{"id":"9b0d99dc-1e4c-490e-b66e-3567aa12e29d","arxiv_id":"2607.25787","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"GraphIDyOM is a validated graph-native Python reimplementation of IDyOM that reproduces Lisp reference outputs with mean information-content differences below 0.003 bits and exposes predictive memories as graphs.","lead":"GraphIDyOM is a Python reimplementation of the IDyOM model of musical expectation, storing its long- and short-term memories as explicit graphs. It matches the original Lisp system to within 0.003 bits of information content and adds memory inspection, network analysis, and a local prediction server.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central 'faithful reimplementation' claim extrapolates beyond the validated viewpoint subset; Section 5's additive-encoding assumption is untested.","rationale":"The paper's measured central claim is well supported: agreement with an external Lisp reference at ΔIC < 0.003 bits and r ≥ .9997 across direct, projected, multi-joint, and multi-source configurations is strong evidence that the implemented prediction core is faithful for those configurations. The parameter count is small and inherited from prior IDyOM work rather than fitted here, and the validation is non-circular because it benchmarks against an independent reference. The main soft spot is the extrapolation from the tested viewpoint family to the full IDyOM architecture. The paper itself flags this in Section 5, listing omitted viewpoints and automated viewpoint selection as limitations. Because the strongest_claim in the paper is phrased as close reproduction under matched settings, the measured claim is not invalidated; however, the abstract's broader 'faithful reimplementation' language and the Section 5 assertion that omitted viewpoints require 'only additional viewpoint encoding and projection logic' go beyond what the experiments show. This is a scoping concern rather than an internal inconsistency or a numerical error. The reader's conditional verdict already reflects this appropriately, so no adjustment is needed.","tokens_in":14151,"tokens_out":8032,"duration_ms":87172,"concrete_test":"Implement one representative omitted viewpoint in GraphIDyOM—e.g., metrical position (beat/bar) or harmonic function—along with its Lisp IDyOM counterpart, and rerun the same 185-chorale, five-fold, seed-0 comparison under matched settings. If ΔIC remains below 0.003 bits and r ≥ .9997 with no changes to the PPM/escape/merge code, the additive-encoding assumption is supported. If the comparison requires touching the core prediction loop, or if agreement degrades, the faithful-reimplementation claim should be scoped to the implemented viewpoint family.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The validation demonstrates strong parity for the 12 configurations in Table 1, all built from pitch, interval, contour, duration, and two timing viewpoints on monophonic chorales. Section 5 concedes that metrical, harmonic, dynamic, and specialised viewpoints plus automated viewpoint selection are not implemented. The paper's bridge from measured agreement to the broad 'faithful reimplementation' claim is the assertion that extending coverage 'should primarily require additional viewpoint encoding and projection logic rather than changes to the validated prediction core.' That is an extrapolation, not a demonstrated property. If any omitted viewpoint changes the core PPM/merge path—for example, a viewpoint whose symbols are not simple atomic tokens, whose projection needs an alternative target alphabet, or whose joint target interacts with exclusion/update-exclusion differently—the validated core may not be sufficient. The reader's conditional verdict already captures this, but the load-bearing assumption is not merely 'more encoding work'; it is that the prediction core's semantics are viewpoint-independent. This deserves an explicit test before the broader claim is accepted.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces GraphIDyOM, a Python reimplementation of the IDyOM model for musical expectation. The implementation represents long-term and short-term predictive memories as explicit directed graphs at each Markov order, while retaining IDyOM's variable-order PPM prediction machinery, viewpoint projection, and LTM/STM merging. The authors validate the system against the original Lisp IDyOM on 185 Bach chorales using five-fold cross-validation: across 12 direct, projected, multi-joint, and multi-source configurations, mean absolute information-content differences are below 0.003 bits and Pearson correlations with the Lisp traces are at least r = .9997 (Section 3, Table 1). They also benchmark computational performance against Lisp IDyOM and IDyOMpy, and demonstrate extensions enabled by the graph-native representation: network analysis of learned memories, expectation-annotated musical networks, recency-sensitive retrieval based on PPM-decay, and a local server interface for interactive applications. The paper claims that GraphIDyOM is a faithful and accessible reimplementation that exposes internal memory structures for analysis and modification, while acknowledging that the full Lisp viewpoint ecosystem and automated viewpoint selection are not yet implemented.","tokens_in":14334,"tokens_out":5958,"duration_ms":66569,"significance":"If the claims hold, GraphIDyOM would be a valuable contribution to computational music cognition: it provides an open-source Python implementation with exceptionally strong numerical parity to the reference Lisp IDyOM on the tested configurations, and it makes the predictive memory inspectable and modifiable, which is a genuine methodological advance over black-box wrappers. The validation design is a clear strength: the implementation is compared against an external reference (Lisp IDyOM), the test configurations cover several architecturally distinct prediction modes, and the free parameters (escape methods, merge bias, max order) are inherited from prior IDyOM literature rather than fitted to the test set. The open-source code and reproducible validation protocol further strengthen the paper. The main reservation is that the broad 'faithful reimplementation' claim is supported only for a subset of the reference architecture's viewpoint ecosystem; the extrapolation to the full ecosystem is plausible but untested.","major_comments":[{"comment":"The central claim of a 'faithful reimplementation' is broader than what the validation supports. Table 1 covers 12 configurations built from pitch, interval, contour, duration, and timing viewpoints on monophonic chorales, but Section 5 explicitly states that metrical, harmonic, dynamic, and specialised viewpoints, as well as automated viewpoint selection, are not implemented. The sentence 'extending this coverage should primarily require additional viewpoint encoding and projection logic rather than changes to the validated prediction core' is an extrapolation, not a demonstrated property. If any omitted viewpoint changes the core PPM/merge path — for example, a viewpoint with non-atomic symbols, a different target alphabet, or an interaction with exclusion/update-exclusion — the validated core may not suffice. This is load-bearing because the abstract and introduction claim a faithful","section":"§5 (Limitations) and Abstract/Introduction"}],"minor_comments":[{"comment":"The validation statistics are reported as aggregate point estimates over the full held-out set: mean absolute ΔIC and Pearson r. There is no per-fold dispersion, standard deviation, or confidence interval. Given the very small differences this is unlikely to change the qualitative conclusion, but reporting fold-level ranges would make the parity claim more robust and would allow readers to assess variability across training splits.","section":"§3, Table 1"},{"comment":"The comparison with IDyOMpy is only approximate for several rows, as the paper acknowledges. For 'Pitch+octave' the IDyOMpy row uses 'pitch' rather than 'cpitch', and the table note says 'length-like rows use the closest aligned subset where required.' This makes the large ΔIC values for IDyOMpy not strictly comparable to the GraphIDyOM values. Please state explicitly which events and settings were used in each IDyOMpy row, or soften the comparison to avoid overstating the contrast.","section":"§3, IDyOMpy comparison"},{"comment":"The recency-sensitive retrieval demonstration applies PPM-decay to the graph counts but does not compare the resulting traces to the published PPM-Decay model of Harrison et al. (2020) or to any reference implementation. The paper already labels this as a methodological demonstration rather than a validation, which is appropriate; however, if the framework is advertised as supporting recency-sensitive PPM, a numerical check against the reference would strengthen the claim.","section":"§4.2, Figure 6"},{"comment":"The absence of automated viewpoint selection is more than a missing convenience: in Lisp IDyOM, viewpoint selection is part of the modelling workflow and affects which predictions are made for a given dataset. This limitation should be stated earlier in the paper (e.g., in the introduction or Section 2) so that readers are not surprised when the validation and examples use manually specified viewpoints.","section":"§5, automated viewpoint selection"},{"comment":"The text says the GraphIDyOM traces are 'visually indistinguishable' from Lisp IDyOM. This is a qualitative statement; the quantitative Table 1 already makes the point. Consider reporting the maximum absolute difference for the illustrated chorale, or at least noting that the figure is illustrative of the aggregate result.","section":"§3, Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid software contribution with a convincing validation on the tested configurations. The single load-bearing issue is the scope of the 'faithful reimplementation' claim relative to the unimplemented viewpoint ecosystem. I would support publication after a revision that either adds targeted validation for representative omitted viewpoints or narrows the claim to the implemented subset. The other issues are presentation-level and should be straightforward to address."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real reimplementation, validated against the Lisp reference with numbers that look good, and the graph-native memory layer is genuinely new. The main caveat is that 'faithful' currently covers a subset of the viewpoint ecosystem; the paper says the rest should be additive, but that's an assumption, not a test.\n\nThe genuinely new part is exposing the LTM/STM as explicit graph objects and letting users inspect, modify, and export them. That is a real enabling step for people who want to do network analysis of predictive memory or plug IDyOM into interactive systems. The validation is the strongest section: 9,227 event-wise comparisons under matched settings, with mean absolute IC differences below 0.003 bits and correlations at or above 0.9997 across direct, projected, multi-joint, and multi-source configurations. That is convincing evidence that the prediction core is correctly ported for the tested cases. The comparison against IDyOMpy is useful and the performance benchmark is reasonable, though the Lisp implementation still wins on speed, which the paper openly states.\n\nSoft spots, in proportion. First, the stress-test concern is real: Section 5 admits that metrical, harmonic, dynamic, and specialized viewpoints are not implemented, and the claim that adding them 'should primarily require additional viewpoint encoding and projection logic' is plausible but untested. It is the load-bearing assumption behind the word 'faithful.' A single additional viewpoint of a different kind, say a metrical one, would help a lot. Second, Table 1 has no error bars or confidence intervals, so we don't know how stable the 0.003-bit agreement is across folds. Third, the extension demonstrations in Section 4 are single illustrative examples—the PPM-decay result is not validated against Harrison et al.'s implementation, and the network analyses are qualitative. These are not fatal for a software paper, but they should be labeled more firmly as demonstrations. Finally, reproducibility is weaker than it could be: no commit hash, no packaged corpus, and the decay parameters for Figure 6 are not reported. The GitHub link is there, but a pinned version would be better.\n\nThe central claim—that GraphIDyOM faithfully reproduces Lisp IDyOM for the evaluated configurations—holds up. The extrapolation to the full viewpoint ecosystem is the only serious gap, and it is explicitly acknowledged.\n\nThis paper deserves a serious referee. It is an enabling tool for computational music expectation and the validation is strong enough to warrant careful review. I would recommend conditional acceptance, with requests for a pinned release, error bars on the main comparison, and at least one test of a non-pitch viewpoint to probe the additive-encoding assumption.","headline":"Solid, honestly-scoped reimplementation with strong validation for the covered configurations; the 'faithful' claim needs a little more evidence for the untested viewpoints.","tokens_in":14887,"tokens_out":1802,"would_cite":true,"duration_ms":19048,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A graph-native Python reimplementation of the IDyOM music-expectation model matches the Lisp reference to within 0.003 bits of information content while exposing its learned memory structures as inspectable graphs.","keywords":["musical expectation","IDyOM","information content","entropy","Prediction by Partial Matching","graph representation","reimplementation","recency-sensitive memory"],"falsifier":"Run the identical Table 1 validation protocol (185 Bach chorales, five-fold, seed 0, maximum order 5, LTM escape method c, STM escape method x, geometric LTM/STM merge with bias 7.0 and no offset) with a different random seed, say seed 1, and check all twelve rows: if any row shows a mean absolute information-content difference above 0.003 bits or a correlation below 0.9997, the parity claim is refuted. Alternatively, port an unimplemented Lisp viewpoint such as a metrical or harmonic viewpoint into GraphIDyOM and observe whether event-wise correlation drops, which would falsify the additive-e","tokens_in":13952,"feed_emoji":"🎵","tokens_out":4745,"duration_ms":47003,"temperature":0.7,"pith_summary":"This paper claims that the influential IDyOM model of musical expectation can be reimplemented in Python with its long-term and short-term predictive memories represented as explicit directed graphs, without losing numerical fidelity. Against the original Lisp implementation, on 185 Bach chorale melodies and across direct, projected, multi-joint, and multi-source viewpoints, event-wise information-content differences stay below 0.003 bits and correlations stay above r=.9997. The point matters because IDyOM's internal memory has been a black box: making it a graph object lets researchers inspect, export, modify, and query the exact evidence behind each prediction. The paper then shows three uses this enables: network analysis of expectation topology, recency-sensitive memory retrieval, and a local server for interactive or experimental control of predictions.","feed_headline":"GraphIDyOM matches Lisp IDyOM to under 0.003 bits","feed_subtitle":"A Python reimplementation makes the model's learned memories inspectable as graphs for analysis and interaction.","key_machinery":"Order-specific n-gram memory graphs: for each Markov order, nodes are context windows and weighted directed edges record observed continuations, so prediction is a lookup from the current context node's outgoing edges. Around these graphs sits the IDyOM pipeline: viewpoint encoding and projection, PPM variable-order combination across orders within each memory (escape methods a, b, c, d, x; exclusion and update-exclusion), and final LTM/STM merge via entropy-weighted geometric combination with a confidence bias of 7.0 in the validated configuration. The work of the graph representation is to make counts and contexts inspectable and modifiable as first-class objects.","core_discovery":"The paper's central claim is that the count-based variable-order prediction machinery of IDyOM—multiple viewpoints, source-to-target projection, PPM escape and exclusion handling, and entropy-weighted LTM/STM merging—can be exactly preserved while representing each Markov order's context–continuation statistics as a directed graph with weighted edges. Validated against Lisp IDyOM under matched settings, the implementation yields mean absolute information-content differences below 0.003 bits and Pearson correlations of at least 0.9997 across all evaluated configurations, including projected and multiple-viewpoint cases. The graph is not a compression or approximation; it is the memory itself,","pith_inferences":["If the parity claim holds broadly, GraphIDyOM effectively makes the Lisp implementation's behaviour auditable: any future discrepancy can be traced to a specific node or edge in the memory graph, a transparency advantage the Lisp system does not provide.","The graph-native memory could serve as a substrate for differential testing: deliberately mutating edge weights, such as removing low-count edges, and observing downstream information-content changes would allow causal analysis of which corpus statistics drive particular predictions.","The paper's 'faithful' label currently rests on a set of 12 configurations over one corpus; if untested viewpoints such as harmonic or metrical ones require changes beyond additive encoders, the extrapolation would fail. A natural test would be to port one additional viewpoint family and re-check parity.","The interactive server architecture suggests a broader class of applications: expectation-aware generative music systems could use next-event probabilities as interpretable constraints, which the graph representation also makes editable per context."],"forward_implications":["Researchers can now reproduce Lisp IDyOM's numerical outputs in pure Python, removing the Lisp dependency for the evaluated configurations.","Because memories are graphs, studies can map surprise and uncertainty onto musical topology, asking whether unexpected events sit on hubs, bridges, or peripheral states.","The optional recency-sensitive retrieval shows that stored context can be reweighted without retraining, enabling tests of forgetting and memory-decay hypotheses within the same architecture.","The local server turns expectation into a queryable, real-time signal, allowing stimuli to be designed to target entropy and information-content profiles, and interactive systems to use probabilities as control signals.","The framework's coverage of direct, projected, joint, and multi-source viewpoints exceeds that of previously available Python baselines while running with millisecond-level prediction latency."],"fun_headline_variants":["GraphIDyOM: IDyOM as graphs, exact to 0.003 bits","Python IDyOM with graph memories, matches to 0.003 bits","Graph-native IDyOM: inspectable memory, bit-exact match","IDyOM's memory as graphs: faithful and interactive","GraphIDyOM: rethink IDyOM as graphs, no approximation"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The parity numbers presume that GraphIDyOM's parser and viewpoint encoders are exactly equivalent to Lisp IDyOM for the evaluated settings; the claim that the full 'faithful reimplementation' extends to the whole Lisp viewpoint ecosystem depends on the assumption that the unimplemented metrical, harmonic, dynamic, and specialised viewpoints would only add encoding and projection logic and would not disturb the validated prediction core.","fun_headline_variants_meta":{"raw":{"variants":["GraphIDyOM: IDyOM as graphs, exact to 0.003 bits","Python IDyOM with graph memories, matches to 0.003 bits","Graph-native IDyOM: inspectable memory, bit-exact match","IDyOM's memory as graphs: faithful and interactive","GraphIDyOM: rethink IDyOM as graphs, no approximation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000169,"raw_usage":{"total_tokens":1097,"prompt_tokens":735,"completion_tokens":362,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":479,"completion_tokens_details":{"reasoning_tokens":264}},"tokens_in":479,"tokens_out":362,"duration_ms":4009,"temperature":1.0,"reasoning_tokens":264,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T01:27:48.138461+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical Table 1 validation protocol (185 Bach chorales, five-fold, seed 0, maximum order 5, LTM escape method c, STM escape method x, geometric LTM/STM merge with bias 7.0 and no offset) with a different random seed, say seed 1, and check all twelve rows: if any row shows a mean absolute information-content difference above 0.003 bits or a correlation below 0.9997, the parity claim is refuted. Alternatively, port an unimplemented Lisp viewpoint such as a metrical or harmonic viewpoint into GraphIDyOM and observe whether event-wise correlation drops, which would falsify the additive-e","supporting_citations":[],"review_version":1}