In the order they are meant to earn their keep. The same log, scoring, and constitution serve all of them.
Check it before you share it
Someone is about to repost a claim. Their assistant checks it in Galedra, finds it recorded or records it, and hands back a plain headline, what to say instead, and a link. The rumour meets a record instead of another repost, and the next person to meet the claim inherits the check.
Example. "The city cut its police budget by 40%." The budget documents show a 4% cut and a reallocation. The claim reads contradicted, the narrower claim the documents support is offered as what to say instead, and the check page is the link to post.
Built. Connect an assistant, or paste a check by hand.
Verify the citations and statistics in an AI draft
Before an assistant's brief goes out, paste it in. Each claim gets its own answer, each citation is checked against a source or marked unverified, and every check is signed and reusable next time.
Example. The memo above: the four-source statistic collapses to one lineage, the journal citation turns out not to exist, and the recommendation is flagged as a value judgment rather than scored.
Built. This is the public demo above.
A private evidence ledger for a team investigation
An incident review, a due-diligence file, or an investigative story accumulates documents, quotes, and interpretations over weeks. Record each as evidence at an exact location, link it to the claim it bears on, and let the record outlive the people who built it.
Example. A vendor's status page says an outage lasted forty minutes; the ticket log shows three hours. Both are recorded at their exact locations, the claim shows as contradicted, and anyone can see which passage did the work.
Built. The same mechanics, on your own sources.
Track one research question over time
New studies arrive, replications fail, qualifiers surface. Pin a snapshot, and later reproduce exactly what the record supported at that sequence number under a named model.
Example. At snapshot 212 a claim was “leans supported” under ledger-default@0.1.0. A failed replication is added at 340 and it becomes “unresolved”. Both remain reproducible byte for byte, and the diff between them is a list of contributions.
Built. Snapshots and replay are in the log.
Keep what a text says apart from what it is taken to mean
Historical and textual sources are read through translation and commentary. The ledger records the passage, the translation, and the interpretation as distinct objects, and counts every interpretive step in the score.
Example. A passage says a figure “taught the making of blades”; a commentary reads this as “introduced weapons”. The first is direct textual evidence with zero interpretive steps; the second is an interpretation with one, and it weighs less until an audit confirms the reading.
Built. An internal stress test exercises it.
A whole speech, claim by claim
A speech, a sermon, or a podcast is too much for one check. Its structure is recorded first, as an outline of sections with one extraction task per leaf, and many volunteers' assistants work the leaves. The outline page shows counts by state, never a score for the speaker, and a step of reasoning across claims can be recorded as an inference, reviewed by someone else.
Example. A two-hour episode becomes twelve leaves. Three assistants extract, verification opens on each claim, and the share line reads "3 claims · 3 self-checked · 0 independently checked · 1 supported · 2 insufficient evidence" — it says plainly how much of that checking was the author's own, and grows as other people work it.
Built. Outlines, extraction tasks, and inferences are in the log.
Next, once these have earned their keep: competing hypotheses with discriminating evidence.
Galedra