A claims table with confidence marks is only worth what the marking discipline behind it is worth. Most of these were made as one-off decisions about a specific problem and then held as standing constraints — this is the first time they've been visible together, on any public surface, as a system.
Generated output enters at ◦ and cannot promote itself. A generated output naming a reference is a lead, not a verification.
One commit per pre-registration document. That makes contemporaneity
machine-verifiable via git log — a reader doesn't have to trust that a criterion
predates the analysis it was applied to.
Never rewrite history to tidy a repository. Rewriting changes every hash, including hashes cited inside the log itself. Fixes go in as ordinary new commits.
Regenerate outputs; never reconstruct a pre-registration. A pre-registration's value is that it was written at a particular time, which can't be recreated. A computed output's value is being the correct result of a deterministic pipeline — which can, and should, be re-run.
Anything produced by a research-synthesis system enters the register as recalled, not verified, regardless of how many citations it carries, and can't be promoted without a primary source being read by something accountable.
Why. The decision log already records a batch of LLM-generated computational artefacts containing fabricated results, hardcoded scores and mathematically guaranteed p-values — all deleted on discovery. A multi-agent system produces exactly that class of object: fluent, structured, cited, and generated without anything having read a source.
References are read at source, and what the reading found is recorded whether or not it was flattering.
What it produced. Twenty-four sources verified; four wrong first authors (Tjhung→Kenchel, Powner→Singh, Aida→Ando, Bzdok→Stanley); multiple framing errors; a citation error on Pizzarello C-2/C-3; over forty findings that changed the drafts. The error count is the credential — a verification pass that found nothing would be evidence that no verification pass happened.
When someone got there first, say so early and by name, credit what's theirs, then state the difference.
Instances. Hartman & Smith 2019 on operational-code priority, credited in §1 rather than a footnote (Chain 2.2). Wołos et al. 2020 had already run the redundancy test four years before this framework existed (Chain 4). Waddington-as-quasipotential is fifteen years old and taken (the escape law, in full). Naming it first converts the strongest available prior-art objection into evidence the paper was actually read.
A pre-registration document is written once and committed once. The consequence is itself a
small result: the single-commit property makes contemporaneity machine-verifiable via
git log.
Git history is not edited, even to remove a personal handle. A published note's credibility rests on its pre-registration log being demonstrably contemporaneous, and commit timestamps and hashes are that evidence. Rewriting history changes every hash — including hashes cited inside the log the note promises is unedited. Fixes go in as ordinary new commits.
Reconstructing a pre-registration would be fabrication — its value is that it was written at a particular time, and that can't be recreated. Serialising a computed output is not: its value is being the correct output of a fixed, deterministic pipeline, and determinism makes that verifiable rather than asserted. A re-run must reproduce already-published counts and matrices exactly, logged as a dated re-run, never backdated.
When a pre-registered criterion is replaced, the original stays in the log, unedited, alongside the replacement and the reason.
The error is written down before it's fixed. Errors are left in the log rather than edited out of it.
A withdrawn claim stays in the claims table as a withdrawn row, with its reason, rather than disappearing. Three instances, all on /claims/: compositional assemblies as an early heritable carrier, the merger obliging an archive, and the bioblob's original performance-breadth prediction — the last of which wasn't in any public claims table until this site.
Corrections to published work appear as a new version carrying a correction notice, not as a quiet replacement of the file. Instance: the set-valued descriptor note, v2, with a correction notice and a new §6 reporting the positive level-2 result.
When a section is rewritten, the superseded passage goes into an appendix with the reason it was withdrawn — record-before-repair applied to a paper's own history. Instance: the bioblob draft's own Appendix A, carrying five withdrawn passages.
Section count, page ceiling, reference count, figure count and permitted new results are fixed before any prose is written. The failure mode this addresses isn't laziness — it's expansion: scope grows, the voice gets stripped, nothing ships.
New ideas are parked, not pursued, and the register is append-only — each entry claim + hook + one bounded probe, so none of them can quietly become a project from inside the register.
Run a five-minute contact probe against the real system before designing anything on top of it, and commit discovered failure modes to the log immediately.
A metadata label is a claim, not a measurement. Instance: 1CR0 is annotated "hexameric" and is a 6₁ screw-filament excerpt (Chain 5.3).
A sequence model may be used for validity, never for ranking, because its likelihood is conditioned on what survived — a survival-conditioned process is not the unconditioned process.
A benchmark specification is written so it can be adopted by people who reject the framework that motivated it.
When a claim outruns its evidence, leave it unresolved and stated as a question rather than resolving it into an assertion that can't be defended. A mystery keeps the game infinite; woo cattens it into a claim that gets dismissed — the unopened box, not the box opened and found empty. This is the rule /open/ runs on.
Material originating in a period whose reliability is in question is held in a separate file so it can't interact with the rest by proximity, and enters only by an explicit promotion check.
An item leaves quarantine only if it survives a bridge case independently of where it came from. The strongest evidence the protocol works: the bioblob's founding idea appeared once, flagged as teleology and killed, then reappeared independently two months later by a different route, with the aim removed — and passed its own check twice.
Decisions made during the compromised period are subject to review from outside it, and the later judgement wins.
Where documents of different dates state a claim at different strengths, the newest governs, and earlier ones are marked as superseded rather than left to be discovered.
Live and unresolved. The 1.6 Ga chloroplast result exists at three different strengths across three documents, and none currently says which governs. The forced-ordering derivation is unaffected either way; what's unsettled is how strongly the coordinate may be stated. Per that uncertainty, this site states it at the safest of the three strengths throughout: the forced ordering is derived, the coordinate is a consistency result.
An internal disagreement is a defect regardless of which side is right.
Live and unresolved. The persistence-capacity partition, between the bioblob's M/T/R sub-classes and a companion ferality paper's cost-sign Class A/B — recorded as a commitment to reconcile before either ships. This is Chain 3's own ending.
Every one of these controls is what a reader checking a claim on /claims/ is actually relying on. If one of them looks like it's been broken anywhere on this site, say so.