TD Trap Detector v3.9.4

Stock promotion risk scan

It doesn't ask whether a stock is good. It asks who is pushing it at you.

Eight signals, each returning hit, miss, or insufficient — with a source attached. Your own wording is scored as evidence. The result is a fixed JSON report and one of four risk levels, not a paragraph of hedged advice.

8 signals
3-state findings
4 risk levels
0 lines of code

See the difference Read the scan

TRAP SCAN RESULT JSON

{

"ticker""300XXX · Hengxun AI"

"trap_score"1

"trap_level""🔴 HIGHLY SUSPICIOUS"

"signals_hit""7 / 8"

── signal 1 ─────────────────────────

id1 · low-quality account amplification

evidence23 near-identical posts, 30–200 followers each

severitycritical

── signal 3 ─────────────────────────

id3 · paid community funnel

evidence5 accounts pushing the same VIP group QR

severitycritical

─────────────────────────────────────

"user_keyword_boost"+4

"recommendation""🔴 Strongly advise avoiding. Any recommendation that asks for a fee first is a prelude to fraud."

}

7 / 8 hit · wording boost +4 · trap_score 1 / 10 fictional teaching ticker

02 · Before / After

One question, two answers — from the same search results.

Both sides below start from the same prompt and retrieve the same 23 articles. The only variable is whether a framework decides what those articles count as.

Shared input · identical on both sides A friend recommended Hengxun AI (300XXX) to me. He said there's inside information, the main force has finished building its position, and the target is a double. Should I get in now?

B Without Trap Detector

—"Inside information is not reliable" — stated, then dropped

—"The stock has recently attracted notable market attention"

—Source quality never examined: 30–200 followers each

—No count of how many posts, or how similar

—Wording treated as tone, not as evidence

—Ends at "decide according to your own risk tolerance"

The failure mode — those 23 articles are the promoter's promotional material. Retrieved without a framework, they read as market consensus and get repeated back as such. The search did not inform the answer; it laundered the campaign.

A With Trap Detector

✓"inside information" scored as evidence: +2

✓Title similarity + account age = signal 1 hit

✓Follower counts recorded: 30–200, registered <2 months

✓23 near-identical posts counted as volume, not consensus

✓8 signals each return hit / miss / insufficient

✓Ends at a tier: 🔴 highly suspicious, avoid

What changed — the same 23 articles were reclassified from sources to exhibits. Noise became evidence; that reframing, not the extra questions, is the actual product.

Dimension	Without		With
Conclusion	"Consider your own risk tolerance"	→	🔴 Highly suspicious · avoid
Signal coverage	No scan performed	→	8 / 8 each given a state
Evidence	0 recorded	→	23 posts + account age + follower counts
Your wording	Dropped as tone	→	Scored: +2 inside · +1 friend / double
Search results	Treated as information	→	Treated as exhibit, turned against the promoter
Output shape	Prose paragraph	→	Fixed JSON + 8-row state list

03 · Capability

Eight signals, three states, and no way to skip one.

Every number below is a structural count read from the source. There are no runtime metrics — this skill has no downloads, installs, or telemetry to report.

Signals scanned

all 8 must return a state

Finding states

hit · miss · insufficient

Risk levels

🟢 🟡 🟠 🔴

Lines of code

pure instruction · no deps

Mandatory full scan

All eight signals must produce a result. None can be silently dropped because search came back thin.

8 / 8 · required · a skipped signal invalidates the run

Three-state findings

Insufficient is a legal answer. It gives the model a way to say "I don't know" instead of defaulting to safe.

hit / miss / insufficient · no silent default

Your wording is evidence

How you heard about the stock is itself a signal. "Inside information" and "guaranteed" score +2; "friend recommended", "must rise", "double" score +1.

+2 tier · +1 tier · summed into the level

Evidence binding

At least one concrete source URL is required, and the recommendation field cannot be left empty.

≥1 source · recommendation mandatory

Four risk levels

Hit counts map to tiers, and at four or more hits the opening words of the recommendation are forced.

0–1 🟢 · 2–3 🟡 · 4–5 🟠 · 6+ 🔴

Structured JSON output

Ticker, score, level, hit array with severity, boost, recommendation, warning phrases — machine-readable downstream.

7 fields · parsable · pairs with deep-analysis

Risk mapping · four tiers

🟢 SAFE0–1 signals hitno organised promotion found

🟡 NOTE2–3 signals hitverify the source

🟠 ALERT4–5 signals hitcaution, wording forced

🔴 SUSPICIOUS6+ signals hitstrongly advise avoiding

Two things read backwards on first contact: trap_score is reversed — 10 is safest, 1 is worst — and signals_hit only carries hits. Missed signals live in the prose state list, not in the JSON array.

Signal composition · 6 : 2

Promotion side1 · 2 · 3 · 6 · 7 · 8 — how it is being sold6

Asset side4 · 5 — fundamentals and price action2

This ratio is the capability boundary, stated as a number. Six of eight signals inspect the promotion, not the company — so what it detects is the organisation of a sales campaign, not manipulation itself. A corner no one is pushing to you will not appear.

Execution note. The two asset-side signals depend on data-fetch scripts from a companion skill that is not installed here. In this environment the honest state for them is insufficient — real coverage is 6 / 8.

04 · Workflow

Say what you heard, then let it be scanned.

Four steps. The first one is where most of the evidence is either preserved or destroyed.

Repeat what you heard

Paste the promoter's words verbatim. Do not tidy them into polite language.

Why — only "inside information" scores +2; "exclusive research channel" scores 0

Scan all eight signals

Each returns hit, miss, or insufficient, with a source. Promotional content is read as an exhibit, not as information.

Trap — no data means insufficient, never "miss"

Weight and classify

Keyword boost is added, hits are counted, and the total is mapped to one of four risk levels.

Read carefully — trap_score runs backwards: 10 safe → 1 worst

Emit report and next step

Fixed JSON plus the eight-row state list, a recommendation that cannot be blank, and forced warning phrases.

Only hits appear in signals_hit; misses live in the list

Important boundary

This answers "is this being pushed at me in an organised way?" — not "is it worth buying?" A manipulated stock that nobody is promoting to you will scan clean. A genuinely bad company with no campaign behind it will also scan clean. The detection target is the sales operation, not market manipulation in general.

05 · Real cases

Three runs, on real tickers.

Designed as a control set rather than three variations of one demo: a negative control, a grey-zone case, and a historical replay. A negative control, a live small cap, and a replay of a case the regulator has since ruled on. All three were run in English; the ticker, the date and the retrieval count are recorded under each one.

Case 01 · Negative control Kweichow Moutai · 600519 2026-09-15 19:34 GMT+8 English prompt → English report

"A friend said it's a guaranteed win." The scan disagreed.

The control case. The prompt is loaded with trigger phrases — "a friend recommended" (+1) and "guaranteed win" (+2) — and the ticker is the most heavily covered large cap in the market. If the scan flagged this, it would flag everything. It returned 0 / 8.

Result

0 / 8 · 🟢 LOW RISK

trap_score 10 · wording boost +3 unused

Eight misses, zero hits, zero insufficient. The +3 wording boost had nowhere to land — with no hits, the score stays at 10. The scan measures promotion, not quality: a company can post its first-ever annual revenue and profit decline and still come back clean.

01 Input

Verbatim prompt, exactly as submitted.

Prompt @skill:trap-detector A friend recommended I buy Kweichow Moutai (600519) and said it's a guaranteed win, told me to just hold it long term. Check whether this is a pump-and-dump. Scan all eight signals one by one — for each one tell me whether it's a hit, a miss, or insufficient data, and give me the source behind every conclusion. Finish with the risk level and your recommendation.

Kweichow Moutai · 600519

Large-cap baijiu. 2025 revenue ¥168.838bn (−1.21%), net profit ¥82.32bn (−4.53%) — the first annual double decline since listing. 2026H1 revenue +1.47%, but Q2 net profit −6.9%.

¥1,272.75−0.41%PE TTM 19.54turnover 0.11%YTD −5.66%institutional visits 0

02 Scan · eight signals

Every signal returned MISS. The basis for each one:

Signal	State	Basis
1 · Low-quality accounts pushing in unison	MISS	All coverage from licensed institutions (25 houses, 92% positive). Institutional visits 0 (rank 116/126). Fund holding 1.18%, −0.69pp over 573 funds.
2 · Templated pitch language	MISS	No phrasing matrix found. One borderline post (2026-09-13) carried a disclaimer and invalidation conditions — logged as an edge observation, not a hit.
3 · Paid VIP group / livestream funnel	MISS	Only generic regulator material surfaced (CSRC 2018 warning,Beijing Stock Exchange case studies). None of it points at 600519.
4 · Fundamentals detached from attention	MISS	Falling earnings, flat-to-neutral sentiment, no institutional following. Both sides move in the same direction — there is no divergence to detect.
5 · Abnormal price action	MISS	Amplitude 1.03%, turnover 0.11%, YTD −5.66%. Low-volume base building, no pump-then-distribute shape.
6 · "Teacher" persona promotion	MISS	Hits were signed analyst reports with SAC numbers, and post-hoc investor commentary. No lead-generation persona.
7 · Cross-platform coordinated push	MISS	All social heat was consumer-level (counterfeit-baijiu controversy, app launch, packaging typo). No ticker campaign.
8 · Fake research / fabricated news	MISS	Fake material exists — franchise-recruitment scams,a debunked wholesale-price rumour — but every instance targets the product, not the share price.

03 Output

TRAP SCAN REPORT600519

signals_hit	0 / 8
signals_breakdown	hit 0 · miss 8 · insufficient 0
user_keyword_boost	+3 (friend recommended +1 · guaranteed win +2)
trap_score	10 / 10
risk_level	🟢 LOW RISK

Safe next stepNo organised promotion detected. The recommendation wording is the risk here, not the stock — treat "guaranteed win" as a signal about the person saying it. Note the company is facing its first annual decline in revenue and profit; that is a valuation question, outside this scan's scope.

04 Evidence

0/8

signals hit

trap_score

wording boost

covering houses

institutional visits

retrievals

Run record · case-materials/trap-detector/01-basics-en/run-log.md · screenshots shot-a-en-chat-01…05.png · 5 retrievals, all English-keyword

Screenshots captured from the live scan — click any frame to open it at full size.

Run · 01 — English prompt verbatim, app-reported 3m17s, the report header, the weighting table and the start o — **Run · 01** — English prompt verbatim, app-reported 3m17s, the report header, the weighting table and the start of the signal table.

Run · 02 — the full 8-signal three-state table, 0/8 hits, 🟢 LOW RISK, trap_score 10, and the Chinese/English d — **Run · 02** — the full 8-signal three-state table, 0/8 hits, 🟢 LOW RISK, trap_score 10, and the Chinese/English divergence section opening.

Run · 03 — the divergence analysis, including the AI content-farm edge case, and the start of the structured J — **Run · 03** — the divergence analysis, including the AI content-farm edge case, and the start of the structured JSON output.

Run · 04 — the JSON body: trap_score 10, boost_detail, warning_phrases, and the 14 sources. — **Run · 04** — the JSON body: trap_score 10, boost_detail, warning_phrases, and the 14 sources.

Run · 05 — the four-item completion check and all honest boundaries. — **Run · 05** — the four-item completion check and all honest boundaries.

Honesty boundary

The Chinese run on the same ticker surfaced a fraud-exposure article naming an "Moutai VIP group" with livestream call-outs. The English run surfaced none of it. English search has almost no coverage of Chinese private-domain exposure content — the false-positive risk disappeared here, and so did the recall. The clean result is partly a property of the language the search ran in.

CASE 01 · BASELINE · 0 / 8 · 🟢 LOW RISK · trap_score 10 · RUN 2026-09-15 19:34

Case 02 · Core value Jinniu Chemical · 600722 2026-09-15 19:41 GMT+8 English report · Chinese sources

Six signals fired. Two of them were the ones that matter.

A small cap whose attention came from a film-title pun. Six of eight signals hit — but four of those six were the same thing counted four times. This is the case that shows both what the scan catches and how its counting inflates one finding into four.

Result

6 / 8 · 🔴 HIGH RISK

trap_score 4 · wording boost +1

Six hits, two misses. The two that actually cost money — a 190× multiple against ¥47.93m of profit, and a distribution structure where institutions sold ¥136.7m to retail on the day — are weighted the same as four signals that all describe "people are talking about this stock".

01 Input

Verbatim prompt, exactly as submitted.

Prompt @skill:trap-detector Lately my Douyin feed and the stock forums are full of Jinniu Chemical (600722) — the attention on it is unusually high. Check it for pump-and-dump risk. I want to know whether this is organised promotion or just organic market attention. Scan all eight signals, give me hit / miss / insufficient for each with sources attached, then the risk level and your recommendation.

Jinniu Chemical · 600722

Small-cap chemical. Market cap ¥10.756bn against 2025 net profit of ¥47.93m. Company's own 2026-03-17 filing puts static PE at 237.88× versus an industry 35.97×.

YTD +166.16%5d −20.03%turnover 17.51%PE TTM 190.98dividend yield 0.00%20–21 dragon-tiger listings

02 Scan · six hits

Signals 1, 2, 4, 5, 6, 7 hit. Signals 3 and 8 missed.

Signal 1 · hit

Three accounts, same copy

2026-09-08/09: near-identical posts from three accounts. A 2026-08-29 forum post describes the mechanism itself — teams running hundreds of accounts on shared templates.

Signal 2 · hit

Shared phrase book

"Scooped up from deep water", "the main force believes this logic", "wash-and-consolidate" — plus one reused support/resistance ladder (15.8 / 17.0 / 20.0 / 21.2).

Signal 4 · hit

190× for ¥47.93m

2025 revenue ¥469.77m (−5.92%), net profit ¥47.93m (−8.92%). 2026H1 profit +31.42% — to ¥35.10m. Zero institutional visits, 100% neutral ratings.

Signal 5 · hit

Distribution, mirrored

On 2026-09-15: institutions net −¥136.7m, retail net +¥136.7m. On the 2026-09-02 limit-down: institutions net −¥299m. Seats include the "Lhasa Tiantuan" / "Wenzhou Gang" hot-money desks alongside Shanghai Stock Connect and Goldman.

Signal 6 · hit

Self-styled "dragon" personas

Accounts giving daily levels and direction ("target ¥23, not leaving until it hits"). Direction without a fee — the promotion is the persona, not a paywall.

Signal 7 · weak hit

A pun, not a campaign

The Niu Lai wordplay spread across Douyin / Bilibili / forums and topped East Money's trending list on 2026-08-16 — on a day the stock was limit-down. Organic, with no evidence of paid placement.

Signal 3 · miss

No funnel found

No paid group or livestream entry for this ticker. The one forum post describing such gangs is descriptive — a warning about the pattern, not evidence it operates here.

Signal 8 · miss

Company denies rumours

2026-03-17 and 2025-06-18 filings both state no undisclosed material information. The circulating "inquiry letter" was a satirical post about the film.

03 Output

TRAP SCAN REPORT600722

signals_hit	6 / 8
signals_breakdown	hit 6 · miss 2 · insufficient 0
user_keyword_boost	+1 (Douyin / forums "full of it", analogised to "someone in the group chat said")
trap_score	4 / 10
risk_level	🔴 HIGH RISK

Safe next stepDo not act on the attention. The position to avoid is buying into a 190× multiple while institutions are net sellers. If already holding, note that the distribution pattern — not the chatter — is the part that has historically preceded the loss.

04 Evidence

6/8

signals hit

190.98

PE (TTM)

−¥136.7m

institutional net flow

+166%

year to date

20–21

dragon-tiger appearances

screenshots archived

Run record · case-materials/trap-detector/02-core-value-en/run-log.md · screenshots shot-b-en-chat-01…06.png · 6 retrievals, 5 Chinese

Screenshots captured from the live scan — click any frame to open it at full size.

Run · 01 — English prompt verbatim, app-reported 4m1s, the report header (ticker / scan time / trigger) and th — **Run · 01** — English prompt verbatim, app-reported 4m1s, the report header (ticker / scan time / trigger) and the first two signal rows.

Run · 02 — the full 8-signal table: 6/8 hits → 🔴 HIGH RISK, trap_score 4. — **Run · 02** — the full 8-signal table: 6/8 hits → 🔴 HIGH RISK, trap_score 4.

Run · 03 — why this is not a textbook case (signals missed), the money-side and content-side read, and where t — **Run · 03** — why this is not a textbook case (signals missed), the money-side and content-side read, and where the 6/8 count inflates.

Run · 04 — the JSON body: signals_hit, breakdown, boost_detail, recommendation, warning_phrases. — **Run · 04** — the JSON body: signals_hit, breakdown, boost_detail, recommendation, warning_phrases.

Run · 05 — the 14 sources in full, the completion check, and the first honest boundary. — **Run · 05** — the 14 sources in full, the completion check, and the first honest boundary.

Run · 06 — all six honest boundaries and the archiving note. — **Run · 06** — all six honest boundaries and the archiving note.

Honesty boundary

Two things this run cannot claim. First: the report is in English, but every piece of evidence behind it is Chinese — an eight-signal scan of an A-share small cap run purely in English returns close to zero. Second: signals 1, 2, 6 and 7 are four views of one fact — people are posting about the stock — counted as four separate hits. 6/8 overstates how organised the promotion is and understates the capital risk.

CASE 02 · CORE VALUE · 6 / 8 · 🔴 HIGH RISK · trap_score 4 · RUN 2026-09-15 19:41

Case 03 · Historical replay Rendong Holdings · 002647 Scanned as at 2020 Retrieved 2026-09-15 19:52

A stock the CSRC later called manipulation. The scan said 🟡.

Run against 2020 material only, with 2021-and-later sources excluded. The operator controlled 83 accounts, drove the price up 380% and lost ¥2.69bn doing it. The scan returned 3 / 8 — 🟡 CAUTION. This is the case that defines the boundary of what the tool can see.

Result

3 / 8 · 🟡 CAUTION

trap_score 7 · wording boost 0

Hits on fundamentals, price action and one persona post. Five of the six misses ask the same question: is anyone promoting this? The answer was no — the operator had 83 accounts and no need for an audience. He raised the price and waited for retail to follow the chart in.

01 Input

Verbatim prompt, with the time constraint stated explicitly.

Prompt @skill:trap-detector Using publicly available material from 2020 only, check Rendong Holdings (002647) for pump-and-dump risk as it stood in 2020. Ignore anything published after 2021 — I want the 2020 picture, not today's. Scan all eight signals against that period: hit, miss, or insufficient for each, with sources. Then give me the risk level and recommendation you would have given in 2020. Finally, in a separate paragraph, tell me what the same scan would return today, and why the two differ.

Rendong Holdings · 002647

2019 close ¥16.72 → 2020-11-20 high ¥64.72. Nine-month 2020 revenue ¥1.754bn (+89.77%) with net profit −¥21.92m (−144.50%); ¥2.142bn of current liabilities due within three months against ¥51m of unrestricted cash.

as at 2020post-2021 excludedboost 05 retrievals

02 Scan · three hits, five misses

Signal	State	Basis as at 2020
1 · Low-quality accounts in unison	MISS	Zero dragon-tiger appearances in the run-up; daily turnover pinned near 1%. Holder count still doubled mid-year (6,638 → 13,090) — retail arrived on their own.
2 · Templated pitch language	MISS	Forum vocabulary ("operator-run stock" / "main force" / "pig-butchering") clusters after the 11-25 crash — spectators, not promoters.
3 · Paid VIP group funnel	MISS	No paid group for this ticker. Adjacent, different: an illegal margin-financing ring involving 002647 was detained on 2020-12-10, after the collapse.
4 · Fundamentals detached	HIT	Three consecutive loss-making quarters; ¥2.142bn due within three months vs ¥51m unrestricted cash; a ¥350m loan in default (filed 2020-10-29). Price +278.73% year to date.
5 · Abnormal price action	HIT	CSRC later measured a 380.48% rise and 400.77% amplitude over the manipulation window. Turnover locked near 1%, minimal intraday movement, uncorrelated to the index — textbook control.
6 · "Teacher" persona	HIT	A 2020-12-11 open letter urging holders to defend their positions, with replies alleging payment. Weak and post-crash — the Chinese run on the same ticker did not surface it.
7 · Cross-platform push	MISS	No coordination during the advance. The nine limit-ups in eight sessions came in 2021-03, outside the window.
8 · Fake research / news	MISS	The core narrative was real (state-controlled entry 2019-07-29, exit 2020-11-18). One genuinely misleading disclosure was flagged by the Zhejiang regulator — logged as a strong edge observation.

03 Output · then and now

TRAP SCAN REPORT · 2020 AS AT 2020002647

signals_hit	3 / 8
user_keyword_boost	0 — neutral phrasing, no trigger words
trap_score	7 / 10
risk_level	🟡 CAUTION
recommendation as at 2020	Some promotion signals present. Verify the source of the information.

What the same scan returns today

Run against current material, the scan is cleaner still — the promotion layer never existed, and by now it has also receded. The divergence is the point: the scan measures a moment, not a company. A 🟡 issued in 2020 and a 🟢 issued today describe the same event from two distances, and neither one is the CSRC's finding of manipulation.

04 Evidence

3/8

signals hit

accounts (CSRC)

380.48%

rise, manipulation window

13,090

holders at Q3 2020

−¥2.69bn

operator's realised loss

screenshots archived

Run record · case-materials/trap-detector/03-differentiation-en/run-log.md · screenshots shot-c-en-chat-01…09.png · 5 retrievals, 2 English-keyword (both returned Chinese sources)

Screenshots captured from the live scan — click any frame to open it at full size.

Run · 01 — English prompt verbatim, app-reported 8m57s, the report header (scan mode / scan time / trigger) an — **Run · 01** — English prompt verbatim, app-reported 8m57s, the report header (scan mode / scan time / trigger) and the keyword-boost note.

Run · 02 — the 8-signal three-state table, rows 1–5. — **Run · 02** — the 8-signal three-state table, rows 1–5.

Run · 03 — signal rows 6–8, 3/8 → 🟡 CAUTION, trap_score 7, the 2020-era recommendation, and the divergence not — **Run · 03** — signal rows 6–8, 3/8 → 🟡 CAUTION, trap_score 7, the 2020-era recommendation, and the divergence note.

Run · 04 — the Chinese/English comparison table (2/8 vs 3/8 · 🟡 vs 🟡 · 8 vs 7) and the “not linguistic” argume — **Run · 04** — the Chinese/English comparison table (2/8 vs 3/8 · 🟡 vs 🟡 · 8 vs 7) and the “not linguistic” argument.

Run · 05 — “the miss is structural”: the regulator’s findings (83 accounts, 218 trading days, 91.63% of trades — **Run · 05** — “the miss is structural”: the regulator’s findings (83 accounts, 218 trading days, 91.63% of trades suppressed, a ¥2.69bn loss) against a 🟡 rating.

Run · 06 — the 2020 retail warning that the eight signals never look for, and the first half of the JSON outpu — **Run · 06** — the 2020 retail warning that the eight signals never look for, and the first half of the JSON output.

Run · 07 — the JSON tail, the completion check, and “what the same scan returns today”. — **Run · 07** — the JSON tail, the completion check, and “what the same scan returns today”.

Run · 08 — the 2020 vs 2026 table and why an adjudicated manipulation scores better today. — **Run · 08** — the 2020 vs 2026 table and why an adjudicated manipulation scores better today.

Run · 09 — all six honest boundaries and the archiving path. — **Run · 09** — all six honest boundaries and the archiving path.

Honesty boundary · two

One: the Chinese run on the same ticker and the same date returned 2 / 8, this one returned 3 / 8 — signal 6 was retrieved by one and not the other. Retrieval is not deterministic; the hit count carries a run-to-run variance the framework does not account for. Two: both runs graded a CSRC-confirmed manipulation as 🟡 CAUTION. Six of the eight signals ask whether someone is selling you the stock. Capital-driven manipulation does not need to sell.

CASE 03 · HISTORICAL REPLAY · 3 / 8 · 🟡 CAUTION · trap_score 7 · SCANNED AS AT 2020 · RETRIEVED 2026-09-15 19:52

Case structure · fixed five-part record

Each run is recorded in the same shape, so results can be compared rather than admired:

01 Inputthe verbatim prompt, the ticker card, the observed market data

02 Scaneight signals, each with its hit state and the source behind it

03 Outputthe JSON report, the assigned level, the safe next step

04 Evidenceretrieved figures, sources, and an honesty boundary for what was not checked

05 Verdictwhat the run got right, and where the framework was weak

06 · Security · Compatibility · Reliability

The scanner itself has no attack surface.

What follows is what the package actually contains — including the parts that limit it.

Security audit · 7 checks

✓0 lines of executable codepure Markdown instructions
✓No scripts, hooks, or install actionsnothing runs on setup
✓No dependencies to installno pip / npm / requirements
✓No outbound network callssearch is performed by the host agent
✓No file, credential, or key accessreads nothing from the machine
✓2 files totalSKILL.md · references/eight-signals.md
✓MIT · v3.9.4source readable end to end

Compatibility · 5 dependencies

✓Host agent with live searchhard requirement — without it, 6 of 8 signals fail
✓Chinese-language sourcesA-share verified; other markets untested
✓Companion skill for signals 4–5not installed here → coverage 6 / 8
✓JSON consumer optionalpairs with deep-analysis downstream
✓Trigger by keyword or explicit requestno parameters, no configuration

Known limitations · 6

Equal weighting. "Paid VIP funnel" and "templated phrasing" each count as one hit, though the first is far more serious. No weights, no confidence scores.
Promotion-biased by construction. Six of eight signals inspect the sales operation. Capital-driven or accounting-fraud manipulation produces little for them to find.
The data-source knot. Signal 4 judges fundamentals from published financials — and the point of a well-built fraud is to make those financials look good. The better the packaging, the weaker the signal.
Result quality tracks the host model's retrieval. The same instruction run on a model that cannot search produces eight "insufficient" rows.
No reference corpus. There is no blacklist or historical case library to compare against, so every run starts from scratch.
Results are momentary. The scan measures whether promotion is happening now. A clean result expires; a re-run is required, not optional.

07 · Who it's for

For the moment before you start researching it.

The scan is a gate, placed upstream of analysis — not a substitute for it.

Retail investors

A-share holders whose information arrives through groups and short video.

Run it on any recommendation before spending an hour on the fundamentals.

New investors

People still calibrating what a legitimate pitch sounds like.

Paste the promoter's message unchanged — let the weighting judge the source.

Checking for someone else

Adult children verifying what a parent was told in a group.

The people these campaigns target and the people who run this tool are usually not the same person.

Advisers

Anyone who needs a first-pass risk note before a human conversation.

Use the JSON as triage material, then apply judgment on top.

Scan the next tip before you look into it.

Quote the message exactly. Paraphrasing it into polite language destroys the strongest signal in the whole framework.

"a friend recommended this to me" "the group says…" "the teacher is leading us into…" "is this a pump and dump?"

Copy-ready prompt Check [ticker or name] for pump-and-dump risk. Here is what I was told, quoted word for word, not rewritten: "[paste the exact message]". Scan all eight signals and give me a risk level, the evidence behind each hit, and a recommendation.