WAYPOINTEVIDENCE

DOES IT ACTUALLY BEAT
SIMILARITY SEARCH?

Every recommender claims good results. This page is the receipt: our planner and a semantic search baseline, run over the same catalogue, scored against learning paths a human expert wrote by hand.

  • 65 SKILLS
  • 126 RESOURCES
  • 9 OF 10 PATHS SCORED
  • EMBEDDING BASELINE

HEADLINE

PREREQUISITE VIOLATIONS

0.0%vs 49.0%

Stops the learner was not ready for when they arrived.

SKILL GAP CLOSED

93.0%vs 64.7%

How much of the distance to the destination the route covers.

REDUNDANCY

16.8%vs 36.3%

Effort spent on skills the learner already had.

AGREEMENT WITH EXPERT

0.392vs 0.168

nDCG against the resources a human chose.

ORDERING CORRELATION

0.459vs 0.111

Kendall tau against the order a human put them in.

HOURS PRESCRIBED

71hvs 232h

Average study time each route asks for.

THE SAME LEARNER, ROUTED TWO WAYS

A violation rate is a number you have to trust. This is the same figure with the routes attached: pick a learner and see which steps each approach puts in front of them before they are ready, and exactly which skill they are missing.

“I want to learn data analysis from zero and I don't know where to start.”ALREADY KNOWS: NOTHING RELEVANT

WAYPOINT

Planned over the prerequisite graph

ALL 7 STEPS READY

  1. 01

    Kaggle Learn: Intro to SQL

    Kaggle · 3h

  2. 02

    Kaggle Learn: Python

    Kaggle · 5h

  3. 03

    Kaggle Learn: Pandas

    Kaggle · 4h

  4. 04

    Kaggle Learn: Data Visualization

    Kaggle · 4h

  5. 05

    NumPy: The Absolute Basics for Beginners

    NumPy · 4h

  6. 06

    StatQuest with Josh Starmer

    StatQuest · 15h

  7. 07

    freeCodeCamp: Scientific Computing with Python

    freeCodeCamp · 25h

SIMILARITY SEARCH

Ranked by closeness to the goal text

5 OF 10 STEPS NOT READY

  1. 01

    freeCodeCamp: Data Analysis with Python

    freeCodeCamp · 30h

    NOT READY — needs Python Basics level 3

  2. 02

    Python Data Science Handbook

    Jake VanderPlas · 45h

  3. 03

    Google Data Analytics Professional Certificate

    Google / Coursera · 180h

    NOT READY — needs Python Data Structures level 2

  4. 04

    Pandas Getting Started Tutorials

    pandas · 6h

    NOT READY — needs Python Basics level 2

  5. 05

    Kaggle Learn: Pandas

    Kaggle · 4h

    NOT READY — needs Inferential Statistics level 2

  6. 06

    Kaggle Learn: Time Series

    Kaggle · 5h

  7. 07

    Kaggle Learn: Data Visualization

    Kaggle · 4h

  8. 08

    Kaggle Learn: Intro to SQL

    Kaggle · 3h

    NOT READY — needs Python Data Structures level 2

  9. 09

    NumPy: The Absolute Basics for Beginners

    NumPy · 4h

  10. 10

    Kaggle Learn: Python

    Kaggle · 5h

Both columns are real output over the same catalogue and the same embeddings. The difference is not the search — it is that one of them knows what has to come first.

HOW TO READ THIS

The ordering correlation is the one worth staring at. A value of 0.111 means the baseline's sequence has no relationship to the expert's — it retrieves reasonable material and then hands it over in an order nobody could learn from. Ordering is the whole problem, and similarity search has no concept of it.

The prerequisite violation rate is zero by construction rather than by tuning: the graph gates which resources are generated as candidates at each step, so an infeasible stop is never considered in the first place.

MEASURED 30 AUGUST 2026 · BOTH SYSTEMS SEARCHED THE SAME CATALOGUE WITH THE SAME EMBEDDINGS

PLOT YOUR OWN ROUTE →