sawz/last30days-skill
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(skill): replace narrative verdict row with evidence-triggered prose beat

Browse Source 
Review showed the "Setting the narrative?" verdict compared across
abstraction levels: a homepage tagline is deliberately broad
("financial infrastructure" covers a chargebacks thread), so
tagline-vs-thread alignment verdicts are unfalsifiable and carry no
information. The signal now ships as PROSE in the entity's narrative
section, fires only when the month's evidence directly bears on the
pitch (supports a specific claim, cuts against one, or is squarely
about the pitched ground), and stays SILENT when the pulse is
orthogonal - omission over a manufactured connection. Claims are
tested at matched altitude (specific claim vs specific thread) and
stay windowed (no trend verbs one 30-day window can't support). The
positioning fetch step survives unchanged and now also grounds the
"What it is" row and brand-noise rejection. All scope gating (people
never, ownerless topics excluded, no pitch from memory) carries over.
This commit is contained in:
Trevin Chow
2026-06-09 17:14:31 -07:00
parent 57860aff1c
commit c1ca1a4e9d
4 changed files with 19 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files
CHANGELOG.md
+1 -1
View File
@@ -9,7 +9,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
### Added
- **Narrative lens for company / product / service topics.** Comparison tables gain a `Setting the narrative?` axis that judges whether each entity's community conversation is actually about what the entity *pitches* (its first-party positioning) or about something else — pricing, rivals, an incident, a ToS change. A new mandatory research step captures each entity's current stated positioning from first-party sources (homepage, docs, pricing) rather than from memory, and single-entity company runs get a `narrative-check` synthesis beat surfacing the same signal. The mismatch is the point: companies usually don't control their own conversation. The lens is scoped to entities with an identifiable first party (companies, products, services): people always get N/A — even founders whose companies would qualify — as do events, abstract concepts, and ownerless topics like Bitcoin, and verdicts require positioning fetched during the run, never from memory.
- **First-party positioning research + pitch-vs-pulse synthesis (company / product / service topics).** A new mandatory research step captures each entity's current stated positioning from first-party sources (homepage, docs, pricing) rather than from memory. The fetched pitch grounds `What it is` descriptions (entities described as they pitch themselves today), helps reject unrelated brand-name noise, and feeds an evidence-triggered prose beat: when the month's conversation directly supports a specific claim, cuts against one, or is squarely about the pitched ground, the synthesis says so anchored to the top thread — and stays silent when the pulse is orthogonal to the pitch, because a manufactured connection is worse than omission. Claims are tested at matched altitude (specific claims against specific threads; broad taglines are never graded against individual items), and statements stay windowed to the 30 days — no trend verdicts. Scoped to entities with an identifiable first party: people are always excluded (even founders whose companies qualify), as are events, abstract concepts, and ownerless topics like Bitcoin; the beat requires positioning fetched during the run, never from memory.
### Fixed