Claims and Stanced Citations

"The corpus holds what you wrote. The evidence layer holds what you concluded — and exactly which words made you conclude it."

Retrieval Finds; Evidence Concludes

Everything up to now has been about finding text. The evidence layer is a different kind of thing entirely: it records what you concluded from what you found. Search answers "where did I write about market panic?" Evidence answers "I believe structure beats willpower — and here are the four specific passages that led me there, and the one that argues against it." One is a lookup; the other is a position you're taking, with receipts.

A Claim Is About the Corpus, Not In It

This is the crucial distinction. A claim is a statement you assert — and it is not a chunk sitting somewhere in your writing. It's a judgment about the corpus, formed by reading it. You might have written around an idea for ten years without ever stating it in one clean sentence. The claim is that sentence, minted by you, now, as a first-class record — a durable node that citations can hang off.

Citations Carry a Stance

A citation isn't a neutral pointer from a claim to a passage. It carries a stance: does this passage support the claim, refute it, or qualify it? That single field is what turns a pile of references into an argument. Without stance, a claim with eight citations tells you nothing — are they eight confirmations or six confirmations and two serious objections? Recording the stance means your evidence keeps its own disagreements visible instead of flattening them into a citation count.

Judgment mints the claim; the engine only records it. Extracting claims from text is an act of interpretation, and interpretation is exactly what a deterministic engine must never do. So Lantern never mints a claim. A person — or an assistant acting as a client, above the API — decides that something is a claim and that a given passage supports or refutes it. The engine records that decision, resolves the citations, and revalidates them mechanically. The 'no model in the path' rule doesn't get suspended just because the layer got interesting.

Evidence Is Precious, Not Derived

Here's why this layer plays by the snapshot rules rather than the index rules: you cannot re-ingest your way back to a claim. Purge the FTS index and re-index — you get every chunk back. Purge your claims and re-index — you get nothing back, because no amount of reading the corpus regenerates the judgment you made about it. Evidence is precious state, in exactly the sense the snapshot store is. So it gets the same protections: deleting a claim never cascades into deleting its passages, and a passage refuses to be deleted while any citation still points at it. Judgment is the one thing in the system that a rebuild can never restore.

Code

Claims are minted above the API; stance makes them an argument·python

# Judgment (a person, or an assistant as a client) mints the claim — never the engine.
claim = create_claim(text="Structure beats willpower for long-term investing.")

# Citations carry a STANCE. Without it, eight citations tell you nothing.
cite(claim, passage=pin_passage(chunk_a), stance="supports")
cite(claim, passage=pin_passage(chunk_b), stance="supports")
cite(claim, passage=pin_passage(chunk_c), stance="refutes")   # kept, not hidden
cite(claim, passage=pin_passage(chunk_d), stance="qualifies")  # "...except in a crash"

# Precious-state rules, same family as snapshots:
delete_claim(claim)          # never cascades into deleting its passages
delete_passage(passage_c)    # -> Conflict: still cited by a live claim

# And what a rebuild can NEVER restore:
purge_index(); reindex()     # every chunk comes back...
# ...but not one claim. No amount of re-reading regenerates a judgment.

Exercise

Write down one thing you believe that came from your own accumulated reading or experience — a claim you'd defend. Now list the specific evidence: for each piece, mark whether it supports, refutes, or qualifies your claim. Be honest and include at least one that cuts against you. Notice what the 'refutes' entry does to the claim: it doesn't destroy it — it makes it a position rather than a slogan.

Hint

If you can't find a single refuting or qualifying piece of evidence, you probably haven't looked — or the claim is too vague to be wrong. A claim worth recording is one where the counter-evidence exists and you chose your position anyway, with the tension recorded rather than hidden.