals ◀ BIGGEST BOTTLENECK

smaller reasons — see LIVE FUNNEL tab

5,459

40.2%

Admission (volume / decision engine)

vol_skip is LIVE on ~36% of flow while its config says shadow_only (#1376 re-test)

3,671

27.1%

Detector gates

per-source entry gates — trending's $1.5k single-tx gate rejects almost all trending (#1350)

2,859

21.1%

Quality-wallet gate: NO verdict (GMGN unavailable)

GMGN asked 4.4x its capacity, ~87% shed -> #1373 scout fix live 12:37 UTC 260923 + #1377 wallet-polling redesign

670

4.9%

Capacity / hard caps

slots full or source over its hard cap (scout 5/24h, #1373)

609

4.5%

Quality-wallet gate: real verdict

no WhaleMover / S-A / KOL / smart / fresh wallet bought >= $1k ($500 RH) in 1h

131

1.0%

Young-token wait (A/B)

tokens < 60 min parked for one re-check at 60-90 min (#1357 3-arm test)

0.5%

ENTERED paper

100

0.7%

§ GMGN capacity — last hour

78% of GMGN requests shed

2,503 asked · 402 served (live+cache) · 1,964 shed · 5 rate-limited (429)

§ Paper slots — open positions

155 / 300 slots used

largest consumer: whalemover_rh 25 open (16%) · share by source on slots.html

🚰 The intake funnel

n=13,550 decisions · trailing 24h · measured read-only from paper_decisions.jsonl

§ Refusal ledger — 24h

What the funnel actually does all day

admission_reject

3,173

23%

mc_below_floor

2,392

18%

strict_gate_reject

1,913

14%

mc_resolved_dexscr

1,783

13%

nonpositive_or_invalid_size

1,284

no_price

739

quality_tx_unavailable

670

decision_reject

498

src_cap

430

fomo_large_tx_reject

207

source_cap

179

quality_tx_reject

131

rescued

young_defer

§ Heatmap — source × refusal reason

Each source dies in exactly one characteristic way

SOURCE	admission<br>rej	mc<br>below<br>floor	strict<br>gate<br>rej	mc<br>resolved<br>dexscr	nonpositive<br>or<br>invalid<br>size	no<br>price	quality<br>tx<br>unavailable	decision<br>rej	src<br>cap	fomo<br>large<br>tx<br>rej
volume_path	2,906	1,064	·	909	1,018	92	37	·	114	·
whalemover_rh	4	725	·	573	·	317	599	·	242	·
trending	2	·	1,913	·	·	·	9	·	·	·
known_conv	31	473	·	235	234	259	·	·	·	·
wc:Reaper-Waldo	·	·	·	·	·	31	·	485	·	·
fomo	·	·	·	·	·	·	·	·	·	207
source_strategy	15	57	·	4	·	10	4	·	74	·
leading_conv	·	·	·	·	·	·	·	·	·	·
oggy_kol	73	23	·	·	·	·	·	·	·	·
wallet_intel	62	·	·	·	·	·	13	·	·	·
control	5	12	·	·	·	6	·	·	·	·
alpha_brief	11	·	·	34	32	·	·	·	·	·

hard refusal no-verdict / missing input capacity cap rescue real verdict opacity ∝ volume

§ The gate that refuses without a verdict

quality_tx: 83.6% of its refusals contain no verdict at all

Traced 260923: the gate shells out to gmgn-cli token traders, which exits 3 (circuit blocking, state-gated) when our own pacing gate sheds it. signal_filters._gmgn() maps every non-zero exit to None, so contention is filed as absence of evidence.
It is not a broken client: 8 probes at 4s spacing returned 4 ok / 4 shed, with the GMGN gate closed and zero families in backoff. The lane is losing a contended budget.
Root cause (#1373, EXCHANGE #240): scout's own MILESTONE brag rows (92% of scout rows, −$1,791 over 128 trades) were read as entries and each token was re-checked ~18× per run, flooding GMGN. Fixed 260923 12:37 UTC: milestone rows never entries, one row per CA per run, scout hard cap 5/24h. Gate patience was deliberately NOT added. Wallet polling (#1377) is the remaining GMGN load.

💰 What enters, what is refused, what it earned

23,432 ledger legs · 630,718 decisions · 155 open positions

§ Window — every number below re-reads on click

✅ Admitted

🚫 Refused

🎯 Admit rate

admitted ÷ (admitted + refused)

💵 Realised PnL

on trades CLOSED in the window

📐 $ per admit

the number that decides what is fruitful

§ Per source — refused (left) vs admitted (right), then what the admissions earned

Source	🚫 Refused ◀ \| ▶ ✅ Admitted	Refused	Admit	Rate	💵 Realised PnL	PnL $	$/admit	Win%

refused / losing admitted profitable bars log-scaled within the window

§ Refusal mix for this window

🚫 Which gate is doing the refusing

§ How to read this

⚠ Two numbers do not mean what they look like

This PnL sum is not the canonical book. Summing pnl_usd over every ledger leg double-counts BAG legs (#1344, ~33% N inflation); the canonical path collapses them and filters resolved outcomes. Rank sources with these bars — never quote the total.
“Refused” is not one thing. It mixes real verdicts with capacity caps (src_cap) and no-verdict refusals (quality_tx_unavailable). A big refusal bar can just mean a position ceiling.
Admits are legs, refusals are decisions. One token makes many decision rows per scan cycle, so admit-rate is a direction, not a probability.
$/admit is the honest ranking column — realised dollars divided by entries that actually opened, so a source cannot look good by being rare.
Decision history in this file starts ·; PnL reaches further back than refusals.

⚗ Filter algorithms — live thresholds and what they cost

read live from admission_gate_config.json · volume_gate_config.json · quality_tx_gate.json · admission_gate_log.jsonl

vol_skipLIVE · rejectingcrowded-volume filter

WHAT IT CHECKSRejects tokens whose buying is already crowded. Two ratios are banded LOW/MID/HIGH: net-buy volume 1h ÷ mcap, and volume turnover ÷ mcap. LOW=+1, HIGH=−1, MID=0. Grade A(≥2) and B(1) pass; C(0) and D(<0) are rejected — so MID+MID is rejected.

LIVE THRESHOLDnet_buy_vol_1h/mcap: LOW ≤ 0.0075 · HIGH > 0.0435 | turnover/mcap: LOW ≤ 2.636 · HIGH > 10.24

KNOB TO CHANGEvolume_gate_config.json → bands.*.low_max / high_min · admission_gate_config.json → gate_on_volume_skip

EXCUSED WHENExcused if an S-tier whale is in the signal (True) or a fresh wallet buys ≥ $1500 AND ≥ 0.5% of mcap.

IMPACT3,173 rejected in 24h (39% of gate traffic) · 21,701 in 7d — the single biggest filter.

38.6% of 24h gate traffic

EVIDENCEIC −0.186 (n=1,612); ratio ceiling lifted expectancy −$12.90 → −$1.91 (n=103). Measured 260726, IN-SAMPLE, before the latency/receipt fixes.

⚠ volume_gate_config.json still says shadow_only:true — STALE. It IS enforced (see the 24h rejects).

weak_smart_volumeSHADOW · logs onlysurvival-thesis filter

WHAT IT CHECKSWould reject when BOTH net smart buy $ in the last hour AND the buy/sell ratio are weak. Missing volume data is UNKNOWN and never a reject.

LIVE THRESHOLDnet_usd_1h < $100 AND buy/sell < 2.0 · mode = shadow_log

KNOB TO CHANGEadmission_gate_config.json → weak_smart_volume.min_net_usd_1h / min_buy_sell_ratio / mode

EXCUSED WHENSame exemptions as vol_skip (S-tier whale / fresh large buy).

IMPACT415 would-rejects logged in 24h · 3,325 in 7d (not blocking).

5.0% of 24h gate traffic

EVIDENCE+3.6pt (n=483) — derived from a POISONED log; being re-proven on clean data before any re-enforce.

quality_txLIVE · shed-blockedquality-wallet ≥$1k gate (Chris, EXCHANGE #200/#204)

WHAT IT CHECKSRequires ≥1 WhaleMover/SMART/KOL/FRESH/S-A or GMGN smart_degen/kol/fresh wallet with an average buy ≥ the floor, active within the evidence window and still holding. v2 also accepts a 30m burst or an entity group with identical amounts; snipe/bundler/sandwich tags disqualify.

LIVE THRESHOLDmin $1000 (robinhood $500) · evidence ≤ 1h · still-holding True · 1-in-5 control bypass

KNOB TO CHANGEquality_tx_gate.json → enabled / min_tx_usd / evidence_max_age_hours

EXCUSED WHEN1-in-5 deterministic control bypass (crc32 of chain:CA).

IMPACTSee the LIVE FUNNEL tab — the overwhelming majority of its refusals carry NO VERDICT.

EVIDENCETraced 260923: gmgn-cli exits 3 (circuit blocking, state-gated) and _gmgn() maps every non-zero exit to None, so our own pacing gate shedding the call is filed as absence of evidence.

⚠ Refusing on contention, not on wallet quality. Fix = bounded GMGN_MAX_WAIT patience + a distinct SHED label (#1373).

blacklistedLIVE · 0 fired 24hbad-deployer veto

WHAT IT CHECKSHard veto: the token's creator is on deployer_intel's blacklist.

LIVE THRESHOLDdeployer in DeployerDB blacklist for that chain

KNOB TO CHANGEdeployer_intel DB · admission_gate_config.json → gate_on_veto

EXCUSED WHENNever excused.

IMPACT0 rejected in 24h.

EVIDENCEConfirmed-bad deployers — categorical, not statistical.

exit_onlyLIVE · 0 fired 24hdistribution-signal veto

WHAT IT CHECKSHard veto: the source's only category is an exit/distribution signal, so it is not an entry.

LIVE THRESHOLDsource category == exit in conviction_gates.json

KNOB TO CHANGEconviction_gates.json categories · gate_on_veto

EXCUSED WHENNever excused.

IMPACT0 rejected in 24h.

EVIDENCECategorical — not a statistical threshold.

§ Scanner filters vs runners (#1300)

A filter that passes ~95% of control tokens is not selecting anything

Filter	Control pass	Runners passed
age_le_4h	99.1%	2/2
mcap_50k_4m	79.1%	2/2
mcap_50k_1m	77.6%	2/2
liq_ge_5k	82.0%	2/2
rug_le_10pct	98.7%	2/2
smart_ge_5	96.1%	2/2
renowned_ge_2	95.7%	2/2
holder_ge_100	97.5%	2/2

n=2 runners — far too few to tune on. Red bars are filters that reject almost nothing.

🗒 The programme — every issue by pipeline stage

status read live from todo.json

§ 🛰 SOURCES & BAKEOFF

#1349DONEP0ASTRA

GMGN + raw-source bakeoff: rank seed sources by measured resolved EV, numeric-first

#1355DONEP1CLAUDE

Source EV ranking rebuilt from tracked script source_ev_rank.py — old untracked ranking wrong in SIGN, not just hit5x

old ranking sign-flipped ~21 rows/signal; no n≥100 source significantly +EV

#1356DONEP1CLAUDE

trending_multi tracker outcomes are LOOK-AHEAD: entry booked median 241s (p90 961s) BEFORE decision — 179,338/179,338 rows fail leakage_audit decision_after_entry

entry booked median 241s (p90 961s) BEFORE the decision — look-ahead

#1350DONEP0ASTRA

trending_multi autopsy: 83.7% of outcomes are SCRATCH+SL_HIT - fix ladder/timing/buy-gate before re-ranking

83.7% of trending outcomes are SCRATCH+SL_HIT

#1358PROPOSEDP1ASTRA

Scout share cap: source occupies 39% of entries despite lowest priority

#1352DONEP0ASTRA

CLASS FIX: signal trade_class intent must reach exit engine - trending advertised SCALP but ran SWING for 80 days

#1353DONEP0ASTRA

SERVER-SIDE universal buy-size admission gate + 50k MC floor - apply >=2x$500 across ALL chains incl Robinhood

§ 📡 FEEDS & CAPTURE

#1364DONEP1BOTH

SYSTEM FIX: durable queryable capture of Dev/Hermes thread (200) + backfill missing messages (class: agent-visible feeds)

#1366DONEP2ASTRA

Build query layer over /opt/data/tg_capture for agent-accessible dev/alpha message search

#1367PROPOSEDP3CLAUDE

Deeper Telegram history backfill (HQ + alpha chats) beyond the 8,000-msg / 2.8-day HQ backfill

#1361PROPOSEDP1BOTH

Whale capture-lag TAIL: p90 657min (11h), only 39% of buys seen <5min — the detection half of the latency thesis

whale capture p90 657min (11h); only 39% of buys seen <5min

#1362PROPOSEDP2CLAUDE

WhaleMover RH alerts show $0 USD on SWING-small buys — token-amount only, USD valuation missing

#1330PROPOSEDP0BOTH

KARPATHY IS WRITE-ONLY TO A HUMAN — ~128 cycles of analysis exist ONLY in Chris's Telegram, nowhere on disk

#1331PROPOSEDP1CLAUDE

Alpha report corpus is unread too — alpha_kb, ALPHA_INSIGHTS, alpha_layers_plan, alpha_feature_scoreboard have no consumer

§ ⚗ FILTERS & GATES

#1321DONEP1ASTRA

Master Filter Dashboard - every feed's filters + measured impact across admission_gate/TDE/sniper/oggy_cap/signal_filters - Chris 260921 open ↗

per-feed filters + measured impact — now merged into this page

#1300APPROVED?

RUNNER-RECALL MASTER FABRIC open ↗

cross-feed filter recall vs top runners — LIVE data dashboard

#1317DONEP0ASTRA

F1: Outcome coverage

#1329PROPOSEDP0BOTH

THE ONLY POSITIVE-EV BUCKETS IN THE SYSTEM ARE THE BUY-PROFILE BANDS — unexploited for 45 cycles

buy-profile bands = the only positive-EV buckets; unexploited 45 cycles

#1322DONEP1ASTRA

OGGY net-buy volume cap MEASURED gate: shadow live, flip to enforce after calibration - Chris 260921

#1312APPROVEDP0BOTH

Convergence detectors STARVE THEMSELVES: one 429 trips the gmgn-gate circuit and blocks the remaining 6-7 calls of the same loop

#1357IN PROGRESSP0CLAUDE

DEFER_YOUNG paper A/B: tokens <60min at entry are 38% of post-fix losses — defer entry to age>=60m

tokens <60m = 38% of post-fix losses; the cap excluded 2/2 runners

#1373DONEP0CLAUDE

quality_tx gate refuses 98.6% of in-scope candidates on MISSING INPUT (GMGN unavailable), not on wallet quality — decide fail-open vs fail-closed

quality_tx refuses 98.6% on a SHED call, not on wallet quality

§ 🧮 JEV SEMANTIC LAYER

#1336PROPOSEDP0ASTRA

EP-TYPESAFE1

#1337PROPOSEDP0ASTRA

Jev calibration harness on frozen Hermes decisions

#1338PROPOSEDP1ASTRA

Jev paper-policy critic + parameter-permutation ranker (shadow)

#1339PROPOSEDP1ASTRA

Jev operations triage pilot: roadmap/Exchange/feed-state

#1340PROPOSEDP2ASTRA

Jev alpha evidence router + claim verifier (shadow)

#1341PROPOSEDP0ASTRA

Jev Lane-0: outcome-labeled narrative/social semantic features on resolved paper ledger (shadow)

v1 was label-leaked (size); leak-free = no signal

#1342DONEP0ASTRA

Jev blast-radius: dedicated key+credit cap, per-lane daily call cap, hash cache (COST BIBLE laws 12/13)

blast-radius: dedicated key+cap+cache — GATES all paid Jev runs

#1346IN PROGRESSP1BOTH

Jev capability expansion map — gated semantic-judgment lane portfolio (broader than token selection)

#1347IN PROGRESSP1BOTH

Alpha-text corpus inventory + timing/provenance tags -> Jev lane routing (whalemover FA, alpha_kb, Karpathy, forensic, dashboards)

#1354PROPOSEDP1ASTRA

Jev probabilistic-decision integration: calibration prior for edge-support + regime-conditional boost test

#1359DONEP1ASTRA

J1 Jev cascade in front of Hermes LLM jobs — verify with Jev, escalate only low-confidence to the LLM

Jev 99% routing @ 0.2s/$0.00003 vs flash 91% — VALIDATED (EXCHANGE #196)

#1360PROPOSEDP1CLAUDE

J2 Telegram alpha router (Jev) — route scanner/relay/retrospective/human; quarantine retrospective as leak, harvest as free outcome labels

#1363DONEP0ASTRA

J1 Jev cascade candidate survey + design proposal

§ ⏱ LATENCY & OUTCOMES

#1328PROPOSEDP0BOTH

WE CANNOT MEASURE OUR OWN LATENCY — every archive verdict is confounded and cannot be de-confounded

signal→entry p50 208s / p95 1550s (26min); event_ts on only 2% of rows

#1327PROPOSEDP1CLAUDE

CORRECTION: kien_convergence measures -4.84pct over N=865 — I argued to revive a detector our own data says to ARCHIVE

#1326PROPOSEDP0BOTH

RECURRENCE ESCALATION: a recommendation repeated N times must raise its own priority — Karpathy band filter is at 47 cycles

#1368PROPOSEDP1BOTH

wallet_outcomes + wallet_scores stale 494h (20 days); inversion_html 137h — System Assessor 4/4 runs, unowned

🔬 Evidence traps

each one has already produced a false verdict on this pipeline

§ Trap 01 · units

📏 Every *_pct is a decimal

5x means max_gain_pct ≥ 4.0, not 400.
Cost: a false “hit5x = 0 across 245k outcomes” that would have steered work to re-selecting tokens when the leak was exits.
Truth in the last 300MB (23,184 rows): 652 hit 2x+, 135 hit 5x+.

§ Trap 02 · per-row vs per-signal

🔢 ~21 rows per signal

signal_outcomes.jsonl writes one row per window, 20m…72h.
A per-row count flipped wallet_buysize_edge from +2.2% to −17.2%.
Aggregate per signal; rerun source_ev_rank.py rather than counting rows.

§ Trap 03 · leakage

☣ Filter on leakage_audit.valid

All 179,338 trending_multi rows book entry a median 241s before the decision.
Jev v1’s selection replay leaked the label through size, rewritten at TRAIL close.
The paper book is unaffected — prefer it for any selection claim.

§ Trap 04 · absent input

⚠ n=0 is a refusal, not a verdict

quality_tx refuses on a shed call and files it as missing evidence.
The admission gate once had 99.72% of its shadow rejects come from an empty volume_snapshot.
Give UNKNOWN its own branch and never let a default stand in for a measurement.

§ Trap 05 · dead producer vs working filter

🕳 An outage and a good filter look identical

The 260909 intake collapse superimposed on the 260915 selection rebuild made “cut >90% of losing intake” unmeasurable.
Check the producer runs before crediting a filter; declare gate epochs forward, never backdated.
Rebuild inventories on output mtime and row counts, never on last_status — wrong 4/4, both directions.

§ Trap 06 · identifiability

🧩 Selection is not separable from latency and exits

Every resolved label = f(selection, latency, exit), and event_ts exists on 2% of rows.
Stratify by entry_latency_s band, and treat “selection improved” as unproven until the receipt lands (#1328).
Beware peak_ret wick inflation — control shows 45.6% peak≥2x, which cannot be traded.

§ Trap 07 · causation

🧬 A missing dependency is not a traced cause

260923: I verified curl_cffi was absent, found a matching roadmap item, and published it as the cause of the quality_tx blockage in four places.
Wrong — the gate uses a subprocess, not that client. Two true facts sat next to each other and adjacency was read as causation.
Run the failing call, read its exit code, and name the module the path actually imports.