Every claim cites its source — and why that beats a chatbot that invents one

A consultant lives and dies by one question after every claim: says who?

A chatbot can't answer it. It produces a fluent, confident sentence and has no durable idea where that sentence came from. Ask it to cite the source and, on a bad day, it will manufacture one — a real-sounding journal, a plausible page number, a quote nobody ever wrote. That's not a bug you can train away. It's what a text generator does.

Pith starts from the opposite end.

The claim doesn't exist until the source does

In Pith, you don't prompt a model and hope. You bookmark what you actually read. Pith reads those saved sources and assembles them into a cited wiki and per-client briefings. Every sentence in that wiki traces back to a source you put there.

The order matters more than it looks. With a chatbot, the sentence comes first and the citation — if any — is bolted on afterward, which is exactly when a model is free to invent. With source grounding, the source comes first. The claim is downstream of it. There is no point in the pipeline where a statement can appear without a saved source underneath it, because the saved source is what produced the statement in the first place.

That's the difference between retrieved and grounded. Retrieval finds something that looks relevant. Grounding means the thing you're reading was built out of sources you chose, and each line knows which one it came from.

Provenance is machinery, not decoration

Here's where Pith deliberately diverges from the tools that wave trust around.

The wiki page is the artifact. You read it like a clean internal memo — no confidence pills, no source-tier badges, no little warning triangles cluttering the prose. The provenance lives underneath: computed for every claim, stored, and exposed where it's actually useful — through the API, through MCP, in your exports and your audit view.

The page is what you read. The provenance is what survives the partner asking says who? three months later.

This is a design choice, not an oversight. Provenance plastered on the page is theatre; it makes a document look rigorous without making it more defensible. Provenance you can query and export is the thing a junior pulls when a deliverable gets challenged in a steering committee. One of those is a UI flourish. The other is the work.

Why this is the consultant's version, not the student's

Plenty of tools cite. Few cite in a way that holds up when money and reputation are on the line.

If you're writing a term paper, a wrong citation costs you a grade. If you're writing a market-entry recommendation or a tech due-diligence memo, a fabricated source costs you the engagement — and possibly the next three. Your output has to be defensible, which means every load-bearing claim has to survive someone hostile, well-paid, and motivated to find the crack.

A general assistant can't give you that, because it can't promise the source exists. The risk isn't that it's wrong often — it's that you can't tell which sentences to trust without re-checking all of them, which defeats the point of using it. That's the trap of AI hallucination: the confident tone is identical whether the source is real or invented.

Pith removes the question by construction. Nothing reaches your wiki that you didn't save. Every briefing is assembled from those saved sources and links back to them. When someone asks says who, the answer is one click, not one prayer.

What this buys you operationally

Three things change once your knowledge base is grounded rather than generated.

Briefings get trusted. A weekly client briefing that cites its own saved sources is something a partner can forward without re-reading every line — because spot-checking takes a click, not an afternoon.

Audits stop being dread. When a deliverable is challenged, you don't reconstruct your reasoning from memory. You export the trail: claim, source, the saved passage it came from.

And institutional memory compounds instead of evaporating. The reading your team does this quarter becomes a cited substrate the next engagement builds on — not a folder of links nobody trusts.

This is also the cleanest line between Pith and a general chatbot. If you've compared the two for real work, you already know the failure mode — see Pith vs ChatGPT. One invents a source to finish the sentence. The other refuses to write the sentence until you've given it a source.

For defensible work, that refusal is the feature. Says who should have an answer before anyone asks.