Why you can trust this

Method

This project exists because reasoning about physics is easy to get wrong and easy to fake. So none of it runs on trust. Every mechanism below is designed to make an unsupported claim impossible to write, rather than to catch it afterwards.

One rule

Prose may not contain citations. Prose contains claim IDs. Claim IDs resolve to ledger records. Ledger records carry verbatim quotes that a script re-finds in the actual source text. A change that breaks any link in that chain is rejected before it is published.

You cannot write "as shown in [some paper], the wall must be thinner than 100 Planck lengths". You write a claim ID, and that claim must already exist with a quote a program locates in the paper. This removes the opportunity to invent a source, rather than trying to detect invention after the fact.

The lifecycle of a claim

1 · Sourced. A paper enters the corpus only after its identifier resolves against arXiv, Crossref and INSPIRE, and only after it is checked for retraction against oracles that are themselves tested on a known-retracted control.

2 · Quoted. Evidence is a verbatim span whose context is cut from the source, never typed. A separate program re-finds every quote in the actual text; a quote that cannot be relocated is not evidence.

3 · Claimed. A claim states an assertion with its regime — what it assumes and where it stops applying — and rests on specific evidence or a runnable derivation. A number that matters is the output of a script committed to the repository.

4 · Attacked. A claim is not "verified" until an adversary has tried to break it and failed, with the attempt recorded. Refutations are kept, not deleted: a disproven claim is marked Refuted and superseded by the corrected version.

An adversary who did not write the claims

The sharpest test is independence. An audit by reviewers who wrote none of the claims or the tooling refuted three findings that the original author's own attacks had missed — including one of the author's own computations, corrected once an outside reader recomputed it from the paper's stated parameters. Those refutations are on the record, visible in the findings, because a research record that hides its corrections is worth less than one that shows them.

Measured failure modes

Each of these is a real error the tooling caught, and each is now guarded by a test:

About 10% of scholarly identifiers recalled from memory were wrong — and the dangerous ones resolve cleanly to a different real paper. Every identifier is now resolved, never recalled.
A search API returned an empty result that looked like a valid "no matches" — so results are counted and inspected, never trusted by status code.
A standard numerical integrator returned exactly zero for the warp energy — silently missing a sharp spike — which would have implied a free warp drive. Every quadrature is cross-checked against a second method.
A text-processing step deleted equations and inline numbers, which once hid a paper's own stated parameter and produced a false published finding. It now preserves them, and the correction is recorded.

What the numbers mean here

Status	Meaning
Verified	Survived at least one documented adversarial attack; sources clean; verified by someone other than the asserter.
Refuted	Disproven by an attack and superseded by a corrected claim. Kept visible on purpose.
Under review	Recorded, sourced, but not yet attacked. Not a conclusion.

The full toolchain and the 26 claims, 23 attacks and 44 quoted spans behind this site are in the open data.