The Sources Log Says What Was Checked, When, and What Could Not Be

Sources and verification

Facts that change belong on this page and not in print. Each entry was checked on 2026-10-07. A check does not mean everything was confirmed: the third column says what that check could not confirm, and each item there is still open.

Source	What the book states from it	What could not be verified
Spence (1973), Akerlof (1970), Nelson (1970), Darby and Karni (1973)	The economics of costly signals, hidden quality, search and experience goods, and credence goods, each stated as the paper’s central idea.	The papers’ full texts were not read in the check; citation details were confirmed through a bibliographic database, and the end pages of Spence and Akerlof are not confirmed.
METR (2025, 2026)	A 2025 randomized trial of experienced developers, and METR’s later statement that selection effects made follow-up measurement unreliable.	The later post’s sign conventions were read through a summary.
AICPA	What a SOC 2 report is; that a type 1 report covers the design of controls as of a specific date, and a type 2 report also covers whether they operated effectively.	Confirmed on AICPA pages on 2026-10-07, which write “type 1” and “type 2.” The pages read do not state how long a period a type 2 report covers, so the book states no length.
FDA, 510(k) clearance	Clearance rests on substantial equivalence and is not approval.	Nothing beyond the point that clearance does not by itself show benefit in a given setting, which is the book’s inference.
10 U.S.C. 4022(f)	The follow-on production provision for competitively awarded prototype transactions.	Verify the current text with your contracting officer and counsel.
FedRAMP, GSA	A standardized security assessment and authorization program for cloud services.	The program’s own site did not carry the descriptive text; GSA’s page was used.
HHS, business associate agreements	A covered entity needs a written agreement with a vendor that handles protected health information for it.	The HHS page was read through search text.
ONC Health IT Certification Program	Certification against adopted standards and criteria.	The program publishes no statement about outcomes; the book’s point is an inference from its scope.
Zhang et al. (2024)	A fresh benchmark showed some models scoring up to 8 percent lower than on a public one, and the leading models showed little drop.	Whether the 8 percent figure is points or relative.
Dell’Acqua et al. (2026)	Gains inside a model’s capability and a loss on a task outside it.	The journal version’s figures were not compared with the working paper’s.
Source code escrow	A neutral third party holds code and releases it on defined triggers.	No government or standards source for the software case was found.

Source

What the book states from it

What could not be verified

Spence (1973), Akerlof (1970), Nelson (1970), Darby and Karni (1973)

The economics of costly signals, hidden quality, search and experience goods, and credence goods, each stated as the paper’s central idea.

The papers’ full texts were not read in the check; citation details were confirmed through a bibliographic database, and the end pages of Spence and Akerlof are not confirmed.

METR (2025, 2026)

A 2025 randomized trial of experienced developers, and METR’s later statement that selection effects made follow-up measurement unreliable.

The later post’s sign conventions were read through a summary.

AICPA

What a SOC 2 report is; that a type 1 report covers the design of controls as of a specific date, and a type 2 report also covers whether they operated effectively.

Confirmed on AICPA pages on 2026-10-07, which write “type 1” and “type 2.” The pages read do not state how long a period a type 2 report covers, so the book states no length.

FDA, 510(k) clearance

Clearance rests on substantial equivalence and is not approval.

Nothing beyond the point that clearance does not by itself show benefit in a given setting, which is the book’s inference.

10 U.S.C. 4022(f)

The follow-on production provision for competitively awarded prototype transactions.

Verify the current text with your contracting officer and counsel.

FedRAMP, GSA

A standardized security assessment and authorization program for cloud services.

The program’s own site did not carry the descriptive text; GSA’s page was used.

HHS, business associate agreements

A covered entity needs a written agreement with a vendor that handles protected health information for it.

The HHS page was read through search text.

ONC Health IT Certification Program

Certification against adopted standards and criteria.

The program publishes no statement about outcomes; the book’s point is an inference from its scope.

Zhang et al. (2024)

A fresh benchmark showed some models scoring up to 8 percent lower than on a public one, and the leading models showed little drop.

Whether the 8 percent figure is points or relative.

Dell’Acqua et al. (2026)

Gains inside a model’s capability and a loss on a task outside it.

The journal version’s figures were not compared with the working paper’s.

Source code escrow

A neutral third party holds code and releases it on defined triggers.

No government or standards source for the software case was found.