A reading inbox is not a publication queue
In short: My WhatsApp self-chat held 168 unique links; only four had cleared analysis. Capture is inventory, not evidence. Put verification between save and publish.
1/ My WhatsApp self-chat is deliberately easy to fill: in the September 11 AndyMental audit, 203 available messages included 170 link-bearing messages and 168 canonical unique links. Only four sources had completed analysis and queue mapping. That is one local workflow result, not a benchmark for anyone else's inbox, and it tells me the publication queue must begin after verification, not after saving.
2/ Frictionless capture is still good design. WhatsApp's own help describes Message yourself as a place to keep track of notes and messages, with messages that work like a regular chat. I want the save gesture to stay that simple. I just give it zero publication authority.
3/ A saved link proves intent to revisit, nothing more. It may be a wrapper around the real source, a repeated link with tracking attached, an outdated claim, or a private URL that should never leave the inbox. Capture creates inventory. It does not create evidence.
4/ Safe ingestion and credible research are separate controls. My local reader opens its SQLite source in read-only mode, which SQLite documents as mode=ro, and also enables PRAGMA query_only. SQLite says that pragma blocks common data-changing statements but does not make the connection truly read-only. Even a careful reader cannot certify the links it reads.
5/ Deduplication has to happen before belief. The 170-to-168 difference in my audit is not proof that URL canonicalisation alone removed two duplicates; I did not run that controlled comparison. The practical rule is narrower: normalise URLs, then cluster different wrapper posts that still point to one underlying claim. Reposts do not multiply evidence.
6/ My verification rule starts at the original. For a repository, inspect the exact code and current documentation. For a video, inspect captions or a transcript plus representative playback. For an article, check the creator, date, full text and load-bearing references. A preview card is a research lead, never a citation.
7/ Limitations belong in the record, not in someone's memory. “Transcript unavailable”, “repository inspected but not executed”, and “number reported by the vendor” are useful fields because they stop a later writer from upgrading uncertainty into fact. Four reviewed items out of 168 unique links is not a failure. It is an honest queue boundary.
8/ The publication queue should begin after verification, not after capture. My state line is captured -> clustered -> verified or blocked -> drafted. Each arrow leaves a receipt, and a blocked source stays visible without becoming copy. That is the original claim here: a reading inbox needs quarantine more than it needs automation.
This also extends my earlier rule that AI research needs evidence lineage. Steal the small version: add a verified field and a limitation field to whatever holds your saved links. Let nothing reach the writing queue while either is blank. The extra state costs less than researching the same weak claim twice, and it keeps a good capture habit from becoming an accidental endorsement machine.
Capture can stay frictionless; publication cannot. Put a verified source and its limitation between every saved link and the writing queue.
