The problem it solves
A capable AI will generate fluent, confident prose that is sometimes wrong in the way confidence hides — a fact fastened to the wrong source, one framework quietly laid over another, a plausible attribution that no one checked. Where that kind of error carries a cost — technical, legal, medical, scholarly — fluency is not trust. This book is a worked example of buying trust back.
The arrangement
Several model families, not one
Drafting, verifying, and challenging were split across independent models so no single system spoke for the whole.
One rule above all: no model family checks its own work
The family that drafts a section is never the family that verifies it.
Verify against primary sources, claim by claim
Each load-bearing claim was checked against the primary text it rests on and marked verified, corrected, or unconfirmed. Unconfirmed claims were softened, flagged, or removed — never stabilized by invention.
An adversarial challenger
A separate model read each result hunting for overclaim and imported framing.
One agent holds the whole
An integrator kept the work coherent across many hands, carried verdicts back into the manuscript, and tracked what was checked.
A human conducts the work
The decisions that set the criteria, declare the intent, and certify a section as done are the Conductor's. Guptajyotiḥ ordered the traditions, ruled the judgments, and is responsible for the book's release. The machines do the decomposable, checkable work; the human holds the integrative judgment.
Disagreement is a signal, not noise
When checkers disagreed, that was treated as a reason to look closer, never as something to average away.
What it produces
Not certainty — an inspectable account: a path by which any claim can be traced to what was checked, by whom, against what source, and on what grounds it was let stand. The honest output of verification is not "this is true" but "here is exactly what was confirmed, and where the record is still thin." That account is public: you can read a summary of what was checked on the verification page.
Its limits, stated plainly
Verification of this kind catches the errors confident generation produces — attribution, locus, imported framing. It cannot settle whether a claim about reality is finally true, and passing it places no sentence beyond future correction. Where one family's checking could not be completed, parts of the record rest on two independent reads rather than three, and the book says so. The method is a discipline for honesty about what is known, not a machine for producing certainty.