What counts as a registered prediction?

A prediction is registered when it is published with a date while the subject is still unclaimed, states a confidence level, and links a method described fully enough to rerun plus a public repository holding every raw measurement. All three can be verified by any reader; a call that appears only after the reveal does not qualify.

Why publish missed predictions?

Because a prediction record that only shows hits is an advertisement, not a measurement. The evidence value of the ledger is the standing commitment to grade every entry under the same rules, in public, before the outcome is known. A missed call published next to its method and raw data still demonstrates the practice; a quietly deleted one destroys it.

How are the predictions made?

From black-box measurements of the subject itself — for the first entry, differential token counting through a public API's billing metadata: 95 probe strings compared against fourteen candidate tokenizer vocabularies, for effectively zero marginal cost. Each entry links its own method paper and raw records, so the basis for every call is inspectable and reproducible.

The Reveal Ledger — Registered AI Attribution Predictions

Q: How does an entry get graded?

Only an official claim grades an entry: a statement from the lab, a model card, or the serving platform renaming the model to a vendor slug. Press reports citing insiders are recorded as notes but do not grade. A claim matching the prediction confirms it; a contradicting claim marks it missed. Silence grades nothing — an unrevealed subject stays pending.

Attribution predictions registered before the answer is known, graded in public when reveals land.

Abstract

The Reveal Ledger is a public record of attribution predictions registered before the answer is known. Each entry names its subject — typically a stealth AI model whose operator is undisclosed — states a prediction with a confidence level, links the published method and raw data behind it, and carries a date.

When the subject is officially claimed, the entry is graded confirmed or missed, and a graded entry is never removed. This page is the ledger.

Why keep a ledger?

A correct call announced after the reveal is worth nothing, because anyone can make one. A prediction carries evidence weight only when it is registered — published with a date, a stated confidence, and a method someone else can rerun — while the answer is still unknown. The stakes are practical rather than sporting: anonymous endpoints attract regulated and proprietary work every day, and knowing who actually serves one is due diligence.

Grading is the other half. A record that only ever shows hits is marketing. Publishing the misses next to the hits, on the same page and under the same rules, is what separates a measurement practice from a highlight reel.

What qualifies as a registered prediction

Three requirements, each one checkable by anyone:

Registered before any claim. The prediction is published, with a date, while the subject's operator is undisclosed and no party has claimed it.
Method published. The entry cites a method described fully enough to rerun — for Entry 001, a white paper.
Raw data public. Every measurement behind the prediction lives in a public repository, written once and never edited.

The list is short deliberately: each requirement does the same job, moving the prediction from trusted to checkable, and a claim that cannot be checked from the page and its links has no place on the ledger.

Grading rules

An entry holds one of three states:

Pending — registered; no qualifying reveal yet.
Confirmed — the subject was officially claimed, and the claim matches the prediction.
Missed — the subject was officially claimed, and the claim contradicts the prediction.

An entry is graded by an official claim and nothing else — a lab statement, a model card, or the serving platform renaming the model to a vendor slug. Credible press reports citing insiders are recorded as notes but do not grade — for earlier stealth models, reporting has preceded the official claim by weeks. Silence never grades an entry either; a prediction with no reveal stays pending, not vindicated.

The ledger

#	Registered	Subject	Prediction	Confidence	Method	Status
001	2026-08-22	`stealth/ox-alpha` (OpenRouter)	"Ox Alpha is a next-generation GLM model trained by Zhipu AI (Z.ai)."	High	Tokenizer fingerprint · raw data	Pending — unclaimed as of 2026-08-23

Entry 001 — Ox Alpha — rests on reported token counts matched against fourteen candidate vocabularies over 95 probe strings: GLM's vocabulary predicted 84 of 95 counts exactly, while the best non-GLM candidate managed 46. A dated follow-up against Zhipu's released GLM-5 vocabulary then scored 95 of 95, upgrading the match from lineage to identity.

The battery ran through the public API's billing metadata alone, at a marginal cost of roughly zero — and it was not alone in its verdict: independent investigations running separate probe sets converged on the same family. The paper states its own limit plainly — a vocabulary identifies a lineage, not an operator — and that limit is exactly what the grading will test.

Status

The ledger is live. Entries are appended as predictions are registered and graded as the answers arrive.
Entry 001 awaits an official claim. Ox Alpha's free preview opened 2026-08-20 with roughly a week announced, putting the window's close around 2026-08-27 — and the reveal lag on earlier anonymous previews has run from days to roughly two months, with Owl Alpha staying anonymous that long before Meituan claimed it as LongCat-2.0.
Grading updates are dated and appended. A graded entry is never rewritten.

The Reveal Ledger

What is the Reveal Ledger?

Abstract

Why keep a ledger?

What qualifies as a registered prediction

Grading rules

The ledger

Status

Links

Frequently asked

Stay Ahead of the Curve