Contents

Solution · Project 2: The Ratchet

This works through Project 2's fixed five-session table using five weight triples, in the order they were tried. Your own weight choices will differ — check the shape of your .jsonl files against this one, not the specific numbers.

Scoring five attempts

Recall the scoring rule from the project: a candidate's score is its weighted genre match, plus its weighted inverse-recency term, minus its weighted skip rate. Accuracy is how many of the five sessions the higher-scoring candidate matches the listener's actual choice.

Attempt	Weights (genre, recency, skip)	Correct sessions	Accuracy
1	(1, 1, 1)	s1, s3, s5	0.6
2	(2, 1, 1)	s1, s3, s5	0.6
3	(1, 3, 1)	s1, s3, s5	0.6
4	(1, 1, 3)	s1, s2, s3, s5	0.8
5	(1, 1, 5)	s1, s2, s5	0.6

Worked arithmetic for attempt 4, the one that actually improves on the baseline — session by session:

s1: Harbor Lights scores 1 + 1/3 − 3(0.10) = 1.033; Faded Static scores 0 + 1/2 − 3(0.40) = −0.70. Predicted Harbor Lights — correct.
s2: Quiet Radio scores 1 + 1/11 − 3(0.20) = 0.491; Neon Drift scores 1 + 1/2 − 3(0.60) = −0.30. Predicted Quiet Radio — correct. (This is the session every other attempt got wrong.)
s3: Neon Drift scores 1 + 1/4 − 3(0.30) = 0.35; Slow Static scores 0 + 1/3 − 3(0.10) = 0.033. Predicted Neon Drift — correct.
s4: Faded Static scores 0 + 1/6 − 3(0.20) = −0.433; Harbor Lights scores 1 + 1/21 − 3(0.05) = 0.898. Predicted Harbor Lights, but the listener actually chose Faded Static — incorrect.
s5: Slow Static scores 1 + 1/2 − 3(0.15) = 1.05; Quiet Radio scores 1 + 1/3 − 3(0.50) = −1.167. Predicted Slow Static — correct.

Four out of five: 0.8.

Applying the ratchet

Attempt	Accuracy	Current best before this attempt	Outcome
1	0.6	none	no baseline yet — automatic pass, becomes `h1`
2	0.6	0.6 (`h1`)	tie, not strictly greater — discarded, logged against `h1`
3	0.6	0.6 (`h1`)	tie, not strictly greater — discarded, logged against `h1`
4	0.8	0.6 (`h1`)	strictly greater — becomes `h2`, beats `h1`
5	0.6	0.8 (`h2`)	worse than the current best — discarded, logged against `h2`

Notice attempts 2 and 3 are both logged against h1, but attempt 5 is logged against h2 — the comparison point moved once attempt 4 raised the bar. A discard log that always pointed at h1 would be lying about what the current best was at the time each attempt ran.

Final file contents

durable-history.jsonl:

jsonl

{"id":"h1","weights":{"w_genre":1,"w_recency":1,"w_skip":1},"accuracy":0.6,"beats":null}
{"id":"h2","weights":{"w_genre":1,"w_recency":1,"w_skip":3},"accuracy":0.8,"beats":"h1"}

discarded.jsonl:

jsonl

{"id":"d1","weights":{"w_genre":2,"w_recency":1,"w_skip":1},"accuracy":0.6,"lost_to":"h1","reason":"tied the current best, not strictly greater"}
{"id":"d2","weights":{"w_genre":1,"w_recency":3,"w_skip":1},"accuracy":0.6,"lost_to":"h1","reason":"tied the current best, not strictly greater"}
{"id":"d3","weights":{"w_genre":1,"w_recency":1,"w_skip":5},"accuracy":0.6,"lost_to":"h2","reason":"scored lower than the current best"}

Checking your own attempt

Does durable-history.jsonl only grow when an attempt's accuracy is a strictly higher number than the last entry's, with no exceptions for "close enough"?
Does every line in discarded.jsonl record which kept entry it lost to, not just a bare score?
If you re-ran your five attempts in a different order, would the ratchet rule still produce a two-entry durable-history.jsonl — or does your logic depend on the order you happened to try them in?