Diligence mark

Multi-agent due diligence

Bull vs bear, side by side.

Diligence reads 10-K, 10-Q, and the earnings call. Five adversarial agents debate the evidence. You see the disputed facts, ranked by materiality.

Loading…

How it works →

Coverage on day one

NVDA
TSLA
PLTR
AAPL
MSFT
META
AMD
AMZN
GOOG
NVDA
TSLA
PLTR
AAPL
MSFT
META
AMD
AMZN
GOOG

The adversarial frame

Bull charges. Bear counters.

Two agents read the same evidence and argue the opposite case. The reconciler ranks where they disagree by materiality.

Bear case · short side

  • Hedging in Q&A14 flags
  • Concentration risk3 customers
  • Margin trajectory−320 bps

Bull case · long side

  • Compute demand+52% YoY
  • Pricing power+18% ASP
  • Operating leverage+410 bps

Reconciler ranks every disputed fact by materiality · 1–10 scale

How it works

One ticker in.
Cited disagreements out.

  1. 01

    Cache the sources

    10-K + 10-Q from SEC EDGAR. Earnings call audio via yt-dlp. Speechmatics transcribes with speaker diarization. Everything saved locally so the agents read the same evidence on every run.

  2. 02

    Five agents debate

    Filing analyst and call analyst extract atomic claims with citations. Bull and Bear build opposing investment cases on top, citing only those claim IDs. The reconciler diffs them.

  3. 03

    See what's disputed

    Every disagreement is ranked 1–10 by materiality. Each side's argument links back to the filing section or transcript timestamp. No invented facts, no unsourced claims.

One row · NVDA · sample output

What a disputed fact actually looks like.

Every row in the dashboard is one place the bull case and the bear case disagree. Both sides cite claim IDs the analysts extracted. Nothing is invented. The reconciler scores how much the disagreement moves the answer.

Disputed fact

One specific topic where the bull case and the bear case argue opposite conclusions from the same primary sources.

Claim IDs (F-014, C-031, …)

F-### = filing claim from the 10-K/10-Q. C-### = call claim from the earnings-call transcript. Every position must cite real IDs; agents cannot invent facts.

Materiality score 1–10

How much this single disagreement moves the final answer. 1–3 cosmetic. 4–6 worth noting. 7–10 decision-altering. Higher ≠ buy or sell — just ‘matters more’.

Topic

Q4 FY26 gross-margin trajectory

Materiality

8/10

Bull position

ASP expansion on Blackwell + Hopper mix is sustainable through FY27. Pricing power compounds the operating-leverage tailwind into ~76% non-GAAP gross margin.

F-014C-031

Bear position

Gross margin guided down sequentially in Q4 prepared remarks; CFO repeated ‘mix and yield’ language three times in Q&A — historically a leading indicator of inventory write-downs.

F-017C-042C-056

Reconciler · materiality rationale

Each 100 bps of gross margin = ~$0.40 of FY26 EPS. The bull and bear paths differ by ~400 bps over the next two quarters. Single-largest swing factor in the model.

On the full dashboard, you'd also see

  • All 15–20 disputed facts ranked
  • Bull pillars · Disputed table · Bear pillars (3-column)
  • Materiality bar chart — click any bar to scroll to the row
  • Diarised call audio with click-to-jump transcript
  • Click any claim ID to land in the source filing section

Research tool, not investment advice. Diligence surfaces contested facts so an analyst can read the evidence faster. It does not say buy, sell, hold, long, or short. Every claim links back to the source filing section or transcript timestamp; the human decides.

What you get back

Materiality, charted.

Every disputed fact lands on a 1–10 materiality scale. The dashboard chart uses the same emerald-and-magenta gradient as the surfaces below — green for the focused disagreement, purple for the highest-materiality rows.