rlat-refresh-facts

name	rlat-refresh-facts
description	Check whether the EXTERNAL (web-fetched) facts inside an rlat knowledge model (.rlat) are still true in the live world, by re-fetching each fact's source and re-reading it yourself — no API key, no metered cost. Trigger when the user asks "are my fetched facts still current?", "did anything I added from the web go stale?", "re-check my external facts", "refresh the facts in my .rlat", "is the external info in my knowledge base out of date?", or wants the corpus's outside facts re-verified against the live world. This is the staleness `rlat refresh` CANNOT fix — refresh only re-syncs corpus files to disk; an external fact's truth lives on the web. NOT for: corpus drift (use `rlat refresh` / `rlat reverify`); finding contradictions (use `rlat-contradictions`); finding gaps (`rlat-gap-scan`).
allowed-tools	Bash(rlat:*), WebFetch, Read, Write, Edit, Glob, Grep

rlat-refresh-facts — re-check external facts against the live world

A .rlat can carry external facts — claims fetched from the web for questions the corpus couldn't answer (the external_fill path). They were verified true when fetched. But the world moves on: "the latest version is X", "the current price is Y", "the record holder is Z" all go stale, and rlat refresh cannot catch it — refresh only re-syncs corpus files to their local sources; it never touches the web.

This skill is that missing check. It re-fetches each external fact's cited source and asks: does the source still say this? It runs free, on your reading of the re-fetched page — no Anthropic API key (the same agent-as-judge basis as rlat-contradictions and rlat-gap-scan).

It surfaces stale facts for review; it does not edit the corpus. Re-authoring from the fresh source or retiring the stale fact is a separate, user-approved step.

---

The loop

1. List the external facts + their sources (free, no fetch)

Find the corpus with Glob "*.rlat"; if several, prefer the project-named one or ask. Then enumerate what was fetched from outside:

rlat audit <km>.rlat --external --format json

Output is a list of {claim_id, state, content, source_urls}. If it's empty, the corpus has no external facts — report that and stop (nothing to re-check; a corpus-only .rlat is fine). Each external fact is backed by two or more cited URLs (the fill required cross-source agreement).

2. Re-fetch each source and judge freshness (you, free)

For each external fact, WebFetch each of its source_urls to get the page's current content. Then decide, from the re-fetched content alone:

Does the current content still support the claim — the same value/fact it was verified against? Wording may differ; additions that don't contradict are fine.

Classify each fact:

• fresh — the current sources still support it. No action.
• stale — the current sources contradict it, no longer state it, or have moved past it (a newer version/value/result). The fact is out of date — surface it.
• unknown — a source 404'd, was blocked, paywalled, or moved, so you could not read the current content. Report it as uncheckable, never as stale (a failed fetch is not evidence the fact changed).

A fact is stale only if the re-fetched content actively disagrees. If you can't re-fetch enough of its sources to tell, it's unknown, not stale.

3. Report (surface, don't fix)

List each stale fact: the claim, the URL(s) that no longer support it, and what the source now says instead (the likely new value). List unknown facts separately as "couldn't re-check". Leave fresh ones unmentioned unless asked.

If the user wants to act on a stale fact, that's the create/fix tier (one fact at a time, on their go): re-fetch and re-author the corrected fact from the now- current sources (it lands low-trust, cited, and earns its place over time), or retire the stale one. Do not silently edit the corpus as part of this scan.

---

Honest limits (state them, never paper over)

• A failed fetch is not staleness. 404 / blocked / paywalled / moved → unknown, not stale. Never flag a fact stale because you couldn't read its source.
• Your freshness verdict is a probability. A page can support a fact in wording you misread, or drop it for unrelated layout reasons. Surface stale facts as candidates to re-check, not settled retirements — which is why this scan never auto-edits.
• Only what was fetched can be re-checked. This re-verifies external facts; it says nothing about corpus drift (use rlat refresh) or facts no one fetched.
• Cost is your fetch budget. One WebFetch per cited URL; a fact with many sources costs more. Re-check the facts most likely to age (versions, prices, "current"/"latest" anything) first.

Why this is separate from rlat refresh

rlat refresh re-syncs corpus passages to their local source files — it fixes corpus drift. An external fact has no local source to re-sync; its truth is on the web. This skill is the only thing that re-checks it. (Corpus drift was deliberately demoted in the self-audit for exactly this reason: refresh already fixes it; world- staleness is the freshness problem that actually needs a re-fetch.)

Reference

• src/resonance_lattice/store/external_freshness.py — the library mechanism: external_claims (the free enumeration this skill drives) and recheck_external_freshness (the metered/automated equivalent of the judge step).
• rlat audit <km> --external — the enumeration surface.
• src/resonance_lattice/curator/external_fill.py — how external facts get ADDED (the cross-source-agreement fill this skill later re-checks).
• rlat-contradictions / rlat-gap-scan — the sibling free self-audit skills.