Skip to content

hypothesis.md

Latest commit

 

History

History
82 lines (54 loc) · 5.74 KB

hypothesis.md

File metadata and controls

82 lines (54 loc) · 5.74 KB
Raw

Hypotheses — cortex-hit-encode

CLOSED 2026-09-14. Two frozen rounds run (v0 previews n=14 pairs; v2 full tracks n=11 pairs). Call rule 1 fired both times: primary test not above the permutation null → H0; H1 not supported. dz = −0.21 (p=0.45) and dz = −0.055 (p=0.84), point estimates on the wrong side of zero in both rounds. H2 never reached, H3 untestable by construction, H4 never triggered. Final call + assumptions relied upon: conclusion.md. Text below is the pre-registration as written before any encode — kept unedited.

Date: 2026-09-14
Instrument: che (vendored VideoCortex encode path) + TRIBE v2 (encoding, average subject)
Sibling craft work (out of band): artist-twin insight/hit_patterns.py (lyric structure only)

Motivation

Commercial “hits” are labeled after social markets act. Encoding models predict how an average cortex responds to a stimulus. The interesting question is whether hit labels, after careful matching, still separate predicted maps — i.e. whether there is a measurable encoding fingerprint associated with hit-labeled songs in a genre — without claiming that fingerprint causes charts.

Primary hypothesis

H1 (encode separation).
Among genre-, language-, and era-matched stimuli, tracks meeting a pre-registered hit label criterion produce systematically different predicted average-subject cortical response patterns under TRIBE v2 / che than matched non-hits — especially in auditory / temporal / language-adjacent ROIs and in temporal dynamics of regional energy — after controlling for loudness, duration, and (where available) vocal presence / cut-rate confounds.

v0 operational gate (same name as plan + protocol): mean_auditory_roi_energy — mean |x| over a frozen Destrieux auditory ROI set, averaged across TRs; permutation test of the hit − non-hit difference.

Operationalization of “hit label” is fixed in the cohort sheet before analysis (examples: national chart peak ≤ N, certification, or streams threshold). One definition per pilot; no peeking.

Null and alternatives

H0 (no separation). After matching and confound regression, hit vs non-hit labels do not separate encode summaries better than a permutation null (label shuffle within matched pairs or within cohort).

H2 (confound). Any remaining separation after the standard v0 loudness-normalize (all stimuli, same LUFS target) collapses when stimuli are duration-matched more tightly, matched to a stricter LUFS residual, or reduced to instrumental stems — i.e. the effect is leftover production energy / density, not a hit-label fingerprint. H2 is a sensitivity, not the Phase 2 prep step.

H3 (fame proxy). Within-artist hit vs deep-cut pairs show little or no separation once artist and era are held fixed — suggesting catalog fame / production era, not a general hit circuit. H3 is a secondary slice; it can be supported even if H1 holds on the broader matched cohort.

H4 (craft bridge — deferred, non-causal). Even if H1 fails, encode diffs between high vs low lyric hit-shaped craft scores (artist-twin Insight) may still separate maps. That would support craft priors, still not chart forecasts. Out of scope for pilot v0 unless H1 is inconclusive and time remains.

Predictions if H1 is supported

  1. Higher mean_auditory_roi_energy (mean |x| in the frozen auditory ROI set) for hit-labeled tracks vs matched non-hits.
  2. Secondary: multivariate distance on the ROI vector, or more sustained time-courses, may also separate — not the v0 gate.
  3. Effect survives the loudness + duration covariates; shrinks but does not vanish under H2-style controls (or we accept H2).

How to call H0 / H1 / H2 / H3

Same rules as experiment-plan.md and protocol.md:

  1. Primary test not above the permutation null → H0. Stop product mythology.
  2. Primary test above null, but H2 controls kill the difference → H2. Do not market as a hit fingerprint.
  3. Primary test above null, H2 survives, within-artist slice null → H1 on the matched cohort; H3 as a fame/era caveat.
  4. Primary + H2 survive and within-artist also separates → H1; H3 not supported.
  5. Encode failures / TRIBE lag misuse / mixed lag modes → invalid run. Fix protocol; do not interpret.

Falsifiers

  • No separation after the pre-registered primary test → H0.
  • Separation vanishes under H2 controls → H2.
  • Within-artist pairs null → H3 (caveat or alternative, per the call rules above).
  • Invalid encodes → no hypothesis call.

What success does not mean

Supporting H1 does not authorize:

  • decoding songs from desired activation maps;
  • claiming high probability of commercial success (hit_probability, “likely a hit,” “% chance,” Billboard / chart forecast, “ensure a hit”);
  • mind-reading or “this is what someone is thinking”;
  • hard-gating artist-twin generation on cortex scores.

It authorizes: “interesting encoding difference in this cohort under this instrument.”

Language: non-goals.md.

Pre-registration discipline

Complete the Phase 1 freeze in protocol.md before the first real encode batch:

  1. Freeze cohort inclusion rules and one hit_definition_id in stimuli/manifests/cohort-v0.csv (copy from the example).
  2. Freeze the same primary endpoint + test in experiment-plan.md and protocol.md: mean_auditory_roi_energy, permutation of the hit − non-hit difference.
  3. Freeze the Destrieux auditory ROI label list (confirm names against the che overlay atlas).
  4. Record che / vendored VideoCortex commit + TRIBE weight hashes in each runs/.../receipt.json (see runs/receipt.example.json).