AN INDEPENDENT RECOMMENDATION LAB

Beyond
the obvious.

Your taste. A little unexpected.
A cinematic feed with a mind of its own.

RESEARCH PROTOTYPE · SYNTHETIC CATALOG

Shape your universe

Tonight’s moodDiscovery level 35%

FamiliarUnexpected

Your next obsession ↘

NEW STORIES · ZERO DEMO HISTORY

Discover a movie before the crowd.

Synopsis-based matching, including precomputed semantic embeddings.

Original Midnight Orchard fictional poster

Original Radiant Transit fictional poster

Original A Morning Elsewhere fictional poster

Fictional movies. Original AI-generated fictional poster artwork. No streaming or trailers. Sponsored predictions are simulated.

THE TRADE-OFF LAB

Not every click
is a good decision.

Compare policies on the same candidates. All ad outcomes below are simulated expected values, not measured clicks or revenue.

Minimum sponsored relevance 0.45

Measured offline experiments

Server mode reads artifacts/metrics.json. These are synthetic recommendation test results, separate from policy simulation.

Real MovieLens benchmark · v2

Validation-only selection. Recommendation relevance proxy, not real advertising outcomes. Full dataset cohort is separate from the earlier 60-user subset.

Use server mode to load the benchmark report.

Candidate retrieval→Hybrid scoring→Diversity reranking→Sponsor constraints

THE COLD-START LAB

New stories.
No history required.

Compare genre, synopsis words and semantic meaning before a fictional movie has any audience interactions. Real-data measurements below are a separate experiment.

Assume you enjoyedMatching method

Loading offline metadata…

Try your own new movie description

Fictional titleGenreSynopsis

Session-only input. No upload or trained-catalog changes. Arbitrary descriptions use genre/TF-IDF fallback; no semantic model or API runs in this static demo.

Loading real-data report…

Temporal sensitivity check · smaller cohort

Corpus: CMU Movie Summary Corpus, Bamman, O’Connor & Smith (2013), CC BY-SA 3.0 US. No raw corpus, MovieLens records or model weights are distributed in this app. Cold-start means unseen by the interaction fit, not a current-release feed. Semantic model: MiniLM ONNX, Apache-2.0.

BUILT ON IDEAS, NOT BUZZWORDS

A small system.
Serious questions.

The question

How do relevance, discovery and sponsored exposure interact? Cinemaflex is an implementation and experiment platform—not a claim of a novel algorithm.

Full source notes, evaluation protocol, model card and limitations are included in the ZIP.