# Listening Party prompt finder: verified method and results

Published public evidence: 15 October 2026

## The product question

Listening Party has a fixed deck of 40 original music prompts. A host may know the feeling or
moment they want, without knowing which card will start the right conversation. The browser-local
finder searches that existing deck. It does not generate cards or send the request to a generative
model.

The released interface makes the result easier to inspect. It shows matched words, the original
card source and an explanation when common words are doing the matching. A host can clear a search,
change category or close the finder, then explicitly choose and deal a card.

## Decision

BM25 remains the default. The compact GloVe comparison remains experimental. Three fixed
evidence-policy thresholds were evaluated against the existing retrieval baselines. Every threshold
removed unsupported suggestions in the development set by rejecting every eligible answerable
request. No candidate qualified, so the selection is locked disabled and no candidate was evaluated
on holdout.

## Frozen evaluation

The evaluation has 74 synthetic, agent-authored English cases: 40 development cases in 20 intent
families and 34 holdout cases in 17 separately declared families. It is an engineering diagnostic,
not user research, a customer study or a production benchmark. Full blinding is not claimed.

Ordinary hit at 3 means an expected card appeared among the first three results for the 16 ordinary
cases. Answerable coverage means any suggestion appeared for one of 22 eligible answerable cases;
it is not a correctness score. An unsupported suggestion means the finder recommended any card for
one of eight no-answer cases.

| Stage and method                | Ordinary hit at 3 | Answerable coverage | Unsupported suggestions |
| ------------------------------- | ----------------: | ------------------: | ----------------------: |
| Development BM25 comparator     |             13/16 |               20/22 |                     6/8 |
| Development semantic comparator |             14/16 |               22/22 |                     8/8 |
| Development evidence-050        |              0/16 |                0/22 |                     0/8 |
| Development evidence-067        |              0/16 |                0/22 |                     0/8 |
| Development evidence-080        |              0/16 |                0/22 |                     0/8 |
| Holdout BM25 baseline           |             12/16 |               22/22 |                     6/8 |
| Holdout semantic baseline       |             13/16 |               22/22 |                     8/8 |

For the 22 eligible answerable holdout cases, BM25 hit at 3 was 18/22 and the semantic comparator
was 19/22. The small difference in synthetic cases does not establish a general advantage or
justify a promotion. The holdout ran baselines only; there is no policy-candidate holdout result.

The older 34-case v1 evidence remains separately available in the parent public folder. Its
denominators and results are historical and are not combined with this evaluation.

## Verified live release

The public site is https://listeningparty.online. On 15 October, a live Chromium run passed 12
recorded checks: matched-word, source and common-word explanations; search/category/clear/close
recovery; keyboard traversal at 390, 768 and 1440 pixels; real-browser optional-model checksum;
and an explicit host deal with a separate guest readback. The release also corrected the optional
model's slow-download deadline from 20 to 90 seconds. Two regression cases failed before the
correction and passed after it: a 58-second simulated download still left keyword results usable,
and a stalled download reached the 90-second bound with retry.

The full game test suite then passed 126 tests, with two opt-in provider tests skipped. Typecheck,
changed file lint and a fresh Cloudflare build also passed. These checks do not cover microphone
input, transcription, paid providers, physical-device behavior or multi-device audio.

## Provenance

- Public site: https://listeningparty.online
- Deployed application source: f497d77fc5c21f0718db8fe90d8f46e710e2dd66
- Release-record revision: abd4329732725481fcc158a14c04bd82ef1cfa62
- Production version verified in the release record: be6216e1-e428-4a9d-b50c-2886fc60d6e1
- Frozen evaluation execution source: 55e7bf73cd7498d0e0e0f8ce93d14146918d0fcd
- Execution-freeze SHA-256: f7a059213e7235e23b361fa1ece8f60538350f9ddeb6bacc1d654c20da662717

This document deliberately excludes raw internal logs, browser storage, credentials, email, local
paths and private repository URLs.