mcptest · live field report SEC FORM 4 · 10.2M LEGS · MID-2003–PRESENT

The ledger a language model can't read

Three insider-trading signals pulled live from the ARGOS sec vertical's insider_ toolset, as of 17 July 2026. Every figure below is real, current, and reproducible — and every one of them requires a materialized database, not a language model's memory, to produce.

Ask any frontier model “which company insiders are quietly buying right now” and it will confidently invent a plausible-sounding answer — a ticker, a name, a dollar figure, none of it grounded in anything filed with the SEC this week. The gap isn't reasoning, it's freshness and arithmetic: nobody's training data contains a Form 4 filed two days ago, and nobody memorized 184,000 individual insiders' five-year buying medians.

The four things below are precisely what a deterministic, materialized pipeline supplies that an LLM cannot conjure on its own — and the three case files that follow are what falls out when you actually query it.

01

Filed this week

Form 4s land within two business days of a trade. No model's training cutoff is that recent, and retraining doesn't happen at filing speed.

02

A baseline per person

“79× their own median” requires five years of history, materialized separately for each of ~184K individual insiders.

03

Signal cut from noise

10.2M classified transaction legs separate opportunistic buying from option exercises, tax withholding, and scheduled 10b5‑1 plans.

04

Deterministic, not vibes

Same inputs, same signal_class, every time — no hallucinated catalyst, no invented ticker.

BREADTH · TSM · Taiwan Semiconductor

Thirty-one insiders bought TSM. Seven of them run the company.

net_selling

The headline number is misleading on purpose — that's the point of showing it. TSM's 90-day net_discretionary_usd reads negative because one vice president's single $14.0M sale outweighs all the buying combined. But that same window shows 31 distinct insiders independently buying in the open market, including the CEO and three separate chief operating officers — a breadth of conviction a raw dollar total actively hides. A concentrated block of that buying, mostly C-suite, landed in a single batch on July 9, with several executives buying at 50–79× the size of their own historical median purchase.

Distinct buyers
31
C‑suite buyers
7
Unusual vs. own history
26 / 31
Buy volume
$756,971
Window
90d
APR 19 JUN 3 JUL 17 JUL 9 · C‑suite batch
● filled = buy   ○ outline = sell    ring = concentrated multi-insider purchase
notable[] · top buyers by value, insider_signal(TSM, 90d)
InsiderRoleBuy dateBuy $× own median
Tien Bor-ZenVP2026-07-09$245,83679.0×
Yuan LipenVP2026-07-09$163,86751.6×
Wei Che-ChiaCEOChief Executive Officer2026-07-09$34,0793.0×
Chin Yung-PeiC‑SUITEChief Operating Officer2026-07-09$16,0703.0×
Mii Yuh-JierC‑SUITEChief Operating Officer2026-07-09$16,0693.0×
Hou Yung-ChinC‑SUITEChief Operating Officer2026-07-09$14,1273.0×
Chuang Tzu-SouVP · also sold $14.0M2026-07-09$10,3923.0×

Reading the top line alone (“TSM insiders are net sellers”) gets the story backwards. Reading the structured fields together — cluster size, C-suite conviction flag, and per-person baselines — is what surfaces the real signal underneath.

SYNCHRONY · QNT

Ten insiders with no buying history. One day. $24.7M.

cluster_buy

Of the 11 insiders who bought QNT stock in the last 90 days, ten had made zero open-market purchases in the preceding five years — flagged individually by first_buy_in_5y on each person's own materialized history, not inferred. All eleven purchases landed on the same calendar day, June 8, totaling $24,661,920 with no sellers at all. That's not one insider's decision reflected once — it's synchronized first-time conviction across an entire board and executive bench.

Distinct buyers
11
First buy in 5y
10 / 11
C‑suite buyers
2
Buy volume
$24,661,920
Sellers
0
APR 19 JUN 3 JUL 17 JUN 8 · all 11 buyers
every dot in this story lands on the same x‑coordinate
notable[] · insider_signal(QNT, 90d)
InsiderFirst buy in 5yBuy $
Barron Haltrue$15,000,000
Jimenez Josephtrue$3,999,900
Bhatia Manish Htrue$1,200,000
Dehoff Kevin ScottC‑SUITEtrue$900,000
Denman Kenneth Dtrue$840,000
Sharan NiteshCFOfalse · 1.6×$390,000

The one buyer who did have history, CFO Nitesh Sharan, still only bought 1.6× his own median — a useful negative case showing the tool doesn't inflate everyone into “unusual.”

SCALE · RIVN · Rivian Automotive

One insider bought a billion dollars of Rivian stock.

net_buying

Volkswagen AG — a strategic investor holding an insider-reportable stake in Rivian — filed 28 separate Form 4s over the window, all discretionary open-market purchases, totaling $999,943,400. That single figure is what pushed the entire “Motor Vehicles & Passenger Car Bodies” SIC sector to +$999M net insider buying in insider_sector_pulse — a sector-level number that would be meaningless without knowing it traces to one filer. Because only two distinct people bought (VW and one other insider, $251,460), the tool correctly withholds the cluster_buy label: this is scale, not breadth, and the classification doesn't conflate the two.

Distinct buyers
2
Filings
28
Is cluster?
false
Buy volume
$1,000,194,860
Sellers
0
APR 19 JUN 3 JUL 17 MAY 4 · VW · $999.9M MAY 19 · $251K
dot area scaled to dollar size — the second buyer is on the chart only because it's drawn at a minimum radius
notable[] · insider_signal(RIVN, 90d)
InsiderLast buyBuy $
Volkswagen AG2026-05-04$999,943,400
Gomez Aidan N.2026-05-19$251,460

signal_class is a deterministic classification of filed activity, not a buy or sell recommendation. Figures aggregate the price-sanitized value_usd field and exclude planned 10b5‑1 sales from discretionary totals, per each tool's documented contract. Compiled by mcptest against the ARGOS MCP sec vertical.