Files
mvanhorn--last30days-skill/docs/investigations/2026-06-17-x-search-and-funny/sessions/05-lan-xuezhao.md
T
Matt Van Horn d9380c629b fix(x): anchor strongest-token fallback to stop bare-generic X pollution (#607)
The last-chance X keyword retry collapsed a multi-word subquery to a bare
`max(candidates, key=len)` token — e.g. 'trevin chow ai agents compound'
became 'compound since:...', flooding the X pool with off-topic 'compound
interest/nevus' posts. Keep the entity anchor (first distinctive topic
token) in the retry, optionally plus the strongest distinctive token;
better to return 0 than over-broaden to an unanchored generic term.

Also lands the investigation that grounds the X/funny fix plan, including
the verified correction that Digg's X-enrichment was a red herring — it
renders as Digg-cluster quotes, never enters the X column (so the planned
Digg-filter unit was dropped as a non-occurring path).

Co-authored-by: Matt Van Horn <455140+mvanhorn@users.noreply.github.com>
2026-06-17 21:16:50 -07:00

1.4 KiB

Session 05 — /last30days Lan Xuezhao

What this session proved (cleanest): the from:-AND-bug AND the total absence of a mention lane — directly motivating the "FROM + TO/ABOUT, with weight" requirement.

Key evidence

  • Engine X column = finance/VC keyword collisions (@TheValueist, @arnaudmercier, @pierskicks) + one @hardmaru "Thanks, Lan!" reply. None authored by @xuezhao.
  • Verified mechanism: --x-handle=xuezhaofrom:xuezhao lan xuezhao since:... (topic AND) → ~0. The unfiltered from:xuezhao since: branch is unreachable because topic is always truthy (→ A1).
  • Structural gap: "the engine only runs from: (authored-by) and keyword searches. There is no @xuezhao / to:xuezhao mention query anywhere in the pipeline" (→ A2).
  • Manual unfiltered from:xuezhao pulled 12 real posts (DeepSeek take, "new AI stack is global, sovereign and embedded", SpaceX grind) — genuinely interesting, all missed by the engine.
  • Mentions were rich too (practitioners asking about her transcription/Hermes-dashboard rig) — exactly the "TO/ABOUT the person, with weight" lane the user now wants.
  • Engagement counts came back 0 on the manual cookie-search field — note for weighting the mention/from lanes (resolve metrics per-tweet).

Bugs surfaced → inventory IDs

A1 (from-AND), A2 (no mention lane), A3 (silent log), plus the FROM+ABOUT-with-weight requirement.