Sample run · public filing · Anthropic S-1

Filing in, checked first draft out.

One real run of our pipeline on a public filing. Every number in the draft points to a row in the fact sheet, every row points to a page, and a separate checker reads the draft before anyone else does. An analyst still edits and signs off.

Questions or a 30-minute call: hello@thelevelheadedai.com

How a draft is made

  1. 1 · SourceThe filing PDF. Nothing else counts as a source.
  2. 2 · Fact sheetEvery claim and number extracted with its page.
  3. 3 · DraftWritten only from the fact sheet, each number tagged.
  4. 4 · CheckerA separate model re-reads the draft against the pages.
  5. 5 · GateNo pass on record for this exact file, no release.
  6. 6 · AnalystEdits, judges, signs off.

The checked first draft

Select a tag to see the fact-sheet row behind it.

Anthropic, PBC · S-1 · H1 2026checker: pass

Revenue rose from $1,163M in H1 2025 to $16,286M in H1 2026, up 1,300% . The H1 2026 net loss of $(80,748)M is mostly a $(72,388)M change in the fair value of convertible notes, derivative liabilities and SAFEs, which the filing attributes to a higher enterprise valuation . That charge is non-cash: the cash-flow statement adds it back, and operations generated $3,929M of cash .

On a GAAP basis the operating loss was $(693)M, a −4% margin ; adjusted operating income, a non-GAAP measure, was $1,618M, a 10% margin . Compute and infrastructure fell from 160% of revenue in FY2025 to 55% in H1 2026 . After the offering, a Long-Term Benefit Trust elects four of the seven directors .

The fact sheet behind it

RowItemValuePage
F1bRevenue, H1 2025 → H1 2026$1,163M → $16,286M (+1,300%)22, 148
F2Net loss, H1 2026$(80,748)M22
F3Change in fair value of convertible notes, derivative liabilities and SAFEs, H1 2026$(72,388)M22, 150
F4Operating loss, H1 2026 (GAAP)$(693)M, −4%22, 24
F5Adjusted operating income, H1 2026 (non-GAAP)$1,618M, 10%24
F6Compute & infrastructure as % of revenue, FY2025 → H1 2026160% → 55%148
F15bBoard after the offering: 7 directors; the Long-Term Benefit Trust (Class T) elects 47 directors; Trust elects 4 (Class T)256
F25Cash flow, H1 2026: fair-value change added back as non-cash; net cash from operations$72,388M; $3,929MF-12

What the checker caught

Real findings from our own runs. The checker is a separate reviewer: the maker never checks its own work.

S1 · dropped qualifierrecap line
Draft"Close to breaking even"
Required fixSay "operating", and show the net loss of $(80,748)M next to it

The draft dropped the word "operating". With a net loss of $80.7B, the unqualified line was false.

S1 · claim not in the sourceopening
Draft"One of the fastest-growing companies on record"
Required fixCut it. The filing (p.5) speaks of the fastest enterprise adoption, a different claim

No fact-sheet row supported it, so it could not stay.

S1 · material omissionheadline
Draft"The founders hold the vote"
Required fixAdd that the trust elects 4 of 7 directors (p.256)

One version of this went out before the check existed. That is why every release now needs a recorded pass.

The release gate

A pass is recorded against a fingerprint of the exact file. Change one character and the fingerprint changes, so the file needs a new check.

What a 2-week build gives your team

One note type, your format

We map one recurring note (results update, initiation, monthly) and build the pipeline to your template.

Your sources, your rules

Filings and data you already use. House style and disclosure rules become checks.

Analysts stay in charge

The output is a checked first draft. Judgment, rating and sign-off stay with your analyst.