4-Layer Editorial System

Arena's core differentiator is a 4-layer editorial system that produces standalone, high-quality clips while other AI tools output incomplete thoughts. Every clip passes through all four layers.

Layer 1: Hook & Moment Detection

The engine scans the full transcript in sliding 2-minute windows, detecting approximately 40 "seeds" — compelling moments like claims, insights, hooks, stories, and conversational peaks. Seeds are deduplicated by time proximity and text similarity to produce ~25 raw candidate windows.

# Enable layer export to see Layer 1 output
arena process video.mp4 --export-layers

# Check: output/editorial_layers/layer1_moments.json

Layer 2: Thought Boundary Expansion

Raw clip windows are expanded outward to natural thought boundaries. Each candidate is analyzed for its premise (setup/context), claim (the core point), and resolution (conclusion or punchline). The boundaries are adjusted so clips capture the complete thought arc — premise to resolution.

This prevents the common AI clipping problem of clips that start or end mid-sentence or mid-thought.

Layer 3: Standalone Quality Gate

The strictest layer. A validation agent checks whether each clip can stand completely alone on a social media feed without surrounding context. It checks for:

• Unresolved pronouns ("it", "this", "that" without clear referents)
• Missing subject introductions
• Dangling references to earlier conversation
• Incomplete premises or resolutions
• Structural issues requiring significant adjustments

Each clip receives a standalone score from 0–10. Only clips scoring 7+ pass. Typical pass rate is around 7–10% of candidates.

Debugging rejections

# Export layer data
arena process video.mp4 --export-layers

# Check rejection reasons
cat output/editorial_layers/layer3_validated.json | \
  jq '.[] | select(.verdict=="REJECT") | {thought_id, rejection_reason, standalone_score}'

# Check pass rate
cat output/editorial_layers/layer3_validated.json | \
  jq '[.[] | select(.verdict=="PASS")] | length'

Common rejection reasons

Layer 4: Viral Packaging

Clips that pass Layer 3 are packaged with professional metadata for social media distribution:

• Magnetic, high-CTR titles
• Detailed keyword summaries
• Platform-specific hashtags
• Hook text for captions

Hybrid scoring

Arena combines two signals to rank clips:

• AI content score — GPT-based semantic analysis of content quality
• Audio energy score — RMS amplitude and spectral centroid analysis detecting enthusiastic delivery

Final score = ai_score × 0.7 + energy_score × 0.3. This ensures clips have both great content and dynamic delivery.

Semantic deduplication

Layer 4 also deduplicates clips using text-embedding-3-small cosine similarity. If multiple clips cover the same topic, the best variant per cluster is selected, preventing repetitive output.

Checkpointing

The CheckpointManager saves intermediate results between layers. If a run fails mid-pipeline (e.g., rate limit or network error), Arena resumes from the last completed layer instead of restarting. Job ID is a hash of the first 500 characters of the transcript.

Cost breakdown

Total for a 30-minute video: $0.15–0.25 with gpt-4o-mini, or $0.40–0.60 with gpt-4o.