PI Index: first results
Can a compact Shelf context stand in for everything a person has listened to? We asked five frontier models to predict what real people will play next week, once from their full 28-day listening history and once from agent memory plus Shelf context. With memory and Shelf, four of the five models beat their accuracy on the full history and the fifth came within half a point, from roughly half the input.
Shelf helps agents know you and act for you, on your terms.
With agent memory and a compact Shelf context, Opus 5.5 predicted next week's listening as accurately as it did from the person's full listening history, from about half the input. You choose the sources, shape what Shelf knows, and control what each agent can access.
29.88% vs 29.41%
Opus 5.5 with agent memory and Shelf context, against the same model reading the full 28-day history.
Half the tokens
Shelf context alone matched the full history, 29.41% each, in a preregistered test, with 9,163 input tokens a call against 18,564.
+3.29 points
What Shelf context adds on top of agent memory for Opus 5.5, 95% interval +1.65 to +5.18. 13 people better, none worse.
What we found
A compact Shelf context does the work of the full history
Our preregistered test asked whether Shelf's short context keeps the value of a person's full 28-day listening history. It did, within the 2-point margin we set in advance: Opus 5.5 scored 29.41% with each, and Shelf context used about half the input tokens. Add agent memory and Opus 5.5 reached 29.88%. Across the five models, memory plus Shelf context beat the full history for four, by up to 6.59 points for Gemini 3.8 Flash, and came within 0.47 points for GPT-6.1 Sol.
Shelf adds to what memory already knows
Our other preregistered test asked whether Shelf adds anything once an agent has its own memory of the person. For Opus 5.5 it did: +3.29 points, with a 95% interval of +1.65 to +5.18. Thirteen people got better predictions and none got worse. Every other model scored higher with Shelf added too, as description only.
Context is what moves the score
With nothing about the person, every model scored between 6.82% and 9.41%, below the 12.24% that global popularity reaches on its own. With context, every model rose by 18 to 22 points. The model matters far less than what it knows about you, and Shelf delivers that knowledge in a fraction of the space.
Sharing is the next thing to get right
Predicting which apps a person will share is a harder problem than predicting what they will play next. The best configuration, GPT-6.1 Sol with memory and Shelf, ranked the apps a person went on to share at 61.2% accuracy, and the models sit within a few points of each other and of simple rules such as global popularity (59.8%). That leaves the most room to improve of anything we measured. The next sharing bench measures the models directly, so we can track how quickly agents get better at knowing what you would want to share.
Results
Predict · Consumption
Which artists will they listen to next week? Round 2.
85 held-out people, prediction week of 2026-09-19. The full history is every listen in the 28 days, newest first. Shelf context is Shelf's fact sheet, placed after the agent's memory when both are given.
Create · Sharing
Which of this week's apps will they share? Round 3.
156 cases over two retrospective weeks, 2026-09-26 and 2026-10-03.
Methodology
- Round 2 (Predict, consumption). Cutoff 2026-09-19. Each model saw 28 days of a person's listening history, or a summary of it, and named five artists from a 200-artist catalog for the following week. 150 people were selected by a seeded hash; 37 were held for development and 113 for the test, of whom 85 were eligible. Scoring policy pi-index-scoring-v2, frozen before the test labels were opened, with a paired bootstrap.
- Agent memory is at most 12 notes about the person, written once by Opus 5.5 from their raw history and cut to 900 bytes. Shelf is Shelf's fact sheet: the person's top 10 artists with recent play counts and recency, plus repeat and new-artist shares and variety, each stated against the cohort.
- Preregistered tests: Shelf added to memory, a superiority test passed at the one-sided 2.5% level; Shelf's fact sheet against raw history, a non-inferiority test with a 2-point margin, cleared by 0.12 points.
- Round 3 (Create, sharing). Retrospective cutoffs 2026-09-26 and 2026-10-03: both weeks had passed before the run, but no model had predicted these people and none of their test labels had been opened. Our research had read both weeks in aggregate beforehand, and measurements from the 2026-10-03 pool informed the choice of the second preregistered comparison. Cases are excluded when the models could not be shown the app screenshots (Android 13 or earlier, or an unknown Android version), or when the person shared none or all of the week's apps: 185 and 228 cases at the two cutoffs. Each model ranked the apps run in the week; ranking accuracy is the share of pairs, one shared app and one not, in which the shared app ranks higher. Random is 50%. The round's report also gives Precision@1, the share of people who shared the model's top-ranked app, as description. Scoring policy pi-index-scoring-v3, cutoff-stratified bootstrap, 10,000 draws.
- All models ran through OpenRouter with pinned endpoints and settings. Predictions were sealed before labels opened, and each test week was scored once.