The White Drum Challenge

The longest book ever written by a human

The longest novel in the world is called Venmurasu — Tamil for “the white drum.” 26 volumes, roughly 22,400 pages. One person, six and a half years.

It is a monument to human endurance — to still remember the characters of volume 1 while writing volume 20 is itself a miracle.

Now AI takes the same test — with extra difficulty

We are adding a difficulty a human author never faces:

“The world state of every chapter must be verifiable.”

At one million characters, the foreshadowing planted in chapter 3 must still be paid off, the injury from chapter 9 must still hurt, and who knew what, and when — every bit of it checkable. Not because the model has a good memory. Because a gate keeps the books.

A human writing a long novel has exactly one ceiling: memory.
AI writing a long novel used to have no ceiling at all — it falls apart at 50,000 characters.

The method: state outside the model's head

We move world state out of the model's context and into persisted objects. Every change is one semantic transaction, passed through five gates; nothing is written unless all five pass; on failure, zero writes and the reason is returned for rewriting.

1. Text non-empty — a chapter with no prose is not a chapter.
2. Structural references — every entity referenced exists.
3. Forbidden zones — constraints that must not be violated.
4. Foreshadowing state machine — planted threads tracked, paid off or still open.
5. Prose-vs-state check — what the text says agrees with what the state says.

This is not “generate a million characters” — a generator can vomit a million characters in a day and fall apart at the 50-thousandth. This is AI writing a million characters without falling apart, the Everest of AI writing.

The honest baseline

Against a self-built benchmark (ShadowBench-W), the method's state write-back accuracy:

	2Origin state layer	Bare model	Vector RAG
State accuracy (33-run series)	98.9%	75.0%	75.0%

⚠️ This is the instrument of our own benchmark — not a third-party evaluation. The definitive comparison after the probe fix (Run #27) is 95.8% ± 5.9 vs 52.1% ± 19.7 / 58.3% ± 11.8, p = 0.0024. Nobody else runs a “state write-back correctness” benchmark — we defined the exam, and the standard-setting power is ours.

Where we stand now (numbers generated by scripts/challenge-stats.mjs, never hand-written)

drift-rate — the rate at which the model misremembers a prior value — was last measured at 8.3% (2026-08-06; run id not retained). It is re-measured at each milestone under the same protocol and published with a run id + date. Unnumbered bare numbers do not go on this page.

Cadence

Publicly marking pauses is worth more than pretending the serial never stopped.

• Loading…

The rules — the challenge must itself be verifiable

1. Each chapter = one semantic transaction (prose + state changes).
2. All five gates must pass before anything is written.
3. Violations write nothing; the reason is returned for rewriting.
4. Every chapter: prose + world state + git commit — all public.
5. Anyone can clone and verify: node adapters/story/cli.mjs state <pkg>

Follow the serials

Humans took 6.5 years. How long will AI take? Unknown. But every chapter is public — prose, world state, foreshadowing graph, timeline — all cloneable, all verifiable. git log is the evidence chain.