Musical Intelligence:
A Glass-Box Architecture
for Computational Models of Music Listening

Developing a model of music listening requires specifying which aspects of sound enter a calculation, how much context it uses and how it depends on other computations. Musical Intelligence addresses this model-building problem.

Formal models make theoretical assumptions explicit and expose their consequences (Guest and Martin, 2021). Cognitive architectures extend this approach to the organisation of components and their interactions (Langley et al., 2009; Laird et al., 2017). MI applies this architectural perspective to computational hypotheses about music listening.

Audio libraries provide reusable sound-analysis algorithms (Bogdanov et al., 2013); probabilistic models specify musical expectation (Pearce and Wiggins, 2012); and TenseMusic predicts continuous tension from audio (Barchet et al., 2024). MI organises acoustic features, temporal context and listening-related calculations within a shared execution. Its primary contribution is an architecture in which these choices can be inspected together and individually revised.

R³ describes sound as named acoustic coordinates. T³ specifies temporal operations requested on those coordinates. C³ combines acoustic and temporal inputs into calculations expressing listening hypotheses. A fixed configuration gives deterministic outputs; saved inputs support reconstruction, and an altered request tests downstream consequences. Cognitive, regional and neuromodulator labels identify proposed interpretations to be assessed against observations.

Architectural evaluation reconstructs an output and follows a controlled change through its consumers. Empirical case studies then compare selected readouts with human observations. Fitted ridge readouts predict film tension with errors close to a fixed acoustic subset, while direct arousal ranks transfer across corpora. The engine stays fixed within each comparison; fitting estimates the observation mapping. Together, these demonstrations assess the architecture’s use for specifying hypotheses, examining computations and relating selected outputs to observations.

The architecture separates acoustic description, temporal context and listening-related calculations, with declared inputs and an explicit execution order. The current configuration has 97 acoustic channels and 1,337 temporal declarations that reduce to 644 unique addresses. Its 84 mechanisms yield 889 output dimensions, alongside 131 weighted readouts and eight selected recursive classes. These counts describe one implemented repertoire within the shared organisation.

Evaluation requires both an engine configuration and an observation protocol. Configuration $\theta$ specifies implementations, equations, requests, initialization and execution order; protocol $\pi$ specifies outputs, observations, aggregation and comparison. Their relationship is written as

$$\tilde{y}_\pi = O_\pi(F_\theta(a), c_\pi)$$

where $a$ is audio, $F_\theta$ the engine, $O_\pi$ the observation mapping and $c_\pi$ a covariate. When $O_\pi$ is fitted, prediction depends on engine and readout. Regression estimates the readout weights separately; an observation mapping need not forecast events.

A mechanism links a computational definition to a proposed interpretation. Its inputs and rule specify what it calculates; an observation protocol specifies how that calculation is evaluated. Depth roles determine processing order, while functional groups bring related proposals together for inspection. Anatomical and neuromodulator links project the outputs onto separately labelled channels with their own interpretation and measurement requirements.

The executor visits each mechanism once, in processing-depth order, retaining discovery order when depths coincide. A mechanism can use only the upstream outputs already available at that point; unavailable inputs invoke its implementation’s fallback. A listed connection therefore affects the calculation only if its source is available and its value is consumed. Both conditions depend on the execution.

R³ transforms each audio frame into 97 named channels across nine families, including consonance, energy, timbre, change, pitch, rhythm and harmony. The channels have no learned parameters or cross-family state. Published computations include Sethares roughness, Plomp–Levelt critical-band integration, Krumhansl–Kessler key profiles, Zwicker–Fastl sharpness and mel-frequency cepstral coefficients. Some channels share or derive inputs.

Each R³ coordinate also has a normalization rule and temporal support, ranging from an analysis frame to the full recording. These definitions accompany both the direct inputs to C³ and the series processed by T³.

The spectral consonance route operates on detected peaks without estimating a fundamental frequency. Peak separation and amplitude determine beating; ratio quality and a beating-dependent gate determine fusion. Pleasantness combines dissonance and fusion with fixed weights. The construction exposes each term and the algebraic link between inharmonicity and fusion. These coordinates cannot be interpreted as independent quantities. The route specifies how spectral relationships enter MI; it does not establish an advantage over pitch-based alternatives.

T³ makes the temporal context of a calculation explicit. Each request names an R³ feature, horizon, operator and window orientation, with no learned parameters. Each mechanism declares its requests before extraction. Matching requests share one stored array, while both callers remain identifiable. The current configuration has 1,337 declarations and 644 unique addresses, drawn from 32 horizons and 24 operators. Horizons span milliseconds to the scale of a long recording. Requests identify both the computed quantity and its consumers, connecting temporal assumptions to the equations that use them.

Window orientation selects preceding, following or surrounding samples. Recording length and boundary handling determine the available support. Changing an address can alter either the samples included or the operation applied. Temporal direction must be assessed across the whole calculation: even a past-oriented T³ window can depend on later audio when its R³ inputs use normalization over the complete recording.

C³ has 84 mechanisms and 889 output dimensions. Each declares temporal requests, a depth role and a functional group. Depth orders execution from relay to hub. Sensory, prediction, attention, memory, emotion, reward, motor and learning groups organise related proposals. These categories organise the model; biological interpretations require independent evidence.

Mechanisms combine acoustic channels, temporal descriptors and available upstream outputs. Parallel belief-update branches read the resulting dimensions. The 131 weighted readouts and eight selected recursive classes form distinct mappings and updates. Following an output through its executed equation and inputs, including fallbacks, exposes how these choices determine its value.

Regional activation mapping projects mechanism outputs onto 26 labelled regions through fixed links. Four further channels provide neuromodulator-inspired proxies. These mappings express biological hypotheses for comparison with measurements. They can be omitted from behavioural tests or evaluated through separate observation models. The regional links and proxy values are model definitions; independent neural or physiological data are required to assess their proposed biological meanings.

ICEM illustrates the acoustic, temporal and mechanism layers together. Nine of the 97 acoustic channels and eighteen temporal requests supply thirteen outputs, labelled from extraction through forecast. Observation, prior and precision terms update a belief, while regional and neuromodulator-proxy branches read the outputs in parallel. Gain and accumulation curves expose the equations’ responses; physiological interpretations remain to be tested.

Two checks follow. The first reconstructs an output from saved inputs. The second changes one temporal request while keeping other calculations fixed. Effects in the recorded example stay within the declared route, showing how an individual modelling choice can be revised and its consequences traced.

The Swan Lake example evaluates whether selected architectural dependencies can be reconstructed and revised in execution. Reconstruction recovers an output from saved inputs; changing one temporal request reveals which dimensions respond. Supplement S1 reports the recorded intervention and its protocol. Reconstruction records require controlled access.

ICEM.E0 was reconstructed from three temporal descriptors and two direct acoustic quantities saved for the 182.597-second Swan Lake recording. Within the pre-selected 30–50-second interval, its value at 40.0022 seconds is 0.25011. Parallel surprise and superior-temporal readouts follow the same execution. Recovering E0 from these inputs establishes the weighted computation behind its unexpectedness label. E0 is a proxy, not information in bits; relating it to experienced surprise requires observations of that human response.

Only ICEM’s centred spectral-flux weighted-mean request is changed, from a 23.2-millisecond horizon of four frames to a 46.4-millisecond horizon of eight frames. Every other input, equation and mapping stays fixed. Over 30–50 seconds, the maximum absolute difference in E0 is 0.003632 and the root-mean-square difference is 0.000643. Seventeen of 889 dimensions exceed the recorded tolerance. All 805 dimensions outside the declared reachable graph remain exactly equal. Observer and identity-adapter controls reproduce the reference outputs exactly. The change is therefore local to the declared route, and the instrumentation is not the source of the difference.

With the other terms fixed, non-expansive clipping and lower rectification give

$$|\Delta E_0| \leq 0.25|\Delta q|, \qquad |\Delta A_{\mathrm{STG}}| \leq 0.85|\Delta E_0|.$$

Of eight reachable mechanisms, only ICEM, PWUP and CHPI change; the illustrated consumers do not read the altered PWUP output. The bounds follow with other terms fixed. This traces local computational effects, without establishing whole-system stability or biological validity.

Equation inspection identifies inputs relevant to testing each interpretation. DAP.P1 combines inverse roughness, tonalness, sensory pleasantness, inverse spectral entropy and an upstream composite, without developmental-history inputs. MMP.P2 combines past warmth stability, a rhythmic summary and centred loudness, without listener-history inputs. AAC.F1 combines acoustic intermediates with a 15-second forward amplitude context; its two-second heart-rate forecast label is unvalidated. Source-linked restatements reproduce these outputs in the worked trace. For AAC, the four-term restatement agrees with all 38 saved TenseMusic excerpts to a maximum absolute error of $3.46 \times 10^{-8}$, conditional on the cached reconstruction inputs.

These definitions lead to specific empirical comparisons. A composite can be tested against its acoustic constituents on the same human outcome. A proposed forecast can be restricted to past information. Familiarity or developmental accounts can incorporate measured listener history. With waveform, configuration, initial state and consumed inputs fixed, a calculation cannot distinguish listener attributes that never enter it. The architecture makes such dependencies explicit, identifying both what the present acoustic representation can test and where an additional input is needed.

The worked checks establish reconstruction and local revision for the inspected pathways. An acoustic quantity, a temporal request and a receiving calculation can be followed within one execution, including whether an available output is consumed. This makes a modelling assumption accessible to a controlled change. The examples do not establish that every connection is necessary, that all mechanisms interact beneficially, or that the full repertoire is needed for a given task. Those questions require comparisons across additional pathways and configurations.

Reusable components and explicit computational graphs have precedents. Essentia supports audio-analysis networks (Bogdanov et al., 2013); the IPEM Toolbox provides perception-based music analysis; and AMT organises auditory models with demonstrations and reproduction experiments (Majdak et al., 2022). The PsyNeuLink environment connects psychological and neural model components, while ModECI specifies model structure and execution for exchange across environments (Gleeson et al., 2023). MI’s proposed contribution lies in its particular organisation of named acoustic coordinates, explicit temporal requests and listening-related calculations with identifiable consumers. The worked intervention demonstrates one use of that organisation. Comparative development effort, efficiency and reuse across new hypotheses remain to be evaluated.

Empirical comparisons relate selected MI outputs to public observations: interval means, excerpt ratings, within-piece trajectories, chord rows and participant-level encoding. Direct associations assess correspondence; fitted readouts assess prediction. Each retains its protocol, observation scale and execution identity.

The fixed pleasantness score, $0.6(1-\text{Sethares dissonance})+0.4(\text{Stumpf fusion})$, correlated with human ratings at Spearman $\rho=0.890$ across 13 harmonic-dyad means and $\rho=0.518$ across 617 chord rows. On the dyads, fusion reached $+0.813$ and roughness $-0.835$. Primary channels carried the expected sign on all nine references. Magnitude criteria were met in full on six and in part on three. Interval-subset ranking above $0.85$ passed on four. The same coefficients, applied to synthesized bell spectra, kept fusion positive across twelve register and synthesis conditions, from $+0.319$ to $+0.830$, highest for sustained tones at 880 Hz.

Fifteen C³ means supported prediction of held-out soundtrack tension. On 110 excerpts, grouped ridge cross-validation gave Pearson $r=0.671$, $R^2=0.431$ and RMSE 1.472; the fixed 15-descriptor acoustic subset gave $r=0.653$ and RMSE 1.485. Generic 970-descriptor models gave $r=0.619$. C³ had lower error on three of five folds against either comparator, a descriptive ordering. Appending C³ to generic descriptors did not reduce pooled error, leaving its incremental predictive value unresolved.

Across 110 film excerpts, all eight emotion pools had Bonferroni-significant associations. Acoustic residualization changed leading correlations: sadness $+0.741$ to $+0.430$, tenderness $+0.722$ to $+0.447$, tension $-0.683$ to $-0.504$. Seven of eight leading readouts retained $|\rho|\geq0.10$; every pool had one. This residual analysis does not test incremental prediction. Three channels recurred in the overlapping 360-excerpt pilot, documenting cross-set correspondence rather than independent replication.

After the MUSIFEAST pilot, channels selected from film ratings were evaluated without test-subset reselection. AAC.F1 arousal correlations were $+0.652$ across 360 film excerpts, $+0.691$ across 356 MUSIFEAST excerpts and $+0.770$ in its nested 65-excerpt subset; PUPF.U0 gave $+0.698$, $+0.676$ and $+0.735$. These overlapping rank comparisons show cross-corpus correspondence. Valence associations weakened: SRP.P1 declined from $+0.622$ to $+0.231$, and CMAT.S0 from $+0.653$ to $+0.172$.

Across 38 pieces rated by 30 listeners, historical AAC.F1 analysis gave Fisher-$z$ mean Spearman $\rho=+0.421$ with consensus tension, positive on 34 pieces. All 15 pool channels met the archived Bonferroni criterion. Physical clock alignment remains unresolved after normalized-time interpolation and lag selection. The listener-agreement value, $+0.386$, follows a different protocol and cannot rank performance. Reconstruction agreed across 38 cached excerpts to $3.46\times10^{-8}$, verifying arithmetic, with temporal alignment and autonomic forecasting untested.

Within $\pm5$ s of chills, MMP.P2 (familiarity label) rose in all seven stimuli: rank-biserial $r=+0.231$, Bonferroni $p=0.007$; its nine-stimulus and alternative seven-stimulus results were nonsignificant (S7).

On matched song-grouped folds, mean held-out $r$ was $+0.457$ for controls including human valence/arousal ratings, $+0.462$ with audio-native predictors and $+0.543$ with symbolic predictors. An additive-score regression reached $+0.615$ with preprocessing before splitting; the $+0.565$ ridge reference used different inputs, preventing a matched added-value claim. The proxy interaction was $-0.158$, with published $-0.124$ inside its interval; the audio-native interaction was $-0.060$, with a chord-row bootstrap interval excluding zero. Across rhythmic versions, HTP and ICEM had median $r=+0.993$ and $+0.828$. The last two values concern model stability; they do not measure pleasure in listeners.

Across seven chord-rating corpora, consonance correlations averaged $\rho=+0.408$, positive on six; the harmonic-timbre readout reached $+0.404$ across 25 intervals. Agreement with seven Hindustani raga pitch-class templates averaged $\rho=+0.565$, all positive. The raga comparison assesses template agreement, not listener responses or classification accuracy.

In 17 participants, 16 of 22 mechanism–region pairs met their BH criterion, versus 11 of 22 random controls; medians were $+0.162$ and $+0.058$. Separate BH families do not test this difference. Four-participant voxel encoding gave top-voxel means of $+0.165$ for MI and $+0.138$ for CLAP, using evaluation-selected voxels and encoder-specific folds that preclude a matched superiority claim. Direct-correlation shuffles passed in 4/4 participants for MI, 1/4 for routing-off and 0/4 for random controls; they do not validate ridge scores. The model-magnitude profiles correlated across audio collections at $r=0.998$; assigned coordinates were within 10 mm of 28/31 published peaks; 52/56 tracks placed the caudate-labelled channel first when the preceding-peak rule was used. These model configuration and timing checks cannot establish biological mechanisms.

The repertoire grew from the Cognitive Consonance Circuit literature-synthesis and exploratory meta-analysis programme. Published findings, extracted claims and theoretical proposals provided starting points for the design. Turning those ideas into computations required further choices about inputs, coefficients and temporal support. Figure 9 and Supplement S5 make that translation inspectable by connecting mechanisms to literature anchors, implemented operations and evaluation roles. Worked examples follow a proposal through its source observation to a specified test. The original quantitative synthesis motivates hypotheses; it supplies neither validated equation coefficients nor performance estimates. Its extraction and dependence limitations remain documented in S5. Unresolved translations identify hypotheses for further adjudication.

Stable mechanism names link the implementation to its design history. Supplement S5b distinguishes topic motivation, component-level background and unsupported extensions of a label. The source-role review examined primary abstracts and accessible full texts, recording access limits and preserving historical anchors. This separates what a source establishes from what a computation proposes. Brainstem, synaptic, memory and transfer interpretations require corresponding measurements; the stored falsification criteria describe tests still to be performed.

The study uses saved executions, acoustic and mechanism arrays, and existing human-rating datasets; no new participants were recruited. Configuration records identify implementations, requests, execution order and output mappings. Mechanisms run in processing-depth order, with discovery order retained at equal depths, and upstream values contribute only when consumed. The Swan Lake trace records ordered sources and inputs. Earlier caches have separate execution records; equal output counts do not identify the code that produced them. Supplement S5 links mechanism names to literature and evaluation roles, distinguishing motivation for a label from derivation of its equation. Independent reproduction requires implementation and reconstruction records, available through controlled access.

At 44,100 Hz and a 256-sample hop, the inspected adapter returns 172.265625 frames per second. Individual R³ analysis windows, longer temporal operations and recording-wide normalization determine the support of each coordinate. The behavioural analyses below use full-excerpt context and output means. Forward and centred operations are part of those offline representations. Numerical coefficients and channel names are documented with their source definitions; fitted observation models are specified separately.

The source Swan Lake recording lasts 182.597 s. Its 30–50-s display interval was selected before inspecting intervention effects. The saved execution identifies the input, selected configuration, outputs and observed consumption. The intervention changes one temporal-context requirement with other computations held fixed. A dimension is counted as changed if any common valid frame exceeds the recorded absolute tolerance of one ten-millionth plus a relative tolerance of one millionth. Observer and identity-adapter controls assess instrumentation. Native-frame results determine the counts; display averaging is used only for visualization. S1 records the scope and results of these checks; the implementation recipe remains under controlled access.

The film-tension comparison retains 110 cache entries grouped into 40 source soundtracks; repeated excerpts and segments remain in the same group. Five outer GroupKFold splits provide held-out predictions. Four grouped inner folds select ridge alpha from 0.1, 1, 10, 100, 1,000 and 10,000 by mean squared error. Mean imputation and standardization are fitted inside each training split. Features use the supplied full clip after omission of the first 172 cached frames. The representations comprise 97 R³ means; 970 acoustic/temporal descriptors; those 970 descriptors plus fifteen C³ means; the fifteen C³ means alone; two training-fold PCA15 representations; a fixed fifteen-descriptor acoustic subset; and a training-target-mean reference. Each R³ channel supplies its mean, standard deviation, 10th/50th/90th percentiles, mean absolute successive difference, normalized-time slope and autocorrelations at 0.35, 1 and 4 s for the 970-descriptor representation. The acoustic subset uses mean, variability and successive change in roughness, pleasantness, loudness, tonalness and spectral flux. The C³ set includes all channels of the existing tension pool. PCA is fitted within each inner and outer training split to standardized predictors, without whitening or subsequent component rescaling.

Pooled out-of-fold RMSE is the primary score; Pearson r and R² summarize the same predictions. R² uses squared prediction error relative to variation around the pooled observed target mean. The five paired folds are reported descriptively. The compact extension retained the earlier outer folds and tuning grid; its alternatives were specified before those fits. Full columns, fold identities, predictions and tuning records accompany Supplement S2.

MUSIFEAST-17 supplies perceived arousal and valence ratings for thirty-second excerpts. Its recorded partition comprises seventeen previously examined pilot excerpts, 274 development excerpts and 65 fixed test excerpts, grouping connected artist/composer, source and identical-audio identities. The total of 356 includes both pilot and 65-excerpt test subset. Channel means use the released excerpt after omitting its first 172 cached frames. Each direct Spearman correlation relates a channel mean to a human excerpt rating, without fitting a regression or removing acoustic contributions from the target ratings.

Four arousal and four valence channel identities were transferred from the prior film tables after the initial pilot, without channel reselection on the test subset. Historical film energy/arousal and valence targets, populations and engine snapshots differ from MUSIFEAST. Normative labels had been accessed for curation and quality control. Supplement S3 retains all eight channels and the partition chronology. The main arousal table reports the complete four-channel arousal family.

Consonance analyses correlate model descriptors with published interval or chord means, retaining each target convention. The fixed pleasantness combination is stated in Results; this evaluates established acoustic cues within MI, not the necessity of its cognitive core. The harmonic-dyad comparison uses thirteen means and the pooled-chord analysis 617 rows. The extended panel uses recorded six-partial harmonic synthesis; flute, guitar and piano identify human rating conditions. Two analytical references are distinguished from seven behavioural targets.

The Carillon evaluation applies one extractor with unchanged coefficients to twelve combinations of four registers and three spectral/decay constructions. These inputs are synthesized dyads. All twelve correlations are retained. The spectral route, normalization, synthesis definitions and rating references appear in Supplement S4. Algebraically related coordinates are presented together as properties of the same representation.

A dataset was eligible for a proposed comparison when it supplied the relevant human target, accessible audio or explicitly documented stimulus construction, recoverable stimulus–response mapping and timing adequate to the observation scale. The bounded dataset review inventoried prior use and considered additional releases, prioritizing 2025–2026. Supplement S6 distinguishes processed, partially available, unresolved and deferred cases with dated source records. It also reports completed comparisons whose increments were small or negative. Availability alone does not establish suitability, and an available but deferred resource is not classified as inaccessible.

These evaluations are exploratory because corpus-guided development shaped the repertoire before comparison. The film corpus and tension pool were already known, and MUSIFEAST transfer followed an inspected pilot. Training-fold preprocessing and nested tuning constrain fitting within a specified model; they cannot undo earlier representation selection (Cawley and Talbot, 2010). Bonferroni and BH corrections concern specified test families, leaving development exposure and effect-size superiority unresolved. Correlated channels and overlapping samples supply dependent evidence. Supplement protocols document negative results, historical execution identities and planning after data inspection. The narrow acoustic-comparator difference has no significance or equivalence test, and selected-output comparisons establish no necessity for all 84 mechanisms. Source definitions, analysis scripts and predictions are indexed in the accompanying record.

The primary contribution of MI is an executable architecture for developing computational hypotheses about music listening. It makes acoustic inputs, temporal context and dependencies among calculations explicit, allowing selected assumptions to be revised and their consequences traced. The Swan Lake example demonstrates this use in a recorded execution: reconstruction recovers a selected output, and changing one temporal request produces effects within the declared route. Empirical case studies assess selected readouts under their own observation protocols. These evaluations address complementary questions about the architecture’s operation and the use of its outputs.

Architectural research asks how a common organisation supports the construction and evaluation of models (Langley et al., 2009; Laird et al., 2017). For MI, the central commitment is to make the acoustic and temporal assumptions of a listening calculation accessible within its executed dependencies. A temporal request can be altered while retaining the surrounding computation; inspection can also reveal that an input required by an interpretation is absent. This is one way formalisation can sharpen theoretical questions (Guest and Martin, 2021). The present demonstrations establish this use for selected pathways. Broader usefulness requires showing that the same organisation supports additional hypotheses and controlled comparisons.

The 84 mechanisms and 889 dimensions form the current repertoire of computational proposals. Their number records scope; it does not establish a minimal set of human processes or a need for every mechanism. Fifteen output means support fitted tension prediction, but their RMSE differs from the fixed acoustic subset by 0.013, and appending them to generic descriptors adds no pooled improvement. These findings limit the predictive case for the larger repertoire. They leave the utility of explicit specification and revision as a separate question. Efficiency, upstream cost and the time required to develop or revise a model remain to be measured.

The observation scale helps explain the empirical pattern. Acoustic cues show consonance correspondence, and arousal ranks transfer across excerpts more strongly than valence. Yet the PMEmo dynamic pilot had no Bonferroni-significant arousal channel among 19 (S7). Strong excerpt associations also need not add prediction: the specified MUSIFEAST arousal family gave $\Delta R^2=-0.016$ beyond its controls, with an interval spanning zero (S6). Matched pleasure prediction gained about $0.005$ in mean fold $r$. Excerpt means discard the timing of changes, motivating tests of temporal support and the observation mapping. These outcomes identify where the representation is useful and where matched alternatives remain necessary. Regional encoding tests a further observation level; interpreting it as brain computation requires evidence beyond encoding success (Kriegeskorte and Douglas, 2019).

Three comparisons would extend the architectural evaluation: replace a selected composite with its acoustic constituents under the same observation protocol; revise temporal context across representative direct, shared and inter-mechanism pathways; and introduce measured listener history where an interpretation requires it. Restricting a forecast pathway to past information must include upstream normalization as well as its T³ window. These are proposed tests. Their value would lie in distinguishing computational alternatives while keeping the remaining assumptions explicit.

A next step is to test whether listener-specific experience improves prediction while the shared MI computations remain fixed. A proposed study would compare a common MI readout, a history-informed personal readout held fixed during the session, and an identically initialized readout updated only from completed listening blocks. All three would predict the same continuous pleasure reports, with predictions committed before each block and transfer assessed on new source recordings. Matched acoustic predictors and simple personal-calibration baselines would test whether any gain depends on the selected MI computations. This design separates the contribution of listening history from that of subsequent feedback. Model personalization would remain distinct from listener learning: a subsequent experiment could manipulate transition histories while matching sound exposure, then measure expectation and pleasure separately for identical probes. Such tests would motivate comparisons between adaptive readouts, adaptive temporal-context weighting and learned representations, using MI’s explicit component boundaries to formulate alternatives.

Independent reconstruction is also central to an architectural contribution. Implementation and reconstruction records are currently available through controlled access, and the public description does not specify all 84 mechanisms. A sufficiently detailed specification and independently executable worked example would strengthen external evaluation, as reproducible auditory modelling and model exchange illustrate (Majdak et al., 2022; Gleeson et al., 2023). MI provides a working basis for this research programme: an organisation in which listening hypotheses can be expressed as calculations, revised at identified points and assessed against appropriate observations.

Public results and analysis records are linked through the Musical Intelligence Results repository and the v2.0.0 archive. The Scientific Supplement supplies the mechanism crosswalk, dataset decisions and evaluation tables. The complete engine is available on request to amace@bu.edu under the stated research-access terms. Detailed source, configuration and reconstruction records require controlled access and are separate from the earlier deposit. The trace demonstrates reconstruction from its retained inputs; it does not establish unrestricted end-to-end reproducibility of the full engine from public files. Raw stimuli and ratings retain their publishers’ licences. OSF C8NE4 links protocol and analysis-plan records; musicalintelligence.app provides the project interface. Independent reruns require the matching implementation, inputs, configuration and access permissions.

Scientific Supplement

Implementation records: controlled access

AI assistance. Claude Code (Anthropic) and Codex (OpenAI) were used to assist software development and analysis, literature searching, manuscript drafting and revision, LaTeX editing, and document checks. The author is responsible for the scientific content and the final manuscript. Funding and interests. No external funding was received for this work. The author developed Musical Intelligence.