/deepen — escalate a source record's reading depth

Take one source record (from 50-resources/sources/) one level up the processing ladder:

captured → processed → connected (a third, reconstruction-grade pass stays connected; synthesized is set only when an area note actually incorporates the source — that's synthesis work, not deepening).

Resolve the argument to a note path (search 50-resources/sources/ by slug/title/URL if given loosely). If it's a bookmark, offer to promote it first: create the source record per /know's schema, add promoted_to: "[[<source record>]]" to the bookmark, then deepen the record.

Contract (binding)

1. Never rewrite existing prose. Fill empty sections and append; if a filled section needs revision, propose it to the user instead of editing. The human's words outrank yours.
2. Own words only. Quotes belong in the evidence package, cited from the record.
3. This builds retrieval, not understanding. After deepening, offer the user one retrieval question drawn from the extracted claims ("From memory: what does X claim about Y?"). Optional, never blocking.
4. One note per invocation. Deepening is expensive by design; batching it recreates shallow capture.

Pass 2 — captured → processed

Actually read the source (fetch fully; for a repo, read README + key entry points; for a paper, the full text if reachable — retain it as an evidence package if you had to fetch something ephemeral).

Video sources: read the full transcript in the evidence package and view the retained frames — slides often carry claims the speech only gestures at; extract slide content (numbers, architecture, tables) into claims. When the transcript references a visual moment that has no retained frame ("as you can see here…", "this slide shows…"), re-fetch the video to the scratchpad (yt-dlp -f "bv*[height<=1440]/bv*"), grab the exact frame (ffmpeg -ss <m:ss> -i <video> -frames:v 1 -q:v 2), view it, and add it to frames/<mm-ss>-<hint>.png if it earns its place. Delete the video afterward. Claims from a video cite their timestamp: … (12:34).

Fill these body sections (create any that are missing, in this order):

• ## Source and evidence quality — who made this, with what incentive; is it experience report, opinion, benchmark, peer-reviewed; how would a claim here be checked.
• ## Claims and evidence — one bullet per claim, own words, each ending with ^claim-<idea-slug> (idea-named, unique in file, never renamed once written). 3–8 claims; fewer honest claims beat exhaustive paraphrase. Note per claim what evidence backs it (assertion / example / data / proof).
• ## Limitations or counterarguments — where it's thin, what it ignores, who would disagree.

Gate before advancing state. Before setting processing: processed, check every bullet in ## Claims and evidence for a trailing ^claim-<slug> anchor. A claim bullet without one isn't addressable — /synthesize can't cite it at block granularity, and /review's promotion count can't see it — so it isn't a completed pass 2 regardless of how much prose surrounds it. Add the missing anchors before advancing state; don't advance state and leave them for later (a 2026-08-16 audit found 30 of 33 processed+ records in this vault missing anchors entirely, all bulk-migrated before this gate existed — don't grow that number further).

Then set processing: processed, bump updated.

Connect pass — processed → connected

Search the vault for related material: existing area notes (40-areas/), concepts (50-resources/concepts/), and other source records' claims. Fill:

• ## Connections — typed inline fields only, from the closed vocabulary:

  supports:: [[Note or Note#^claim-slug]] — optional qualifier
  extends:: [[...]]
  contradicts:: [[...]] — on <what, specifically>
  special-case-of:: [[...]]
  supersedes:: [[...]]

  Write relations on this note pointing at the older/more canonical note; don't hand-maintain inverses. Prefer claim-block targets when the relation is about one claim. A few labeled edges beat many bare links — if you can't type it, it isn't a connection worth recording.

• ## Synthesis usage — leave empty unless an area note already cites this source; then link it.

If 00-meta/questions.md exists and a claim answers an open question, add the ID to answers: [] and note it in your report. Then set processing: connected, bump updated.

Pass 3 — reconstruction (stays connected)

Only on explicit request. Append a ## Reconstruction section:

• Repo/system: how it actually works (data flow, key mechanisms) and how you would have built it — where your design would differ and why.
• Paper: work the argument — restate the method, check whether the conclusions follow, identify the load-bearing assumption.

This pass often produces new claims (add them to ## Claims and evidence with block IDs) and new open questions (add to 00-meta/questions.md if it exists; otherwise list them in the section).

Report

• Append one row to 00-meta/log.md: | YYYY-MM-DD | ingest | /deepen: <title> → <new processing state>, <n> claims extracted |.
• Summarize: new state, claims extracted (with slugs), relations added, questions touched.
• If the source turned out shallower than its capture suggested, say so — downgrading expectations is a valid outcome.