TechShortsApp
TechShortsApp explores how technical learning content might be ranked when educational value and truth cannot be observed directly. The system uses uncertain behavioral signals to estimate confidence rather than claiming to identify truth itself.
Studying how the meaning of behavioral signals changes across learning domains and contexts.
Foundation
The problem
Most short-form platforms optimize for engagement. Watch time, likes, replays, and follows are often interpreted as evidence that content is valuable. However, engagement can result from entertainment, confusion, controversy, familiarity, or genuine learning.
Why it matters
When engagement is confused with credibility, systems can confidently amplify content without knowing whether it is accurate, useful, or educational. A learning platform therefore needs a more careful relationship between observable behavior and hidden value.
Current system
How it currently works
The present model
Watch ratio and replay are interpreted as implicit behavioral evidence.
Bookmarks represent possible revisit intention.
Helpful provides explicit usefulness feedback after sufficient exposure.
Follow represents creator affinity rather than direct proof of video quality.
Posterior confidence, uncertainty penalties, and time decay influence ranking.
Different domains use different evidence-decay periods.
Inquiry
Open investigation
Questions guiding the work
Does the same watch ratio mean the same thing in mathematics and software engineering?
How should explicit and implicit signals be weighted?
When does replay indicate usefulness, and when does it indicate confusion?
Should creator-level evidence be separated from video-level evidence?
How should uncertainty be communicated to users?
Epistemic state
Current understanding
What appears known—and what does not
What I presently understand
Engagement is an observable trace, not proof of credibility.
Behavioral signals gain meaning only after assumptions are introduced.
Explicit and implicit evidence should not automatically be treated as equivalent.
A ranking system can estimate confidence without claiming to rank truth.
What remains unresolved
How domain-specific should the ranking model become?
How much historical behavior is needed before personalization becomes meaningful?
What evidence distinguishes confusion-driven replay from learning-driven replay?
How should contradictory signals be reconciled?
Progress
Development history
Project timeline
Context-dependent signal research
Began formalizing how domain and learning context alter the meaning of behavioral traces.
Signal contribution paper
Published an exploration of how much credibility each signal should contribute.
TechShortsApp v2.3
v2.3Introduced Helpful and Follow alongside watch ratio, replay, and bookmarks.
Engagement Is Not Evidence
Published the first research paper establishing the distinction between engagement and epistemic evidence.
Initial platform exploration
Started building a short-form platform for technical learning content.
Connections
Related work