Frequently asked questions

What it is

Is Galedra against misinformation?

Yes, in the only way that lasts. A false claim spreads faster than anyone can check it, and every check that is made is lost in a thread. Galedra keeps the check: signed, public, and cumulative, so the claim meets a record the next time it comes round instead of another repost. It does this without verdicts, scores of people, or a different procedure for any side: it shows the evidence for and against and lets the reader see. A record that shows its reasons is harder to dismiss, and harder to capture, than a verdict.

Is Galedra a fact-checker?

No. It stores claims, not verdicts. For each claim it keeps the evidence for and against, where each piece came from, who linked it and why, what audits found, and what a named scoring model concludes from all that at a given point in the log. It tells you why you might believe or doubt something. It never tells you what is true.

What is the “log”?

An append-only record. Every write, whether a new claim, a piece of evidence, a link, an audit, or a moderation action, is a signed contribution with a gap-free sequence number and a hash that chains to the previous entry. Claims, evidence, and scores are projections of that log: drop them, replay the log, and the same hashes and scores come back.

Why “a heap of witness”?

From Galeed, the cairn Jacob and Laban piled up at the end of a bargain neither trusted the other to keep. Two parties, no shared authority, and a durable marker left in the open so that later neither could claim the agreement had been otherwise:

And Laban said, This heap is a witness between me and thee this day. Therefore was the name of it called Galeed.

Galedra is that cairn for claims rather than covenants: a durable public marker of what was asserted and what bore on it.

Scores

What does the number mean?

It is always shown as “0.86 under ledger-default@0.1.0 at snapshot 212”: a probability produced by a named, versioned model from the evidence graph as it stood at one sequence number. Change the model or the snapshot and you may get a different number. It is deterministic and reproducible. It is not objective, and it is never “86% true”.

Why do I see a headline like “Unresolved” and no number?

Answers come first and numbers on request. Every claim page leads with a short assessment and a cited explanation. The probability, the full trace, and the review checklist are behind “Show calculation”. Two states carry no number at all: Insufficient evidence, because nothing counted bears on the claim, and Not assessed, for value judgments, metaphysical claims, forecasts, and other statements a probability would misrepresent.

Why did four sources count as one?

Repetition is not corroboration. Evidence that traces to the same origin, such as a press release and three articles quoting it, is placed in one independence group, and only the strongest member counts. The suppressed members are still listed on the claim page, marked as dependent.

What are “review checks”?

A count of four things a careful reviewer would do: read the primary source, search for opposing evidence, review independence, and review qualifiers. They are reported as “3 of 4”, never as a percentage, and they are separate from the probability. A claim can be well reviewed and unresolved, or barely reviewed and provisionally supported.

Why are there two scoring models?

So that no single model can be mistaken for the truth. Both ship in the log, both run on the same graph, and the compare view shows where they disagree and why. A claim whose assessment changes between models is labelled as depending heavily on modelling choices.

People and agents

Does Galedra run an AI?

No. Galedra runs no model. It is a deterministic framework: signed records, closed vocabularies, replayable projections, and versioned scoring, through which people and the AI assistants they bring collaborate on a durable record of claims and the reasons for them. Every judgment that needs a model is made outside Galedra, by a person or by an assistant acting for one, and enters as a signed contribution open to audit. The server's only outbound requests fetch sources already in the log, to hash them and check quoted passages.

Can an AI add evidence?

An AI agent can lease a signed task packet, do the work, and submit a signed result. The result is validated against a closed set of operations, accepted into the log, and left permanently open to audit. An AI's own output is never evidence by itself: it weighs zero. Evidence is a passage at an exact location in a source.

What is reputation?

Audited reliability at specific tasks in specific domains, computed from audit history and reproducible from it. It is not authority, it is not a score of the person, and in this version it is not an input to any claim score.

Who moderates, and can moderation hide things?

Moderators are designated keys, listed publicly at the metadata endpoint. Quarantine withholds content but leaves a visible stub. Takedown removes bytes but keeps the hashes and the chain. Every action is a signed contribution in the public moderation log. Disagreement with a claim is never a reason.

Is my data public?

In this version, yes: the log is a single shared record and every contribution in it is readable. Personal assessments that live outside the shared log are planned and not yet built. Claims about identifiable private individuals are out of scope, and you affirm that for every claim you add.

Epistemology

Is this a theory of knowledge?

No. It is bookkeeping for justification. Epistemology asks what knowledge is and when belief is warranted. Galedra takes no position on the first question and a modest one on the second: a belief is better placed the more its reasons can be traced, examined, and challenged. It records reasons, not knowledge, and it never promotes a claim from “justified” to “true”.

Where does it stand on justification?

Closest to evidentialism: the standing of a claim is a function of the evidence that bears on it, and of nothing else. Who submitted it, how many people endorse it, and whether it matches consensus carry no weight. The classical regress problem, that every reason needs a further reason, is not solved here. It is made visible. Every chain of support bottoms out in a passage at an exact location in a source, and you can decide for yourself whether that floor is firm enough.

What about testimony?

Most of what anyone knows comes from other people, so testimony matters. The ledger treats it carefully rather than dismissively. A person's or agent's audited reliability at a specific task in a specific domain is recorded and reproducible, but it never stands in for evidence and never enters a claim score. Authority is a reason to look, not a reason to believe.

Why is “unknown” treated as a result?

Because suspending judgment is an epistemic act, not a failure of one. A claim with nothing counted for or against it reports insufficient evidence and carries no probability. The prior alone never yields a verdict. Fallibilism is built in: no conclusion is immune from re-examination, corrections are new entries, and history is never rewritten.

Are the probabilities credences?

They are model-conditional, not personal. A number is what a named, versioned scoring model concludes from the graph at a sequence number. It is not anyone's degree of belief and not a frequency. Two models ship, and where they disagree the claim is labelled as model-dependent. Determinism is not objectivity, and the display says so. Repetition is not corroboration either: evidence with a shared origin is grouped so that dependent reports do not count as independent confirmation.

How does it handle disagreement and values?

Contradictions coexist; nothing is overwritten. Disagreement is localized to the specific link or interpretation in dispute rather than to whole claims or whole people. Value judgments are kept off the probability scale entirely. A recommendation is not made empirical by being stated confidently, and the ledger refuses to score it as if it were.

Hermeneutics

What does interpretation have to do with a ledger?

Most claims are about texts: what a study reports, what a witness said, what a law requires, what a passage means. Hermeneutics is the discipline of interpreting texts, and its central insight is that reading is never neutral. The ledger does not pretend it is. It makes each interpretive act an explicit, attributed, auditable object.

How does it separate what a text says from what it is taken to mean?

Textual claims and interpretive claims are different types. Evidence is a passage at an exact location, quoted and hashed. The link “this passage supports that claim” is a separate object with a count of interpretive steps, and the more steps, the less it weighs. A direct quotation carries zero steps. A reading that infers motive, context, or implication carries one or more, and it says so.

Does it choose between readings?

No. Competing readings of the same passage are recorded side by side, each with its own provenance and audit history. Ricoeur distinguished a hermeneutics of trust, which listens for what a text is trying to say, from a hermeneutics of suspicion, which asks what it conceals. The ledger asks for both: an opposing-evidence search is one of the four review checks, and a claim is not well reviewed until someone has looked for what would count against it.

What about the hermeneutic circle?

Parts are understood through the whole and the whole through its parts. Atomizing a document into claims is itself an interpretive act, and the ledger treats it as one: proposals are edited by a person, attributed, and left open to challenge, and the source is kept intact so any claim can be read back against its context. Gadamer's point that every reader brings prior understanding is not resisted. It is recorded. An interpretation names its interpreter, its steps, and the assumptions that connect the passage to the conclusion.

Where are the limits?

Translation, textual criticism, historical context, and theological meaning are distinct kinds of question, and the ledger's claim types can hold them apart. It does not yet have scoring models fit for legal or theological interpretation, and it says so rather than forcing them onto an empirical scale. Its aim is narrower than hermeneutics as a whole: that when someone asks “why should I read it that way?”, the answer can be followed all the way down to the passage.

Using it

How do I start?

Most people start from their own assistant: connect it, then say "galedra:" followed by what they were about to share. Without one, create an account and paste a paragraph under Analyze text. A signing key is generated and held for you. Galedra does not break the text into claims itself, since it runs no model: it gives you a prompt for your own assistant, opens the breakdown to volunteers' assistants, and offers a form to do it by hand. Each claim gets an answer card as evidence arrives.

My assistant does not seem to know a new Galedra feature.

Your assistant's host keeps a copy of Galedra's tool list from the moment you connected. Galedra sends its working instructions inside every reply, so most changes reach your assistant at once; a new tool or a changed field appears only after you disconnect and reconnect Galedra in your assistant's settings. Your account, your connection, and everything recorded are unaffected.

How do I use it with the AI I already have?

Give ChatGPT, Claude, or any assistant that speaks MCP the address on Connect an assistant. When it connects, Galedra asks you one question: connect under your name, or continue anonymously. Then say "check this in Galedra before I post it". The assistant searches first, reads the sources itself, records what it found in one signed call, and hands back Galedra's headline, what to say instead, and a link. Its own opinion is never evidence and never enters the assessment.

Do I need an account?

No. Anonymous use needs nothing at all. Anonymous work is signed under a fresh key, checked by auditors more often, and labelled as such on every card. If you later want it under your name, every anonymous result carries a link that puts it there with one click while signed in; the adoption is itself a public log entry.

What if my assistant cannot connect at all?

Ask it to draft the investigation bundle as JSON and paste it at Record an investigation. It is recorded the same way, no account needed.

How do I help check what is already recorded?

Connect your assistant under your name and say "work the open tasks in Galedra". Every recorded claim opens tasks: search for opposing evidence, look for omitted qualifiers, verify that a passage supports its claim, group sources that share an origin. Your assistant leases the highest-priority one it may hold, never one on your own claims, reads the sources, answers, and reports each with a link. A search that finds nothing is a result too. What it submits is signed for it, counts once accepted, and stays open to audit. To run your own agent instead, the example client in examples/agent/ uses the same task endpoints; any task page shows the command.

What is a snapshot?

A sequence number in the log, optionally pinned with a label. Any score can be recomputed at any past sequence number under any released model, byte for byte, and two snapshots can be compared claim by claim.

How do I check that nothing has been tampered with?

Every contribution page shows whether the client and server signatures verify and whether the chain is intact. The API exposes the same checks, the constitution's integrity hash, and the public keys, so you can verify the whole log with your own code.

What will it never do?

Determine ultimate truth, build one universal probability model, arbitrate values, replace courts, scientists, journalists, or analysts, score people or parties, prevent every coordinated attack, act as a blockchain, or index the web.