Skip to content

plan.md

Latest commit

 

History

History
103 lines (79 loc) · 4.14 KB

plan.md

File metadata and controls

103 lines (79 loc) · 4.14 KB

XFEED Implementation Plan

Feed Quality Improvements (Jan 2026)

Features Added

1. Serendipity / Exploration Mode

  • Includes high-quality tweets from unknown/new authors at configurable rate
  • Bounded by: minimum quality threshold, author cooldown, diversity sampling
  • Exploration candidates are sampled AFTER base exclusion filtering
  • Maximum N exploration posts per refresh (default: 10% of feed)
  • Authors marked [EXPLORE] in reason string for transparency

2. Reasoning Quality Scoring

  • LLM prompt now includes factor-based scoring instructions
  • Boost factors: mechanism, tradeoffs, evidence, uncertainty, assumptions
  • Penalty factors: vague, unsourced, rhetorical, overconfident
  • Each tweet's reason includes factor breakdown: "Great insight [+mechanism, +evidence]"

3. Contrarian-but-Serious Allowance

  • New factor dissent_rigorous flags contrarian content with strong reasoning
  • Dissent bonus only applied when rigor score meets threshold
  • Bonus capped to prevent over-rewarding contrarian views
  • Low-rigor dissent treated normally (no boost, may be penalized)

Files Touched

File Changes
src/xfeed/filter.py New SYSTEM_PROMPT with factor scoring, exploration sampling logic, dissent bonus calculation, _build_explanation() helper
src/xfeed/config.py Added config knobs: exploration_rate, exploration_min_quality, exploration_cooldown_hours, dissent_min_rigor, dissent_bonus_cap, reasoning_boost_max, reasoning_penalty_max
~/.xfeed/objectives.md Added sections: Serendipity, Reasoning Quality Scoring, Contrarian-but-Serious, Interpretation Rules
objectives.md (repo) Mirror of user objectives for reference
ROADMAP.md Added "In Progress" section documenting these features
tests/test_filter.py Unit tests for new scoring logic

Config Knobs Added

# Exploration / Serendipity
exploration_rate: 0.1          # % of feed for exploration candidates
exploration_min_quality: 7     # Min score for unknown authors
exploration_cooldown_hours: 24 # Hours before repeating new author

# Reasoning quality
reasoning_boost_max: 2         # Max boost for quality factors
reasoning_penalty_max: 2       # Max penalty for quality factors

# Contrarian handling
dissent_min_rigor: 6           # Min rigor for dissent bonus
dissent_bonus_cap: 2           # Max boost for contrarian content

TODOs / Future Work

  • Persist exploration author cache across sessions (currently in-memory)
  • Add CLI command to view/clear exploration cache
  • Track author reputation over time (requires SQLite)
  • Add --no-exploration flag to disable exploration mode
  • Expose factor breakdown in mosaic UI (currently only in reason string)
  • Add metrics: % of feed that is exploration, avg rigor score, etc.

Test Plan

Run tests with:

source .venv/bin/activate
pytest tests/ -v

Test Cases

  1. Exclusion rules still apply in exploration

    • Rage bait from unknown author → excluded (score 0-1)
    • Engagement farming from unknown author → excluded

  2. Contrarian without rigor NOT boosted

    • Contrarian take with no evidence → no dissent bonus
    • Contrarian take with vague factor → may be penalized

  3. Mechanism/evidence posts are boosted

    • Post with mechanism factor → explanation shows [+mechanism]
    • Post with evidence factor → explanation shows [+evidence]

  4. Explanation strings include factors

    • All scored posts have reason with factor breakdown
    • Format: "Base reason [+factor1, -factor2]"

  5. Exploration sampling is deterministic

    • Same seed → same exploration candidates selected
    • Candidates interspersed at regular intervals

  6. Author cooldown works

    • Author shown as exploration → marked in cache
    • Same author within cooldown → not shown again as exploration

Architecture Notes

  • Scoring is still done by Claude Haiku (no local ML)
  • Factor extraction is LLM-based, not heuristic
  • Exploration sampling uses seeded RNG for reproducibility
  • Author cache is in-memory (resets on process restart)
  • O(n) complexity over candidates maintained