TSM
8-K
$29.4B capital expenditure commitment
Why it matters: Capex at this scale signals multi-year capacity bets. It moves supply expectations across the semiconductor chain.
🔺 DELTASIGNAL
SEC Filings Intelligence
DeltaSignal reads SEC filings for the moves companies do not advertise. Every filing below was flagged as material by two independent graders working blind. When they disagree, a third adjudicator decides. This graded record trains the DeltaSignal Decider.
The DeltaSignal Decider · Training corpus · October 2026
Signals flagged
18
Signal rate
9.9%
Independent graders
2
Grader agreement · targeted batch
88.9%
Cohen's kappa · targeted batch
0.77
Signals by form type
Where the signals come from
- 8-K Current reports 7
- Form 4 Insider trades 5
- Proxy actions 4
- Form 144 Planned sales 1
- Form 425 Combinations 1
Signals by category
What kind of event
Corporate actions6
Insider transactions5
Capital and balance sheet4
Leadership and workforce3
Most signals cluster in corporate actions and insider flow. Routine filings, the other 90 percent, stay quiet.
How the corpus was built
Labeling phases, October 2026
1
Original corpus
47 filings
Hand-labeled single-filing verdicts from the desk-note record: 16 SIGNAL, 31 NO SIGNAL. The seed set.
2
Expanded grading
135 newly graded
Two graders, blind to each other. Raw agreement on the filings both graders share: 95.5 percent.
3
Combined signal set
18 unique signals
Union of both graders. A 9.9 percent signal rate against a 90 percent routine baseline.
4
Agreement test
27 targeted filings
Event-driven filings (officer changes, material agreements, debt actions), graded blind by both graders. 88.9 percent agreement, Cohen's kappa 0.77 (substantial agreement), against a 0.75 bar set before grading began.
Flagged signals
Each card is a filing our graders flagged as material. Cards marked under review had grader disagreement and went to adjudication.
HUT
Form 4
$58.3M insider monetization
Why it matters: Insider sales of this size can signal management's view on valuation. Size relative to float is the read.
HUT
Form 4
$58.3M insider monetization, second filing
Why it matters: A second large monetization in the same window strengthens the signal. Pattern over single event.
MSTR
Proxy
Daily dividend proxy actions
Why it matters: Proxy actions around distributions reveal capital structure decisions. Repeated filings suggest an ongoing program. Four related filings flagged.
NVDA
Form 4
$300M discretionary director sale
Why it matters: Nine-figure discretionary sale by a director. The market reads insider conviction from size and timing.
CIEN
Form 4
$1.34M discretionary sale
Why it matters: Smaller but discretionary. Direction of insider flow at an inflection point carries weight beyond the dollar figure.
AMZN
8-K
Bond tranche issuance
Why it matters: Debt issuance terms reveal how cheaply a mega-cap can fund. Tranche structure signals duration appetite.
NET
144
$8.8M planned sale size step-up
Why it matters: Step-up in planned sale size indicates accelerating insider liquidity. Watch the Form 4s that follow.
CVX
8-K
C-suite reshuffle
Why it matters: Leadership changes at a supermajor reset strategy expectations. The market reprices governance risk on the announcement.
CVX
8-K
Hess Midstream divestiture
Why it matters: Asset sale sharpens portfolio focus. Proceeds use, buyback versus capex, drives the read.
HUBS
8-K
7% workforce reduction
Why it matters: Cuts of this size signal margin defense. The open question is whether growth reaccelerates after.
HUBS
8-K
Chief legal officer resignation
Why it matters: A chief legal departure can precede regulatory or litigation overhang. Context determines severity.
BMNR
8-K
$17.4B ETH holdings disclosure
Why it matters: Balance sheet exposure to ETH at this scale turns the equity into a crypto proxy. The disclosure moves the multiple.
AAPL
Form 4
CEO share sale
Why it matters: CEO transactions draw outsized attention. Even routine sales move sentiment at this scale.
DJT
425
Under review
Business combination communication
Why it matters: Form 425 appears only around business combinations. One grader flagged it, one did not. Currently in adjudication.
How signals are flagged
Two graders, blind. Each filing is reviewed independently by two graders who cannot see each other's verdicts or any prior labels.
Disagreements adjudicated. When graders split, a third reviewer decides with grader identities hidden.
Agreement measured. On the 133 filings both graders share, raw agreement is 95.5 percent. On the targeted batch of 27, agreement is 88.9 percent and Cohen's kappa is 0.77, against a 0.75 bar set before grading. The corpus mixes a routine filing stream, where flags are rare, with a targeted event-driven slice where material disclosures concentrate. Agreement was measured on the targeted slice: 15 of 27 flagged by both graders, 9 cleared by both, 3 split and sent to adjudication.
Outcome tracking in progress. These are flagged signals, not graded predictions. Price outcomes will be added as the record builds.
Record updates
Oct 15, 2026. Remaining corpus graded. The 179 filings of the main corpus were graded blind by both graders under the same protocol as the targeted batch: 88.0 percent agreement on decided filings, Cohen's kappa 0.57, and 25 splits and abstentions ruled by blind adjudicators. The corpus stands at 206 filings with final labels, 44 SIGNAL and 162 NO SIGNAL, including the 27 targeted filings. The 200 figure used earlier in the program was a planning frame, not a count; 206 is the count.
Oct 14, 2026. Correction and update. The original 47 filings are hand-labeled single-filing verdicts, 16 SIGNAL and 31 NO SIGNAL; the phase list's earlier wording belonged to a different corpus and has been corrected. The Riot Platforms re-extraction from EDGAR was completed Oct 13.
Oct 12, 2026. Adjudication. The three filings the two graders split on were ruled by a third judge, blind to label origin: Mercury Systems NO SIGNAL, Riot Platforms SIGNAL, Marathon Digital NO SIGNAL. The batch stands at 16 SIGNAL and 11 NO SIGNAL with no unresolved splits. The Riot ruling was made from the filing at EDGAR after our captured copy proved truncated; the corpus copy was re-extracted in full from EDGAR on Oct 13. The three rulings set the grading precedents the remaining 173 filings are graded under. (the post)
Oct 12, 2026. Named. The model this corpus trains is the DeltaSignal Decider. When it goes live, its filing calls will run for Alpha Circle, each one graded in public. (the post)
Oct 12, 2026. Agreement figures labeled by set. The 88.9 percent agreement and Cohen's kappa 0.77 measure the targeted batch of 27 filings. The 95.5 percent figure is raw agreement on the 133 filings both graders share. The earlier presentation did not distinguish the two sets. (the correction post)
Oct 12, 2026. Agreement result published. The two graders agreed on 24 of 27 targeted filings, kappa 0.77 against the 0.75 bar set before grading. The 3 split filings go to a blind adjudicator. (the post)
Oct 11, 2026. Corpus locked for blind grading, Step 1 of the Decider program. (the post)
Chapters
The graded record, in order
Chapter 1: The Seed 47 hand-graded filings, 16 SIGNAL / 31 NO SIGNAL.
Chapter 2: The Targeted 27 27 event-driven filings, 16 SIGNAL / 11 NO SIGNAL.
Chapter 3: The Locked Corpus 179 blind-graded filings, morning lock October 16: 25 SIGNAL / 152 NO SIGNAL / 2 ABSTAIN; relock same day: 19 SIGNAL / 157 NO SIGNAL / 3 ABSTAIN.