gavindiaz/gmgn-skills
1
0
Fork 0
mirror of https://github.com/GMGNAI/gmgn-skills.git synced 2026-11-02 07:46:11 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix(wallet-analysis): only a fully read book can lift the count veto

Browse Source 
- A clear reading on a truncated book no longer lifts a launcher's
  count veto: unread older positions can hold any amount of own-token
  gains offset by losses the all-time net cannot reveal, so no sample
  ratio can bound them. Such a launcher now gets G1 ⚪ and
  "HOLD OFF · its launches look minor, but its older record was not
  read"; a clear reading keeps paging toward the book's natural end.
- Tag keys are stripped again, so " wash_trader " still matches.
- A wash check blocked by an unreadable launch list advises re-running,
  not configuring a key.
- An empty own-token set while the wallet launched is "not measured".
- stats_30d catches a CLI timeout as a gap.
- Docs: verdict row 10c, sixteen outcomes, paging and timeout wording.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
chenjunfeng
2026-10-15 15:05:02 +08:00
co-authored by Claude Opus 5.5
parent 76f9408198
commit 76abc2ca9b
2 changed files with 52 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files
skills/gmgn-wallet-analysis/SKILL.md
+10 -9
View File
@@ -59,7 +59,7 @@ detail.
Verdict is a pure function of the gates — G1 and G2 are vetoes, G3 and G4 change *what you do* rather than whether you act:
Fifteen outcomes, short-circuited in this order:
Sixteen outcomes, short-circuited in this order:
| # | Gates | Verdict | Fixture |
|---|-------|---------|---------|
@@ -72,6 +72,7 @@ Fifteen outcomes, short-circuited in this order:
| 5 | G1 ❌ via `token_num < 5` | ⚪ NO READ · only N tokens traded | `thin-sample` |
| 6 | G2 ❌ | 🔴 DO NOT COPY · it has stopped making money | `cooled-star` |
| 10b | G1 ⚪ via a self-launch range straddling 50% | 🟡 HOLD OFF · over half its profit may come from its own launches | — |
| 10c | G1 ⚪ via a clear reading on a truncated book of a launcher by count | 🟡 HOLD OFF · its launches look minor, but its older record was not read | — |
| 10 | G1 ⚪ | 🟡 HOLD OFF · a wash-trading flag we cannot check | `unverifiable-wash` |
| 7 | G3 ❌ **and** G4 ❌ | 🟡 WATCH, DO NOT COPY · you cannot get its fills, and it never cuts | `unreachable-and-no-cut` |
| 8 | G3 ❌ | 🟡 WATCH, DO NOT COPY · you cannot get its fills | `sniper-bot` |
@@ -185,7 +186,7 @@ then can they move the verdict.
| lo ≥ 50% | ❌ | `DO NOT COPY · most of its profit came from tokens it launched` |
| lo < 50% ≤ hi | ⚪ — only when no measured G1 failure applies; a launcher by count still fails on the count | `HOLD OFF · over half its profit may come from its own launches` |
| 20% ≤ hi < 50% | ✅ — but a launcher by count still fails on the count | A warning line with the share (a range when suspects remain) and what it made on other people's tokens |
| hi < 20% | ✅ | The launch count stops vetoing — G1 says the profit is from trading, and G3/G4 are evaluated instead of marked not-applicable. On a truncated sample this needs the sample's gains to cover at least half of the all-time realized profit; otherwise the count still vetoes |
| hi < 20% | ✅ — when the whole book was read | The launch count stops vetoing — G1 says the profit is from trading, and G3/G4 are evaluated instead of marked not-applicable. On a truncated book a launcher by count gets ⚪ instead: unread older positions can hold any amount of own-token gains, offset by losses that the all-time net cannot reveal, so no sample ratio can clear them |
| not measured — fewer than 3 winning positions, `created-tokens` or `holdings` unavailable, or the paging cut off before page 6 (a failed call, an empty or malformed page that still points onward, a repeated cursor) | — | Falls back to the count test, unchanged |
Only positive realized gains are attributed. A launcher by count whose own tokens netted zero or
@@ -201,12 +202,12 @@ The rules that hold it together:
--order-by last_active_timestamp`. Every other sort returns at most 400 rows, and for a wallet
with more positions than that those are not the global top, so the largest closed gains can
be missing. Paging stops at page 6 at the earliest, once the last 3 pages agree on the verdict
class and — for a clear reading — the sample covers half of the all-time realized profit; at
20 pages at most. A cut-off (failed call, malformed page, repeated cursor) is never read as
the end of the book. A truncated sample names its start date. For a wallet that is not a
launcher by count, a share that could not be measured is said so in G1. EVM contract
addresses are compared case-insensitively; base58 addresses exactly, the creator check
included.
class, at 20 pages at most; a clear reading on a launcher by count keeps paging, because only
the book's natural end can lift its count veto. A cut-off (failed call, malformed page,
repeated cursor) is never read as the end of the book. A truncated sample names its start
date. For a wallet that is not a launcher by count, a share that could not be measured is said
so in G1. EVM contract addresses are compared case-insensitively; base58 addresses exactly,
the creator check included.
- **Those closed rows also feed the conviction share — minus the wallet's own launches.** A
wallet's biggest wins are usually closed, so the open book alone under-reads it and can
corroborate `wash_trader` on a record that its closed positions refute. Own and suspected
@@ -330,7 +331,7 @@ issues zero API calls and finishes in seconds.
Restore it rather than testing on live wallets alone. Live data moves — the 1d window is
rolling and `activity` is a sample — so two runs minutes apart differ for reasons unrelated to
your change, and a full dossier costs weight 23 or more against a bucket of 20, so a handful of
verification runs rate-limits the account. Most of the fifteen verdict branches are also
verification runs rate-limits the account. Most of the sixteen verdict branches are also
close to impossible to reach on demand: you would need to find a wallet whose wash-trade flag
is corroborated *right now*.
skills/gmgn-wallet-analysis/analyze.py
+42 -21
View File
@@ -186,6 +186,9 @@ ZH = {
" (positions active since {0})": "（统计 {0} 之后活跃过的仓位）",
"per-token realized P&L unavailable: {0} — the share of profit from its own launches was not measured": "按币的已实现盈亏未取到：{0} —— 自发币盈利占比本次未评估",
"🏭 launches tokens, profits from trading": "🏭 发币方（但盈利靠交易）",
"launched {0:,} tokens; in the positions read, at most {1} of realized gains came from them — the profit there is from trading {2}{3}. Its older positions were not read, so they are not verified": "发过 {0:,} 个币；在读到的仓位里，最多只有 {1} 的已实现盈利来自这些币 —— 这部分盈利来自交易 {2}{3}。更早的仓位没有读到，无法核实",
"HOLD OFF · its launches look minor, but its older record was not read": "先别跟 · 自发币占比看起来很小，但更早的记录没有读完",
"Re-run once its launch list can be read. Do not size off this record first.": "等发币列表能取到时再跑一次。在那之前别按这份战绩下注。",
"where the profit came from cannot be checked — its launch list could not be read, so its own tokens cannot be told apart from the rest; the {0} in this window is neither confirmed nor refuted": "盈利来源无法核验 —— 它的发币列表没有取到，自己发的币和其他币分不开；这个窗口的 {0} 既不能确认也不能排除",
"launched {0:,} tokens — the share of profit from them was not measured (it needs at least 3 winning positions and the per-token P&L pull)": "发过 {0:,} 个币 —— 来自这些币的盈利占比本次未评估（需要至少 3 个盈利仓位，并取到按币的盈亏数据）",
"DO NOT COPY · one token made all the money": "别跟 · 全靠一个币赚钱，复制不了",
@@ -860,7 +863,7 @@ def read_tags(raw_tags):
"""Return [{key, emoji, sev, name, meaning, known}] — unknown tags kept verbatim."""
out = []
for t in raw_tags or []:
key = clean_text(t, 32)
key = clean_text(t, 32).strip()
row = TAGS.get(key.lower())
if row:
emoji, sev, name, meaning = row
@@ -935,10 +938,9 @@ SELF_MIN_PAGES = 6
SELF_STABLE = 3
SELF_MAX_PAGES = 20
SELF_ROWS_PER_PAGE = 100 # `--limit 50` asked for; a page claiming more is cut, not trusted
# A clear reading on a truncated sample only lifts the count veto once the sample's gains
# cover this share of the wallet's all-time realized profit — otherwise older own-token
# gains beyond the last page read could be what the record is made of.
SELF_COVER = 0.50
# A clear reading lifts the count veto only when the whole book was read. On a truncated
# sample, unread older positions can hold any amount of own-token gains, offset by losses
# that net all-time profit cannot reveal — so no coverage ratio can bound them.
# A token address from an API row goes into a `token info` argv. It is third-party data, so it
# must look like an address before it gets there — never like a flag.
TOKEN_ADDR_RE = re.compile(r"^(0x[0-9a-fA-F]{40}|[1-9A-HJ-NP-Za-km-z]{32,44})$")
@@ -1172,7 +1174,7 @@ def collect(chain, wallet, gaps):
try:
d["stats_30d"] = first_row(cli(["portfolio", "stats", "--chain", chain, "--wallet", wallet,
"--period", "30d"]))
except Gap as e:
except (Gap, subprocess.TimeoutExpired) as e:
d["stats_30d"] = {}
gaps.append(f"stats_30d: {e}")
@@ -1191,7 +1193,9 @@ def collect(chain, wallet, gaps):
ct = d["created_tokens"]
own = own_token_set(ct)
complete = launch_list_complete(own, ct, i(common.get("created_token_count")))
all_time = f(((d.get("profits_all") or {}).get("total_realized_profit")), 0.0)
pnl7 = d["stats_7d"].get("pnl_stat") or {}
created_n = i(common.get("created_token_count"))
dev_by_count = created_n > 0 and created_n > 0.5 * max(1, i(pnl7.get("token_num")))
rows, cursor, classes, cursors, cut = [], None, [], set(), None
# Every page is the same critical-auth route that just failed for want of a key —
# skip it rather than fail again and print the same gap twice.
@@ -1237,9 +1241,10 @@ def collect(chain, wallet, gaps):
classes.append(cls)
# Stable includes stably unmeasurable (None): past the floor, a book with too
# few winners will not grow a verdict, so reading 20 pages would be waste. A
# stable "clear" also needs the sample to cover the all-time record.
covered = cls != "clear" or (all_time > 0 and sp["gains"] >= SELF_COVER * all_time)
if page + 1 >= SELF_MIN_PAGES and len(set(classes[-SELF_STABLE:])) == 1 and covered:
# "clear" on a launcher by count keeps paging: only the book's natural end can
# lift its count veto.
settled = cls != "clear" or not dev_by_count
if page + 1 >= SELF_MIN_PAGES and len(set(classes[-SELF_STABLE:])) == 1 and settled:
break
if cut is not None:
pages_read, why = cut
@@ -1254,7 +1259,6 @@ def collect(chain, wallet, gaps):
cursor = cursor or "cut"
d["realized_rows"] = rows
d["realized_truncated"] = bool(cursor)
d["realized_all_time"] = all_time
# Resolve the largest suspects only when they could move the verdict: a range that
# straddles a line. Weight 1 each; an unresolved one stays a suspect.
sp = self_launch_split(rows, own, complete, wallet, {})
@@ -1742,9 +1746,11 @@ def compute(d, latency_s, my_size):
# ── self-launch attribution (see self_launch_split) ──
m["self_lo"] = m["self_hi"] = None
m["self_clear"] = False
m["self_clear"] = m["self_clear_partial"] = False
rr = lc["rows"]
if lc["ct"] and rr and not d.get("self_partial"):
# An empty own set while it launched tokens means the list was not read: own tokens cannot
# be told apart, so nothing is measured (and nothing can be cleared).
if lc["ct"] and lc["own"] and rr and not d.get("self_partial"):
sp = lc["sp"]
m["self_lo"], m["self_hi"] = sp["lo"], sp["hi"]
m["self_own_usd"], m["self_ext_usd"] = sp["own"], sp["ext"]
@@ -1756,12 +1762,12 @@ def compute(d, latency_s, my_size):
m["self_since"] = (time.strftime("%Y-%m-%d", time.gmtime(min(ts)))
if d.get("realized_truncated") and ts else None)
# A launcher by count whose profit demonstrably comes from other people's tokens is
# a trader who also launches. The count stops vetoing; it is still printed. On a
# truncated sample "demonstrably" needs the sample to cover the all-time record.
all_time = f(d.get("realized_all_time"), f(pall.get("total_realized_profit"), 0.0))
covered = (not d.get("realized_truncated")
or (math.isfinite(sp["gains"]) and all_time > 0 and sp["gains"] >= SELF_COVER * all_time))
m["self_clear"] = m["self_hi"] is not None and m["self_hi"] < SELF_WARN and covered
# a trader who also launches: the count stops vetoing; it is still printed. Only a
# fully read book is demonstrable. A clear reading on a truncated one is kept as
# `self_clear_partial`: it holds G1 open (⚪) instead of vetoing or passing it.
clear_reading = m["self_hi"] is not None and m["self_hi"] < SELF_WARN
m["self_clear"] = clear_reading and not d.get("realized_truncated")
m["self_clear_partial"] = clear_reading and bool(d.get("realized_truncated"))
# ── friction: the numbers that decide whether the edge survives being copied ──
# Per-trade net is the yardstick everything else is measured against. A wallet netting
@@ -1913,7 +1919,7 @@ def gates(m):
pct(m['self_lo']), usd(m['self_own_usd']), m['self_own_n'],
joinsym(m['self_own_top']), self_window(m)),
)
elif m["is_dev"] and not m["self_clear"]:
elif m["is_dev"] and not m["self_clear"] and not m["self_clear_partial"]:
g["G1"] = (
False,
T('launcher wallet: created {0} vs traded {1} — its win rate and entry timing are self-authored, not a market read', m['created_tokens_n'], m['token_num']),
@@ -1940,6 +1946,14 @@ def gates(m):
T('{0}–{1} of realized gains may come from tokens it deployed — its launch list is truncated and the rest could not be confirmed{2}',
pct(m['self_lo']), pct(m['self_hi']), self_window(m)),
)
elif m["is_dev"] and m["self_clear_partial"]:
# Recent positions clear it, older ones were not read: neither the count veto nor a
# pass is earned. Ranked with the other open questions, below every measured failure.
g["G1"] = (
None,
T('launched {0:,} tokens; in the positions read, at most {1} of realized gains came from them — the profit there is from trading {2}{3}. Its older positions were not read, so they are not verified',
m['created_tokens_n'], pct(m['self_hi']), joinsym(m['self_ext_top']), self_window(m)),
)
else:
if m["pcr_represents_record"]:
pcr_txt = T('profit concentration {0}', pct(m['pcr']))
@@ -2176,9 +2190,16 @@ def verdict(m, g):
return ("🟡",
T('HOLD OFF · over half its profit may come from its own launches'),
T('Check its launch record before sizing off this one.'))
if (p["G1"] is None and m.get("self_clear_partial")
and not any(t["sev"] == "veto_g1" for t in m["tag_info"])):
return ("🟡",
T('HOLD OFF · its launches look minor, but its older record was not read'),
T('Check its launch record before sizing off this one.'))
if p["G1"] is None:
return ("🟡",
T('HOLD OFF · a wash-trading flag we cannot check'),
T('Re-run once its launch list can be read. Do not size off this record first.')
if m.get("conviction_blocked") else
T('Configure GMGN_PRIVATE_KEY and re-run. Do not size off this record first.'))
if p["G3"] is False and p["G4"] is False:
return ("🟡",
@@ -2433,7 +2454,7 @@ def archetype(m):
tags = []
if m["self_lo"] is not None and m["self_lo"] >= SELF_VETO:
tags.append(T('🏭 launcher (marks its own homework)'))
elif m["is_dev"] and m["self_clear"]:
elif m["is_dev"] and (m["self_clear"] or m["self_clear_partial"]):
tags.append(T('🏭 launches tokens, profits from trading'))
elif m["is_dev"]:
tags.append(T('🏭 launcher (marks its own homework)'))