# HumanFB Initial Music Evaluation

## Blinded deep-track listening, contemporaneous voice capture, interpretation testing, and personal-preference prediction

**Project:** HumanFB<br>
**Document date:** August 15, 2026<br>
**Status:** Initial evaluation protocol<br>
**Stimulus domain:** 1980s and 1990s heavy metal and hard rock

## Executive summary

The initial HumanFB evaluation uses 100 comparatively lesser-known album tracks by prominent heavy-metal and hard-rock artists. The tracks come from the 1980s and 1990s and are presented in shuffled order. Artist, song, and album names are concealed in the working manifest with a simple monoalphabetic substitution cipher so that casually seeing the list does not create expectations before listening.

This is a useful first test because music is temporally rich, emotionally legible, easy to replay, and available without constructing a new physical apparatus. A song can evoke several reactions at identifiable moments, while contemporaneous speech can refer naturally to features such as a riff, vocal entrance, transition, timbre, groove, solo, production choice, or change in intensity. The domain therefore exercises the core HumanFB loop: stimulus presentation, low-friction voice capture, temporal anchoring, AI interpretation, user-specific profile updates, and prediction on later stimuli.

The central evaluation is not whether an AI can summarize a memo after the fact. It is whether the complete system can preserve what was heard and said, interpret the speech into useful and correct dimensions, and learn enough about this individual to predict reactions to tracks that have not yet been heard in the study. Internet commentary about the same deep tracks can provide a valuable evaluator-blind comparison corpus after each response is frozen, but it is contextual evidence rather than ground truth about the user's preferences.

The prepared stimulus manifest is [`humanfb_80s_90s_metal_hard_rock_blinded_100.txt`](humanfb_80s_90s_metal_hard_rock_blinded_100.txt). It contains exactly 100 unique tracks, evenly divided between the two decades. Its SHA-256 digest is `6851ec1bb64b695b1a3c7e07d7a3cd7982f093142c865579bb2dfc48c4d7532f`.

## 1. Evaluation objectives

The study should answer five separable questions.

1. **Capture:** Can HumanFB reliably associate a voice memo with the exact track, playback interval, and listening context without requiring the user to navigate a form?
2. **Interpretation:** Can it convert free speech into a faithful, editable response vector while preserving the memo and transcript as primary evidence?
3. **Temporal attribution:** Can it distinguish a reaction to the whole track from a reaction to a particular entrance, section, transition, or sound?
4. **Personal learning:** Can earlier exposures improve prediction of the user's later liking, aversion, replay desire, and descriptive reactions before those later reactions are observed?
5. **Downstream usefulness:** Can the resulting profile rank or search for additional music that the user prefers more often than simple baselines?

These questions should not be collapsed into one score. A transcript may be accurate while its interpretation is poor. An interpretation may be faithful while a preference model fails to generalize. A model may predict ratings while failing to explain which musical attributes matter. The evaluation records each layer independently.

## 2. Stimulus-set design

### 2.1 Selection principle

The set consists of deep album tracks rather than canonical hits. The artist or album may be familiar, but the selected song should be less likely to carry a strong pre-existing reputation for the listener. That choice creates a useful middle ground: there is enough public metadata and discussion for independent analysis, while the evaluator is less likely to know the expected reaction in advance.

The current manifest uses these constraints:

| Property | Current design |
|---|---|
| Total unique tracks | 100 |
| Decades | 50 from the 1980s; 50 from the 1990s |
| Domain | Heavy metal and hard rock, interpreted broadly enough to include adjacent metal and hard-rock styles |
| Artist profile | Predominantly prominent or historically established artists |
| Track profile | Album tracks selected to avoid a list dominated by signature hits |
| Order | Deterministically shuffled rather than chronological |
| Visible manifest | Encoded artist, title, and album; numerical trial ID remains readable |

This is a curated evaluation pool, not a claim that every track is unknown to every listener. Familiarity is measured after exposure and used as a covariate. A recognized track remains useful evidence but should not count as a clean blind trial.

### 2.2 Blinding mechanism

The manifest uses the following substitution:

```text
PLAINTEXT : ABCDEFGHIJKLMNOPQRSTUVWXYZ
CIPHERTEXT: QWERTYUIOPASDFGHJKLZXCVBNM
```

The cipher is intentionally not cryptographic protection. Its purpose is to prevent accidental recognition when the user or operator glances at the queue, a log, or a trial identifier. A playback component or facilitator decodes only the next row needed for presentation. The participant should not manually browse or decode later rows.

The cipher key is included in the file header as requested, which makes the artifact self-contained and machine-readable. Stronger concealment, if later needed, should be implemented as access control or a player-side opaque manifest rather than by inventing a more complicated home-grown cipher.

### 2.3 Exposure identity

Each trial should retain both a stable `stimulus_id` for the recording and a separate `exposure_id` for the particular listening episode. The stimulus record should identify the exact release or audio master when possible. A remaster, live version, radio edit, and original album version are not assumed to be interchangeable.

## 3. Hypotheses and observable outcomes

The study begins with explicit hypotheses so that interesting transcripts do not substitute for a system evaluation.

| Hypothesis | Primary observation | Comparison |
|---|---|---|
| Natural voice capture is usable | Completed, correctly attached memos | Button-driven or manually labeled fallback |
| AI interpretation is faithful | User acceptance or correction of extracted axes | Transcript-only and generic sentiment baselines |
| Context improves modeling | Held-out prediction error | Same model without mood/context inputs |
| Temporal anchors add information | Correctly resolved section-level references | Whole-track-only representation |
| Personal history predicts later reactions | Held-out liking and replay prediction | Mean rating, genre prior, and popularity/metadata baselines |
| The learned profile improves discovery | Preference for profile-selected new tracks | Familiarity/popularity or random deep-track selection |

The most important distinction is between **interpreting the current reaction** and **predicting a future reaction**. A same-trial voice memo contains the answer and therefore cannot demonstrate preference prediction for that track. Predictive evaluation freezes the model before the held-out track is played and scores the prediction against the later response.

## 4. Minimal system under test

The first evaluation requires only a narrow vertical slice of HumanFB.

```text
encoded manifest
      ↓
one-row-at-a-time resolver and audio playback
      ↓
playback clock + stimulus identity + session context
      ↓
local voice trigger or simple automatic post-track recording window
      ↓
audio memo + transcript + temporal anchors
      ↓
versioned interpretation vector
      ↓
explicit confirmation/correction and held-out prediction report
```

The listening system does not need a complete cross-domain personal profile, biometric stack, or recommendation service. It does need immutable trial identities, trustworthy timing, raw memo retention, versioned interpretations, and a clean separation between training and held-out trials.

For the earliest run, an automatic audio cue near the end of playback or immediately afterward is acceptable. A later run should compare this prompted mode with locally detected natural evaluative speech. The prompt itself is logged because it may influence the response.

## 5. Listening protocol

### 5.1 Session structure

Use ten sessions of ten new tracks rather than one long sequence. This limits fatigue and makes session-level mood and listening conditions measurable. Session order follows the shuffled manifest unless a technical failure requires rescheduling a trial.

At the beginning of a session, record a lightweight baseline:

- energy or fatigue;
- mood valence;
- stress or agitation;
- desire for stimulation versus calm;
- current familiarity with the artist if revealed only after the trial;
- playback device, volume, location, and notable environmental conditions.

The baseline should take less than a minute. A burdensome questionnaire would change the natural interaction the system is intended to test.

### 5.2 Trial sequence

1. Resolve only the next encoded row and assign a new `exposure_id`.
2. Record the source identity, audio version, start time, player clock, device state, and session baseline reference.
3. Play the track without displaying artist, title, album art, lyrics, recommendations, or comments.
4. Permit spontaneous voice triggers during playback and preserve several seconds of playback pre-roll around each trigger.
5. At track end, open a short voice-memo window automatically if no sufficient memo was captured.
6. Collect compact explicit targets: overall liking, replay desire, familiarity, confidence, and whether the user wants to add anything.
7. Freeze the raw evidence and first-pass interpretation before retrieving public commentary or updating the model used to predict later trials.

The user may speak during the song, immediately afterward, or both. During-song speech provides strong temporal evidence but changes the listening experience. Post-track speech is less intrusive but depends more on memory. HumanFB records the capture mode so that the two can be compared rather than treated as equivalent.

### 5.3 Optional repeat trials

After the 100 unique-track sequence is established, select approximately 10–15 tracks for unannounced repeat exposure in later sessions. Repeats measure response stability, context dependence, memory effects, and whether the interpretation layer produces similar abstractions from different wording. They are additional exposures, not additional unique songs in the manifest.

## 6. Voice memo and interpretation record

The user should be allowed to speak naturally. A prompt can suggest topics without forcing a questionnaire:

> What happened for you in that track? Mention anything you liked, disliked, noticed changing, wanted more or less of, or would want to hear again.

The first-pass interpretation should include both general and music-specific dimensions.

| Layer | Example fields |
|---|---|
| Literal evidence | Audio object, transcript, word timing, speaker confidence |
| Whole-track judgment | Valence, liking, dislike, replay desire, familiarity, certainty |
| Affective response | Energy, tension, pleasure, irritation, curiosity, excitement, boredom |
| Perceptual description | Heaviness, brightness, density, spaciousness, roughness, polish, dynamics |
| Musical focus | Riff, rhythm, vocals, solo, arrangement, production, melody, transition |
| Temporal anchor | Playback interval, section label, confidence, quoted referring phrase |
| Comparative intent | More like this, less like this, preferred alternative, exception |
| Context | Session baseline, volume, device, interruption, fatigue, prior-track carryover |

Every derived field carries a model version, confidence, and evidence link. The user can correct an interpretation without overwriting the original transcript or earlier model output. Corrections are especially valuable evaluation data because they reveal systematic interpretation errors.

## 7. Analysis plan

### 7.1 Capture and alignment

Report memo completion rate, trigger success, false triggers, missing pre-roll, speaker/playback confusion, and alignment error. Review a sample of references such as “that change,” “the riff just before the solo,” or “when the vocal came back” against the playback timeline.

### 7.2 Interpretation accuracy

Ask the user to accept or correct a compact interpretation for a stratified sample of trials. Measure agreement separately for polarity, scalar axes, mentioned musical objects, and temporal anchors. Do not use transcript sentiment alone as the reference; a memo can praise one feature while rejecting the song overall.

### 7.3 Preference prediction

Before each held-out track is played, generate a prediction using only earlier evidence and stimulus-side features available before exposure. Score overall liking, replay desire, likely positive and negative attributes, and uncertainty. Compare against simple baselines:

- the user's mean rating so far;
- artist, decade, or broad-style averages;
- a stimulus-feature model without voice-derived profile parameters;
- a popularity-oriented or generic recommendation baseline.

Use chronological or session-aware splits so that later trials are genuinely predicted from earlier ones. Report calibration as well as rank correlation or error: a system that is uncertain should be allowed to say so.

### 7.4 Discovery test

After the profile has sufficient evidence, use it to rank an unseen candidate pool or compile a targeted search description. Compare a small profile-selected set against a matched baseline set without telling the user which route selected each item. The practical outcome is whether the learned profile finds more tracks the user wants to replay, not merely whether it produces plausible adjectives.

## 8. Use of independent internet commentary

Independent commentary on deep tracks is useful because it may describe perceptual or compositional features unknown to the evaluator. It can test whether the user's memo and the interpretation system detect features that other listeners also noticed. It can also supply vocabulary, section descriptions, and disagreements that help separate broadly audible properties from highly personal response.

The commentary corpus should be retrieved only after the trial response and first interpretation are frozen. Otherwise, song descriptions, reputation, and fan consensus can leak into the participant's response or the evaluator's coding.

Commentary is best represented as a set of source-linked observations rather than a single consensus label. For each passage, retain source, date, quoted or paraphrased proposition, target track or section, and extraction confidence. Compare it with the user's record on at least three levels:

1. **Descriptive overlap:** Did both identify the same riff, transition, production feature, vocal quality, or structural event?
2. **Evaluative agreement:** Did both like or dislike that feature? This is informative but not expected.
3. **Personal divergence:** Did the user respond idiosyncratically to a feature others noticed, or notice something the commentary corpus did not?

Public commentary is not ground truth about subjective quality and should not be used to train against the held-out user's answer before scoring. Its strongest role is post-hoc triangulation and vocabulary expansion.

## 9. Trial data contract

A minimal exposure record should contain:

```text
trial_id
stimulus_id and exact audio-version identity
exposure_id
manifest row and randomized position
session_id and baseline-context reference
playback start, pause, seek, completion, and end events
voice trigger, memo start/end, and playback-relative timing
raw memo object and transcript artifact
first-pass interpretation artifact and model version
user corrections or confirmations
explicit liking, replay desire, familiarity, and confidence
pre-exposure model prediction, uncertainty, and model snapshot
post-trial profile update reference
independent-commentary artifacts retrieved after freezing
```

The pre-exposure prediction and post-trial interpretation must be distinct artifacts. This prevents the system from appearing predictive by incorporating the response it is supposed to predict.

## 10. Principal validity threats

| Threat | Treatment |
|---|---|
| User recognizes the song | Record familiarity immediately; analyze blind and recognized trials separately |
| Cipher is decoded or queue is browsed | Resolve one row at a time; treat the text cipher as glance protection only |
| Streaming UI reveals metadata | Use a metadata-suppressed playback view or facilitator-controlled display |
| Loudness dominates preference | Normalize playback policy and record actual volume/device state |
| Fatigue or prior-track contrast | Use short sessions, baseline state, and prior-trial identifiers |
| Speech disrupts listening | Distinguish during-track from post-track capture and compare them |
| Model learns artist identity rather than musical preference | Evaluate artist-held-out subsets and feature-based predictions where possible |
| Commentary leaks before response | Freeze evidence and first interpretation before retrieval |
| Same-trial memo is used as a predictor | Freeze predictions before exposure and maintain temporal data lineage |
| Deep-cut obscurity is inconsistent | Treat recognition as measured data rather than assuming perfect unfamiliarity |

## 11. Initial exit criteria

The first evaluation is successful enough to justify a broader prototype when all of the following are true:

- At least 95% of completed playbacks have correctly attached raw memos and trial identities.
- Timing is sufficient to resolve most explicit section references without manual reconstruction.
- The user accepts the majority of high-confidence interpretation fields, and corrections can be stored without loss of provenance.
- Held-out preference predictions improve materially over the user's running mean or another declared simple baseline.
- Confidence is directionally calibrated: low-confidence predictions fail more often than high-confidence predictions.
- A blinded discovery comparison shows that profile-selected tracks produce higher replay desire than a matched baseline often enough to warrant replication.
- The entire result can be regenerated from immutable exposure records, model versions, and manifest identity.

The exact statistical threshold should be declared before formal scoring once the number of sessions, repeats, and candidate discovery tracks is fixed. The first run remains a single-person system test, so its main purpose is to validate the measurement and learning loop rather than claim population-level generality.

## Conclusion

The blinded heavy-metal and hard-rock track set is a strong initial HumanFB evaluation because it combines controlled digital presentation with rich, temporally anchored subjective response. It can test capture, interpretation, context modeling, personal prediction, and recommendation without requiring a physical robot or a large participant cohort.

The decisive design rule is to freeze each prediction before exposure and freeze each user response before consulting outside commentary. That ordering separates genuine prediction from retrospective interpretation and separates the user's reaction from public reputation. If the system can preserve those boundaries while improving held-out track selection, the result will demonstrate a reusable personal-feedback architecture rather than merely a polished voice-journaling interface.