Corrections & Refinements — the visible trail¶

This page exists so that the accountability a byline normally supplies is supplied instead by a record anyone can inspect. EveryClaim carries no author's name (by design); in exchange, it owes readers something a named outlet is rarely asked for — a permanent, dated, public account of every time it was wrong in a way a reader could have seen. This is that account.

The commitment, in one line: corrections here are permanent and conspicuous. Nothing is silently edited away. When this site changes a published fact, the change is logged below with the date, what changed, and why.

The correction policy¶

What counts as what. We distinguish three things, because collapsing them is itself a way to hide errors:

Correction — a factual error in published material: a wrong number, a wrong date, a misattributed source, an unsupported claim. Every correction is logged below.
Refinement — a calibration or clarity change that does not reverse a fact: tightening a verdict, adding a caveat, sharpening a definition. Logged when it changes what a reader would take away.
Update — new events or newly available evidence added to a living page. This is not an error; it is noted on the page itself with an as-of date, and is not logged here as a correction. (Ongoing-event figures always carry explicit dates for exactly this reason.)

Retraction. If a claim cannot be supported, it is struck — visibly, with a note explaining why — never quietly deleted. A removed claim leaves a mark.

How to report an error. If you find a claim that fails its tag — a source that does not say what we claim, a figure that is wrong, a dispute we have miscalibrated — report it through Suggest a Topic. Substantive challenges receive a response, and where the challenge is right, a logged correction. Finding an error here is a contribution to the project, not an attack on it.

What this log does and does not include. Most errors in a project built this way are caught before publication, during the source-audit and verification passes that precede every document — for example, the "Muslim Brotherhood black-list count" was corrected in research before the terror-finance document was written, and the Journal of Palestine Studies' own correction to the "160 children died digging tunnels" figure was incorporated in drafting, so neither ever reached a reader. This log is deliberately not for those. It is for the other kind: errors that were live, that a reader could have seen, and that were changed afterward. Padding a corrections log with pre-publication catches would defeat the only purpose it has.

The log — post-publication corrections¶

Most recent first. Each entry is tagged by type.

An independent AI review credited the site's transparency, versioning, source registry, and refusal to overclaim — and then drew a fair line: "Its transparency is meaningful; its independence and neutrality are not yet demonstrated." Rather than bury that, the About page now quotes it and answers it plainly, distinguishing two things the critique compressed: neutrality in conduct (demonstrated and checkable — the two-directional corrections, the frame audit, the one-standard claim-checker) versus institutional independence (genuinely not demonstrable from inside a single-curator, anonymous, self-governed project). No fact was changed; what changed is that the site now names its own structural limit and points to the fix — external review (a standing adversarial panel; published dissents). An honest site amplifies its fair critics; this records that it does.

2026-08-24 · [TOOL] · Check-a-Claim graded a hedged "a factor" claim as if it claimed "the cause"¶

A reader asked the claim-checker whether Islam is a factor in the region's catastrophes against minorities — explicitly adding "not the only cause." The engine returned PARTIALLY TRUE, then spent its answer refuting a claim the reader never made: that Islam is the sole or primary cause ("the agent is often a specific state," "much was intra-Muslim," "attributes too much agency to Islamic rule alone"). Every caveat answered the strong claim; against the modest claim actually asked, the correct verdict was VERIFIED. Straw-manning a claim upward and then refuting the stronger version is how a tool that feels neutral smuggles in a thumb on the scale — the mirror image of the certainty-upgrade the engine already forbids.

What changed: a new engine rule (Rule 12) requires the verdict to grade the claim at the strength it asserts — "a factor" / "involved" / "contributed" is not "the cause" / "the sole cause" / "every," and must not be refuted as if it were; a documented contributing factor grades VERIFIED even where the excerpts also show it was not the only cause, with that nuance moved to amplification. The rule is symmetric: a maximal claim ("the sole cause," "all," "every") is still graded at full strength and can still fail — it prevents rescuing a claim downward as much as inflating one upward. The PARTIALLY TRUE definition was tightened to match ("missing context" excludes caveats the claim already states). Separately, Doc 53 gained a "First, separate two questions" section that answers the modest question plainly — is Islam a factor? yes — distinct from the slogan it corrects (did Islam wipe them out? no), so the affirmation leads and the engine has explicit text to ground it. The fix cut, as it should, toward stating a true thing plainly — not toward any side.

Follow-on the same day — including a correction to the correction, logged because that is the rule here. Testing first revealed a phrasing inconsistency: the question form ("is Israel guilty…?") and the Hamas parallel returned DISPUTED while the statement form returned CONTRADICTED. The first fix over-corrected — it forced every "guilty of" claim to CONTRADICTED / "not guilty," which flattened Hamas and Israel to the same verdict. That was itself an error: a false symmetry that laundered a real, documented asymmetry — Hamas's genocidal intent is openly declared in its founding charter; Israel's alleged intent is contested and leans against. Treating them identically is exactly the failure this site exists to prevent, in the opposite direction. The final rule grades a "guilty of" claim on the intent evidence for each party: where intent is declared and acted upon (Hamas, Hezbollah, the Houthis) the claim is graded on its substance and is never answered "not guilty"; where intent is genuinely contested (Israel and the crime of genocide) the claim is not established — CONTRADICTED, "No — not guilty." Verified live: "Hamas is guilty of genocide" → PARTIALLY TRUE (declared charter intent + the October 7 mass atrocity + the ICC extermination charge; the sole precision is that the specific crime "genocide" vs "extermination" was not separately adjudicated — documented guilt, not exoneration); "Israel is guilty of genocide" → CONTRADICTED; "Israel is committing genocide" → DISPUTED. The lesson, logged for ourselves: one standard means the same test, not the same verdict — identical results across unlike evidence would be the bias, not the cure.

2026-08-24 · [REFINEMENT] · The Gaza genocide answer now separates the verdict question from the factual question¶

A reader challenged the Gaza genocide FAQ: its "disputed" framing, they argued, let an accusation sit as a peer of Israel's documented conduct, when the plain answer to "is Israel guilty of genocide?" is no — no court has convicted anyone, and an accusation is not a verdict. The challenge was substantially right, and it exposed that the page was answering one question ("is genocide occurring?") while readers were asking another ("is Israel guilty?"). Merging the two is what produced the appearance of false balance.

What changed: the FAQ now opens by splitting the two questions. Guilt is a verdict, and none exists — so by the presumption of innocence and the burden of proof, Israel is not guilty of genocide, stated plainly. This is one standard, not a lean: the Nazism/Hamas FAQ already holds that Hamas, too, is not "guilty" of genocide despite the clearest stated intent of any actor — no court has adjudicated the charge against anyone here. The factual question — is genocide occurring? — stays DISPUTED on the evidence, with the honest addition that its crux, intent, leans against the charge under the governing "only reasonable inference" standard. The refinement also demotes the population-growth argument from proof to supporting point: it rebuts the maximalist "exterminating a people" rhetoric but not the legal charge (genocide is "in whole or in part" — Srebrenica at ~8,000 dead), so the "not guilty" rests on the burden of proof and the unmet intent bar, not the demographics. No documented fact was reversed; what changed is that "not guilty as a verdict" and "unproven as a fact" are now stated as the two distinct answers they are.

We ran a new kind of audit on ourselves. The pre-launch audit that found ~390 issues checked claims — is any statement false? It cannot catch two subtler biases: frame bias (which contested vocabulary a site adopts as its unmarked, "neutral" default) and omission bias (what is never covered, so a reader never learns a question exists). A one-standard site's blind spots on these are invisible precisely to the readers without background — the audience it exists to serve. So we swept the Israel/Palestine core for both.

The audit's finding was, on balance, reassuring: on claims the corpus is close to one-standard, and where it leans it leans pro-Israel and openly marks those leans (UNRWA heritability, the right-of-return arithmetic, the "Nakba Rethink," modern-identity dating). But it found one bounded, invisible counter-tilt — the baseline territorial vocabulary. The corpus used "West Bank"/"occupied" as unmarked, VERIFIED fact (and International Law literally tabled the West Bank occupation as the "clean" framework), while never once surfacing the Israeli naming ("Judea and Samaria") or the Israeli legal counter-position ("disputed territory") — even though the same document flags smaller disputes (genocide, apartheid) as contested.

What changed: the International Law table row no longer calls the occupation designation "clean" — it states "occupied" as the majority view (ICJ, ICRC, UN, and Israel's own Supreme Court) and notes the serious minority "disputed" position (the missing-reversioner jurists; the Levy Report); the Israel profile now surfaces the naming choice on first use; and a new FAQ — Is it "the West Bank," "Judea and Samaria," or "occupied territory"? — lays out both frames, the backstory, and the honest scorecard. This is not false balance: occupied remains where the weight of legal authority sits, and the pages say so. The change is that the site no longer presents that verdict as if no choice of frame were made. (The full frame-and-omission audit also flagged omitted Jordanian-rule history and some 1967 war-context gaps; those fixes are queued.)

2026-08-13 · [TOOL] · Check-a-Claim returned inconsistent, under-calibrated verdicts on "apartheid" questions¶

The live claim-checker returned "partially true" for "Is Israel an apartheid state?" without engaging Israel's documented record of Arab-citizen political rights — and returned a different verdict for the same question about Lebanon depending only on how it was phrased. Two fixes followed: a new page applying the 1973 Apartheid Convention's legal elements to every state in the region at one standard, and a new engine rule requiring legal-category claims (apartheid, genocide, and the like) to be graded by their definitional elements rather than freelanced. The whole-state claim now returns "contradicted," with the citizen record stated in the answer. The correction cut in two directions at once: against an anti-Israel over-claim, and against the tool's own inconsistency.

2026-08-11 · [CONSISTENCY] · Five documents published without the standard summary box¶

Documents 32 through 36 went live missing the plain-language "Short Version" summary that every other document carries — a defect a reader comparing two pages would see immediately. The summaries were written and added to all five. The gap came from a document template that had not been updated as the collection grew; the template was fixed so it cannot recur.

2026-08-08 to 2026-08-10 · [TOOL] · Check-a-Claim gave imprecise verdicts on comparative and measure-dependent claims¶

The live engine graded claims such as "malnutrition is worst in Gaza" as a vague "partially true" instead of performing the comparison the word "worst" demands. After stress-testing exposed the pattern, rules were added requiring the engine to extract and rank the comparable figures before ruling, to name which measure a verdict depends on, and to treat percentages and absolute counts as different rankings. Retested against the same claims, the engine now ranks the figures and rules accordingly.

2026-08-02 · [PRIVACY] · The project owner's name appeared on live pages¶

A user-experience review found the project owner's full name present on several live pages, as residue from internal working files. Because the project's design is deliberate anonymity (About), the name was removed the same day it was found, and a full sweep followed to remove it and other internal identifiers across the site. This entry is here because a correction record that hides its most awkward item is worth nothing.

Before first publication — the pre-launch hardening¶

These did not reach readers, and so are not "corrections" in the sense above. They are recorded here so the scale of the process behind the site is visible, not to inflate the count.

The claim-by-claim audit (July–early August 2026). Before the site went public, the corpus was audited against its sources. The audit flagged roughly 390 problems across about 1,760 tagged claims — including about 25 fabricated or misquoted sources (some of them AI-generated citations in the project's own drafts) and about 65 factual errors. Every one was corrected against real sources, rewritten to what the evidence supports, or removed, before publication.
The adversarial red-team review. A separate hostile review of the near-final site produced twelve high-severity findings — an inaccurate quotation, instances of one-sided labeling, an anachronism, among others — all corrected before wide release.

The detailed pre-launch correction logs live in the project's source repository.

The standing rule¶

From here forward, every post-publication correction is logged on this page — dated, with what changed and why — and no published fact on this site is ever changed silently. That is the deal this project offers in place of a name: you cannot check who wrote it, but you can check everything it ever got wrong.

See also: How We Verify · Verification Standards · Sources & Links · About