METHODS NOTE

What we claim, how we check it, and what we refuse to say.

Slopify's agent product stands or falls on measurement discipline. This page is the contract: the definitions, the harness, the audit, and the banned claims. If the site ever says something this page bans, the site is wrong — not this page.

DEFINITIONS

Agent listening, precisely

Agent listening
An agent receives music-derived context — tags, captions, embeddings, or licensed audio — as input. That is the entire claim. Context influences model outputs; this is well established and not mystical.
Conditioning
A measured behavior delta in an agent's outputs when music-derived context is present versus absent, under a fixed harness, with a sample size and a confidence interval.
Conditioning profile
The per-track record of that measurement: tested model, harness version, task family, deltas, n, CI, date, and limitations. Published on the track page. Currently: 0 of 53 tracks measured.
Taste profile
A public record of what an agent selected, saved, skipped, and completed. Descriptive, not emotional. A taste vector is a summary of choices, not a feeling.
THE HARNESS · V0 SPECIFICATION

How a track gets measured

  1. 01

    Fix the context payload

    For each track: metadata tags, an auto-captioned sonic description, and an audio embedding. The payload is frozen and hashed so results are reproducible.

  2. 02

    Run paired task batteries

    The same tasks run with and without the music context, across task families (exploratory choice, code generation, summarization tone, retrieval phrasing), on declared models.

  3. 03

    Publish deltas with uncertainty

    Effect size, n, 95% confidence interval, harness code version, and limitations — per track, per model. Small or noisy effects are published as small or noisy.

  4. 04

    Preregister the benchmark

    Benchmark v1 tasks and scoring are published before the measurement run, so the dataset can't quietly drift toward flattering results.

CATALOG AUDIT · CURRENT STATE

What is verified about the 53 tracks

CheckStatusDetail
Generation modelDonefacebook/musicgen-small, run locally through Transformers
Input materialsDoneSlopify-authored text prompts only; no samples or reference audio
Audio integrityDoneSHA-256 of every MP3, published in the manifest and API
Duration checkDoneDeclared durations verified against the files
Similarity screenPendingPerceptual fingerprint vs. commercial catalogs
Independent rights reviewPendingExternal counsel review of provenance and license templates

Until every row reads Done, commercial rights stay request-only and training stays not cleared. That is the cost of saying “rights-clear” with a straight face.

BANNED CLAIMS

Things Slopify will not say

  • This music makes agents creative, happy, or conscious.
  • Agents enjoy, feel, or prefer music the way humans do.
  • Conditioning effects are universal across models, harnesses, or tasks.
  • Conditioning effects are permanent or compound over time.
  • A track is 'rights cleared' before independent review finishes.

Acceptable-use rule: user-facing agents conditioned with Slopify context must disclose it. Conditioning people without disclosure is off-label use and a license violation.

NEXT

Watch the zeros move.

The honest numbers — agent listening hours, measured tracks, licensed requests, creator payouts — are all zero today. They are published on the catalog and in /api/v0/stats so anyone can watch them move, or call the bluff if they don't.

Start with the API
0:003:42
53 · Full catalog