Editorial policy

The shape of an article

Articles are built on the same skeleton: what the subject is, what has been demonstrated, what the unsolved problems are, who is working on it, and what the objections are. The last section is not optional. An article that cannot say why its subject might fail is incomplete, and it is written until it can.

Classification is validated at build. A status outside the six values below, a TRL outside 1 to 9, an unknown portal, or a missing summary fails the content check rather than defaulting to something plausible. So a filter on this site returns a set you can trust: nothing is missing because someone forgot a field.

The six evidence grades

Status is the most load-bearing field on the site. It answers one question: how much of this has been shown, as opposed to argued? The definitions below are read from the taxonomy module at build, so this page cannot drift from the code that classifies the 205 articles it describes.

Grades the specific application this article describes, not every use of the underlying platform. A technique can be routine in one tissue and speculative in another. An article on a platform is graded on the application it describes. CRISPR–Cas9 is graded on an approved ex-vivo therapy, not on the idea of editing genes; the grade would be different on an article about editing embryos, and there is one.

Readiness measures how far something has been built and demonstrated. It is not a measure of how well it works, and it is silent on whether the demonstration was in a person or an animal.

Evidence grades and horizons are not printed on articles about people, organizations, or completed events. “How much of this has been demonstrated” is a question about a claim, and a person is not a claim. Those articles carry a classification and a portal instead.

Two need a note. Contested is not a midpoint between established and speculative; it is a different axis. It marks claims under active dispute or practices widely condemned, so a contested subject may be very well demonstrated indeed. Historical covers completed events and superseded ideas. It records that the matter is closed, not that the idea was wrong.

Rank exists so the corpus can be sorted by how solid it is. A low rank is not a verdict: plenty of the best articles here are speculative, and they say so at the top of the page.

Readiness: TRL 1 to 9

Where a subject is concrete enough to have one, an article carries a technology readiness level — the nine-point scale NASA introduced for flight hardware and the European Commission adopted for research funding. It measures distance travelled from principle to operation, and nothing else.

Status and TRL are independent. A technology can sit at TRL 8, qualified and in routine use, and still be contested: deep brain stimulation is deployed hardware and an unsettled question about mechanism. Another can sit at TRL 3 with nobody disputing anything, because nobody has claimed anything yet.

TRL is left blank where it would be nonsense: concepts, people, events, and organizations have no readiness level. Personal identity is not at TRL 4. The readiness matrix crosses the two fields, which is the fastest way to see whether a portal holds deployed systems or arguments.

The horizon field

Horizon answers a narrower question than it looks like: when could this plausibly matter at scale? The buckets are historical, present, late 2020s, 2030s, 2040s, 2050s+, and indefinite.

The field claims one thing: estimates made by people cited in the article body fall in that range. It is not a prediction by this site, not a probability, and not a schedule. Where an article’s sources decline to give a date, the horizon is indefinite — an honest answer rather than a hedge, and the answer for most of what depends on whole-brain emulation.

The timeline is built entirely from this field. Reading it as a roadmap is a misreading.

Sourcing

Citations point at real work — peer-reviewed papers, trial registrations, regulatory decisions, or a primary statement from the organization concerned. A claim that cannot be traced to one of those is cut, not softened into anonymity. Where the source is a press release or a conference abstract, the sentence says so. That difference is frequently the whole story.

Claims are written down to the evidence, never up to the headline. A cohort study that found lower all-cause mortality stays a cohort study that found lower all-cause mortality; it does not become extends life. Effect sizes carry their sample size and whether anyone has replicated them. Where the literature is genuinely split, both readings appear and the split is described rather than resolved.

This site takes no funding from any organization it covers, and no article is written or reviewed by a party with a commercial interest in it. Company claims are attributed to the company by name.

Every citation is labelled with what kind of thing it is — paper, preprint, trial registration, regulatory decision, report, book, journalism, statement, or law. The label answers exactly one question, who, if anyone, checked this before it was published, and deliberately says nothing about quality: a preprint can be excellent and a paper can be wrong. The distinction that matters most is statement. A company describing its own results records a claim; it never establishes one.

A citation may also carry a short note saying why that source is the one cited — the population it studied, the thing it did not measure, the reason a well-known paper is weaker or stronger than its reputation. Those notes are optional and meant to stay optional. One invented to fill the slot would be worse than none, because it would look like verification.

Saying where the evidence runs out

Two of the five callout types exist only to admit something. A caution marks a claim weaker than its coverage — a result in mice, a sample of twelve, a surrogate endpoint standing in for the thing anyone cares about. A debate marks the different situation where the evidence is in and competent people still disagree about what it means.

Written one article at a time, each is visible only to whoever happens to be reading that page. Collected, they become open questions — every stated limit in the corpus at once, which is a view of the field that no single article can give and that secondary coverage rarely attempts. Popular writing aggregates the findings; the caveats are what get dropped in the retelling.

The same page publishes this site’s own backlog: gaps each article declares in itself, and the count of articles that state no limit at all. For a person or a completed event, carrying no caveat is usually correct. For a technology it is more often a gap in the writing than a fact about the world, which is why the number is printed rather than presented as a clean bill of health.

A mouse is not a human

Animal results are never presented as human results. Every sentence reporting a preclinical finding names the organism, and the same paragraph says whether it has been tested in people. This is the field’s most common failure mode.

Rapamycin extends median lifespan in mice by a margin replicated across independent labs. That is a real and important result. It is not evidence that rapamycin extends human lifespan by any margin, and an article that lets a reader come away believing otherwise has failed however carefully it hedged elsewhere.

The rule extends downward. Cell lines are not tissues, organoids are not organs, and an effect in a dish is a hypothesis about an animal, which is in turn a hypothesis about a person. Every step is named.

Held as a writing habit, that rule gets honoured unevenly and cannot be checked. So it is also a field. Every technology and intervention states, in one sentence, what has been shown in people — and where the honest answer is nothing, the sentence says nothing. It sits in the infobox directly under the evidence grade, because a grade reading Experimental and a line reading no human study has been run are not the same claim, and the second is the one that stops a mouse result being read as a human one.

Three lines every intervention carries

An evidence grade and a readiness level between them answer “how well established is this” and “how far has it been built”. Two questions a reader actually has are left over, so each article about a technology or an intervention answers them on the record rather than somewhere in the prose.

Accesssays who can obtain the thing today and on what terms: approved and reimbursed, trial-only, or sold by clinics operating outside approval, and roughly what it costs where that is public. Almost no reference work answers this, and it is the question that follows immediately after “does it work”. It is a statement about availability, never a recommendation.

Reversibility is the axis neither of the others can express. A daily pill at readiness 9 and a heritable edit at readiness 3 differ on something more consequential than development stage: what happens if the science turns out to be wrong. The five values run from reversible, where stopping restores the prior state, to irreversible, which includes anything heritable. Context dependent is not a hedge — some interventions genuinely reverse in one tissue and not another, and the article has to say which. The readiness matrix crosses the two axes, and the cell worth looking at first is the one holding things that are both unproven and permanent.

None of the three appears on an article about a person, an organization, or an event, for the same reason those articles carry no evidence grade. George Church is not a claim, and asking how reversible he is would be a category error.

Wanted pages

Article text links freely to topics that have no page yet. Those links render in a different colour with a trailing question mark, so nobody is sent to a dead end without warning, and each one is collected on wanted pages, ranked by how many articles asked for it. Right now 8 topics are in that queue, led by alzheimers disease, requested by 1 article.

The point is to let the corpus argue with itself about what is missing. A wanted link is never deleted to tidy up the red: the demand signal is worth more than the tidiness.

Relatedness and backlinks

Neither is hand-curated. Backlinks are read off the link graph: an article appears under what links here when its text contains a wikilink to the article you are reading. Nothing else qualifies, and the count is exact.

Related articles are scored. A mutual link, where two articles cite each other, is worth the most, because reciprocity is the strongest available signal that two subjects are genuinely adjacent. A one-way link is worth about half that. Shared tags come next, since tags are hand-assigned and specific. Shared portals count least, because a portal holds dozens of articles and proves almost nothing. A small bonus goes to pairs sharing several outbound neighbours, which surfaces siblings that never link to each other directly.

So a related list describes the link structure of this corpus, not the field. Thin lists mark thin areas, which is information too.

Corrections

Errors are fixed in place and the article’s updated date changes with them. Where a correction alters a claim rather than a typo, the change is described in the body rather than folded in silently. Recent changes lists every article by date; the RSS feed carries the same list.

A claim resting on a retracted paper is removed, not downgraded. A forecast overtaken by events stays on the page, dated, and marked as overtaken: a field’s record of missed predictions is part of what a reader needs to weigh the current ones.