Who gets seen?#

Platform Architecture and the Elo Hierarchy

SLOP1276 · Week 02 · Alex’s story

Fictional Alex has a place in the toy ranking queue. The exposure record is missing. A position on a list is suddenly a poor audience report.

Investigate: What proves the profile was seen?

What you will be able to do#

Trace a profile through retrieval, filtering, ordering and exposure.

Separate a published claim from an assumed model.

Calculate an Elo update and a rate with a named denominator.

Explain how an exposure rule can sustain a hierarchy.

How this pack fits the week#

60 minutes · Monday’s selected lecture slides.

90 minutes · Thursday’s four practical blocks.

About 210 minutes · independent preparation, remaining examples and consolidation.

Timed selections and tasks.

The complete 180-minute route is a resource guide; its break is for a full sitting.

Alex returns from week 1#

Fictional adult CS student; board games and terrible puns.

Friday 5–7 pm · $20 total · bus travel.

Wants to meet someone, with the possibility of a relationship.

Today: how could a candidate reach Alex’s screen?

A profile has a route#

An ordered candidate can remain unseen if the viewer stops scrolling.

Three candidates, one screen#

Which proposed outings pass Alex’s supplied time and budget checks?

Filtering changes the winner#

Retrieve: P1, P2, P3.

Filter: remove P2; 8 pm is outside Alex’s window.

Order by our invented score: P1 (8), then P3 (6).

If only one profile is shown, P1 gets exposure. Bus feasibility and mutual interest remain unknown.

Draw the boundary · 8 min#

Sketch the five stages for Alex and the three candidates.

Place availability, ordering score, shown on screen and interest in Alex.

Label each as supplied, assumed, observed or hidden.

Pair discussion or solo annotation; record one uncertainty.

Boundary debrief#

Supplied: Alex’s time and budget.

Assumed: the candidates, their scores and the one-profile display limit.

Observed in a log: which profile was actually shown.

Hidden: motivation and mutual interest. An eligible candidate is still a person who can decline.

Instrument the right event#

A toy log needs viewer ID, profile ID, time, exposure and explicit choice.

Count an exposure when the profile is shown, not merely retrieved.

Deduplicate repeated event IDs; define the observation window.

An absent reply is not a recorded dislike.

Read a disclosure like an engineer#

Record claim → primary source → scope → unknown.

Keep the page title, URL and date checked.

“Uses recent activity” names an input. It does not reveal a weight.

The next three slides report company descriptions checked on 6 September 2026.

Tinder: retire the folklore#

Tinder’s matching explainer says it no longer relies on Elo.

It describes activity, profile information and Likes/Nopes as signals.

Supported: this is Tinder’s published description.

Unknown here: production weights and a numerical swipe penalty.

Hinge: a recommendation feature#

Most Compatible describes mutual preferences, recent activity and patterns of liking.

Scope: one named recommendation feature.

Unknown here: its scoring equation and individual connection probabilities.

The feature name is not a probability estimate.

Bumble: inputs, not coefficients#

Discover documentation says its highlighted recommendations use profile information and previous matches.

Scope: the “Recommended for you” selection within Discover.

Unknown here: relative weights or an Elo-style update.

Do not extend a feature description to every screen in the app.

Audit three claims · 6 min#

A: “Tinder’s explainer says it no longer relies on Elo.”

B: “Our simulator subtracts eight points after a chosen event.”

C: “All three apps apply an eight-point mass-swipe penalty.”

Classify: published disclosure, model assumption or unsupported claim. Add the evidence needed.

Claim audit: worked answer#

A · disclosure. Cite Tinder’s page and the date checked.

B · assumption. Cite our rule and a reproducible trace.

C · unsupported. These sources provide no such number.

An unknown coefficient stays unknown. Plausibility is not a source.

Why teach Elo here?#

A small rating model makes expectation, surprise and feedback inspectable.

We can change one rule and reproduce every update.

Our binary outcomes are simulated comparison results.

They are not measurements of a person’s attractiveness or an app’s recovered algorithm.

Five symbols, one update#

R: current rating · Q: comparison rating.

S: simulated result, 0 or 1 · K: positive update size.

E: model expectation, between 0 and 1.

Expectation follows the rating gap#

Equal ratings mean E = 0.5, not “a 50% chance of romance.”

Work an equal-rating win#

R = Q = 1200 · K = 32 · S = 1.

E = 1 / (1 + 10⁰) = 0.5.

Change = 32 × (1 − 0.5) = +16.

New rating = 1216. Keep the expectation in the trace.

Change one input at a time#

Reset to R = Q = 1200 for each case.

Loss, K = 32: 1200 + 32(0 − 0.5) = 1184.

Win, K = 64: 1200 + 64(1 − 0.5) = 1232.

K changes responsiveness. A larger movement does not make the model more accurate.

Surprise gets a larger update#

R = 1200, Q = 1600, K = 32.

E = 1 / (1 + 10¹) = 1/11 ≈ 0.0909.

Win: 1229.09. Loss: 1197.09.

Explain why the unexpected win moves the rating much more than the expected loss.

Updates have state#

Start at 1200. Win against 1200 → 1216.

Then lose against 1200: recompute E using 1216, not 1200.

E ≈ 0.5230 → final rating ≈ 1199.26.

A win then a loss need not cancel. Round for display, not between steps.

Implement and trace · 12 min#

Write update(R, Q, S, K) using the two equations.

Reject non-finite ratings, non-positive K and S outside 1.

Run an equal win, an equal loss and the two-step sequence.

Save inputs, E and output. Pseudocode plus a hand trace is a valid starting point.

Test against known answers#

Reset cases: 1216 · 1184 · K = 64 win 1232.

Unequal ratings: win against 1600 → 1229.09.

Sequential case: win then loss against 1200 → 1199.26.

K = 0 must fail. Compare unrounded calculations within a tolerance; the notes include runnable assertions.

A penalty is a separate hypothesis#

Assume a simulator rule subtracts 8 after a chosen event.

Equal-rating win: 1216 − 8 = 1208.

Keep rating_after_elo and assumed_penalty as separate fields.

You changed your implementation. You have learned no platform’s hidden penalty.

What counts as a result?#

Here S is an explicit simulated binary result.

An unseen profile supplies no result.

An unanswered message is not automatically S = 0.

Before adapting a model, define its event, missing-data rule and observation window.

Pause for the full route · 10 minutes#

For an uninterrupted full-pack session; skip this slide in Monday’s selection.

Leave one question beside your trace.

When we return: two profiles receive the same number of likes.

Did they have the same opportunity to receive them?

Same likes, unequal exposure#

Calculate each like rate before drawing a conclusion.

The denominator changes the story#

Equal counts conceal a tenfold rate difference. Neither rate measures human worth.

Even the ranking can reverse#

Synthetic scenario: rank_reversal.

A: 80 likes / 1,000 exposures = 8%.

B: 20 likes / 100 exposures = 20%.

A leads on count; B leads on rate. Name the metric before announcing the winner.

A rate still has assumptions#

Were the audiences, time windows and exposure definitions comparable?

Could one viewer appear many times?

If ordering selected the audience, these rates do not isolate a profile’s effect.

Today we audit measurement. Week 4 adds bio experiment design.

The feedback hypothesis#

The arrow from likes to allocation creates the loop. Without that rule, the story changes.

A seed advantage can persist#

Add the seeds: 66 vs 44 likes. The share stays 60:40 in this expected-value calculation.

Change the allocator · 6 min#

Keep seeds 6 and 4, a 1,000-exposure budget and equal 10% response probabilities.

Allocate 500 exposures each, regardless of previous likes.

Compute expected new likes and cumulative totals.

Which assumption changed? Does the result prove a real app should use it?

Feedback debrief#

Equal exposure → 50 expected new likes each.

Cumulative likes become 56 and 54: A’s share ≈ 50.9%.

The response probabilities stayed equal; the allocation rule changed.

Our proportional loop preserves a seed advantage. It does not prove that every feedback loop widens the share gap.

Name the population#

“Reciprocal” means mutual eligibility under supplied constraints, not observed attraction or a match.

Two valid questions#

Among available profiles, what fraction is reciprocally eligible?

40 / 150 ≈ 26.7%.

Among active profiles, what fraction is reciprocally eligible?

40 / 300 ≈ 13.3%. Neither is a measured probability that Alex will get a date.

Audit the census · 8 min#

Open the synthetic CSV.

Sum North, South and East. Check 0 ≤ reciprocal ≤ available ≤ active in every row.

Add 50 active-only profiles to East; hold other fields fixed.

Recompute both fractions. Write which denominator changed and why.

Missing active profiles matter#

Available remains 150, so 40 / 150 ≈ 26.7% is unchanged. This is a scenario, not a discovered census error.

Stop bad rows before bad claims#

Try a row: 12 active, 15 available, 4 reciprocal.

Reject it: available cannot exceed active in this nested dataset.

Require whole, non-negative counts; detect duplicated zones.

If the denominator is zero, report undefined, not 0%.

Repair an audit paragraph · 8 min#

“Alex has an Elo of 1216, so Tinder considers Alex attractive. Profile A gets more likes, so its bio is better. There is a 13.3% chance of a date.”

Rewrite all three claims using today’s evidence.

Keep the useful numbers. Replace each unsupported inference with its actual meaning.

A defensible rewrite#

1216 is the output of our equal-rating simulated win.

In the reversal fixture, A has 80 likes, but B has the higher observed rate: 20% vs 8%.

13.3% is reciprocal eligibility among 300 active synthetic profiles.

No platform score, bio effect or date probability was measured.

Assemble the platform audit#

Boundary map with observed and hidden state.

Claim/source/unknown table with scoped citations.

Elo trace and tests, with assumptions separated.

Exposure and census log, with denominator definitions and sensitivity results.

Exit check · 3 min#

1. Which pipeline stage is missing when a candidate is ranked but never seen?

2. From equal ratings of 1200, what does K = 64 give for a simulated win?

3. Is reciprocal eligibility evidence of mutual attraction?

Answer before turning the page. Cite one assumption you will carry into week 3.

Carry the boundary forward#

Answers: exposure; 1232; no.

Alex can now distinguish ranking from exposure. The next unknown is the image itself. Week 3 opens the photo folder.

Tutorial and deliverable · Worked notes and sources