These are the explanations that Claude Sonnet 5 gave in the extraction session next to its ATWL code: the note on the extraction approach and, where the session gave it, the table of key extraction decisions; after the review, the summary of the changes. Only the text around the code is shown; the ATWL representations are in atwl/. The session processed the five papers one after another (UTOPIAN, EventThread, NodeTrix, DPVis, SOMFlow). GPT-5.6-Terra did not explain its decisions.
UTOPIAN
Note on extraction approach: UTOPIAN is a tool-focused paper (not a single case study). Per extraction guidelines for tool papers, I synthesized a representative composite workflow from the three usage scenarios (TV Reviews, Car Reviews, InfoVis-VAST), which share a common analytical pattern: fit an initial topic model, visually inspect and interpret topics, and iteratively refine them through keyword- or document-based semi-supervision until the topic structure is satisfactory.
After the review
Summary of changes applied
| # | Fix | Status |
|---|---|---|
| 1 | Template now includes generate-knowledge (specify refinement) inside the loop |
✅ Fixed |
| 2 | Added embedment: set to D_topics |
✅ Fixed |
| 3 | T_project input changed to D_documents, F_document_terms, F_topic_label (bag-of-words/similarity-based modified t-SNE, not document-topic weights) |
✅ Fixed |
| 4 | V_topics now encodes proximity links explicitly in form/encoding |
✅ Fixed |
| 5 | V_topics added as input to T_specify_refinement |
✅ Fixed |
| 6 | Removed unsupported "refinement history" claim from T_synthesise/K_findings; K_findings now references P_topics, K_topic_quality matching its actual inputs |
✅ Fixed |
All other elements (document encoding, build-model/define-unit/characterise decomposition, explicit assessment loop with assign) remain unchanged, as they were confirmed correct in the review.
EventThread
Note on extraction approach: Unlike UTOPIAN, EventThread describes a single, well-defined analytical pipeline (tensor-based event sequence summarization) that is applied consistently across all three usage scenarios (COPD cohort, vehicle maintenance, academic behaviour). The computational pipeline (filtering → alignment → stage folding → tensor decomposition → thread extraction → layout → interactive clustering) is fixed; only domain-specific parameters and interpretations vary. The workflow below represents this single methodology rather than a composite across divergent modes.
After the review
Summary of changes applied
| # | Fix | Status |
|---|---|---|
| 1 | T_visualise_categories/V_threads now include F_thread_stage_likelihood, D_sequences_filtered, and D_stages as inputs; encoding describes opacity/dashing from stage likelihood and raw-event/entity provenance for contextual views |
✅ Fixed |
| 2 | T_decompose/M_tensor_model now state the orthogonality constraint applies only to the stage factor matrix, not all factors |
✅ Fixed |
| 3 | T_filter now uses "TF-IDF-based scoring" terminology, matching the paper's method |
✅ Fixed |
| 4 | D_sequences features replaced with f_event_types and f_event_times, removing unused f_length/f_duration |
✅ Fixed |
| 5 | F_tensor description clarified as a "three-way event-stage-entity incidence tensor" |
✅ Fixed |
| 6 | K_findings now references P_patterns, K_cluster_quality, matching its actual transform inputs |
✅ Fixed |
All other elements (filtering/alignment/stage-folding pipeline, CP decomposition structure, thread definition, optimisation-based layout, and the explicit assessment-driven clustering loop with assign) remain unchanged, as they were confirmed correct in the review.
NodeTrix
Note on extraction approach: NodeTrix is a tool/technique paper (hybrid node-link/matrix visualization), not a single empirical case study. Per extraction guidelines for tool-focused papers, I synthesized a representative composite workflow primarily from Section 6 (the InfoVis co-authorship case study), which demonstrates the tool's full analytical cycle: initial layout, interactive aggregation of dense subgraphs into matrices, hybrid visual exploration, and interpretation of collaboration patterns and actor roles (tasks T1–T3).
Key extraction decisions
| Decision | Rationale |
|---|---|
Researchers modelled as individual entities with embedment: relational structure |
The paper's data model treats researchers as vertices in a graph; per the entities/embedment distinction, when each researcher is a separate entity, the network connectivity becomes embedment rather than internal structure |
F_collaboration as relational configuration feature |
Captures the co-authorship edges and their weights without modelling the whole network as a single formation entity |
Community aggregation = define-unit with actor: hybrid |
Matches the lasso-select-and-aggregate interaction (human selection + machine attribute propagation/rendering), analogous to clustering transforms that jointly produce groups and membership labels |
Attribute propagation modelled as separate characterise transform |
Reflects the paper's explicit description of combining/aggregating nominal and numeric attributes up to aggregate nodes |
Loop follows Pattern B (no conditional; until governs termination) |
All transforms (adjustment, re-visualisation, interpretation, assessment) execute every iteration of exploratory interaction, mirroring the inner exploration loop in bin_class_diagnostics.atwl |
assign targets limited to D_communities, F_community_membership, F_community_attributes, V_nodetrix, P_patterns |
These are the artifacts genuinely carried forward and refined across iterations; K_exploration_quality is freshly produced each iteration and not reassigned |
Three analytic tasks (T1–T3 in the paper) folded into P_patterns/K_findings |
Community identification, central-actor identification, and role/position analysis are all expressed through the same interpret → assess → synthesise structure, consistent with how the paper treats them as facets of one exploratory process |
After the review
Summary of changes applied
| # | Fix | Status |
|---|---|---|
| 1 | Added T_arrange_communities/A_community_layout as arrangement(D_communities); V_nodetrix now uses A_community_layout instead of A_node_link |
✅ Fixed |
| 2 | Separated aggregation (T_adjust_aggregation) from spatial repositioning (T_reposition_communities, producing A_community_layout_r with assign) |
✅ Fixed |
| 3 | V_nodetrix/V_nodetrix_r now reference D_researchers, F_community_membership, F_community_attributes to match their encodings |
✅ Fixed |
| Imp.1 | Corrected matrix-ordering claim: membership determines composition; order is manually adjustable | ✅ Fixed |
| Imp.2 | T_visualise_r reframed as an interactive visualisation update consuming prior V_nodetrix state, with actor: hybrid |
✅ Fixed |
| Imp.3 | T_layout manner updated to "energy-based cluster-prominent graph layout" |
✅ Fixed |
| Imp.4 | K_findings now references P_patterns, K_exploration_quality |
✅ Fixed |
| Minor | F_community_attributes value type extended to {numeric, categorical, text} |
✅ Applied |
Template updated to reflect the two contextualise transforms (initial layout and aggregate-community arrangement) and the contextualise (reposition aggregates) step inside the loop, per Issue 2's correction.
DPVis
Note on extraction approach: DPVis is a tool-focused paper (not a single empirical study). Per extraction guidelines for tool papers, I synthesized a representative composite workflow from the usage scenario (Section 5.2) and the five analytic tasks (T1–T5): training and selecting a hidden Markov model, characterising and interpreting discovered disease states and their transition pathways, and iteratively building, visualising, and assessing patient subgroups (cohorts) to uncover associations with outcomes and patient characteristics.
Key extraction decisions
| Decision | Rationale |
|---|---|
D_visits (elementary, {set, time}) and D_subjects (sequence, set) as two levels of entities |
Mirrors the ATWL D2/D2_seq pattern: visits are the raw observational units used for HMM state inference, while subjects (per-patient visit sequences) are the units for trajectory and cohort-level analysis (T2–T5) |
Model selection via cross-validation folded into a single build-model manner clause rather than a separate loop |
Per abstraction-level guidance, standard hyperparameter search is implementation detail; the paper's own analytically significant iteration is the interactive cohort-refinement loop, which is retained explicitly |
F_state_sequence as feature(D_subjects) derived via a dedicated characterise transform |
Separates visit-level state inference (F_state_label) from the subject-level trajectory representation needed for Pathway views and trajectory-group abstraction |
D_cohort with internal structure: group/cluster and no embedment |
Each cohort artifact instance represents a single selected subgroup (one group), not a collection of peer groups, so embedment is omitted per the single-entity rule |
T_build_cohort with actor: hybrid |
Matches State Sequence Finder's interaction model: human specifies graphical query constraints, machine executes the filter |
| Loop follows Pattern A (gated refinement via conditional) | Matches the paper's sensitivity-analysis narrative: cohort criteria are refined only when the analyst judges the current cohort/association unclear; otherwise exploration concludes |
K_findings references P_trajectory_groups, P_associations, D_cohort |
Reflects T5's emphasis on retaining and transferring cohort definitions alongside substantive insights, matching T_synthesise's actual inputs |
No explicit arrangement artifacts |
DPVis's views (matrix, beeswarm/trajectory, sankey, chord, density plots) derive layout directly from entities and features without an intermediate computed/external context, consistent with the ATWL allowance for visualisations without arrangement |
After the review
Summary of changes applied
| # | Fix | Status |
|---|---|---|
| 1 | Added f_subject_id to D_visits; T_train_candidates/T_fit_final_model now use D_subjects, D_visits as input, making subject grouping and sequential ordering explicit model inputs |
✅ Fixed |
| 2 | Split model building into T_train_candidates (build-model) → T_assess_model_fit (assess, produces K_model_fit) → T_select_model (generate-knowledge, produces S_hmm_selected) → T_fit_final_model (build-model, produces final M_hmm) |
✅ Fixed |
| 3 | Replaced single replaced D_cohort with retained collection D_cohorts (group/cluster, embedment: set) and overlapping F_subject_cohort_memberships (list-valued); loop now builds/updates/retains cohorts via assign rather than replacing one subgroup |
✅ Fixed |
| 4 | T_visualise_cohorts/V_cohorts now include F_state_label, F_posterior, and F_subject_cohort_memberships, matching the encodings they drive |
✅ Fixed |
| Imp.1 | Added T_visualise_static_vars/V_static_vars before the loop; added as input to T_build_or_update_cohorts alongside V_transitions, V_state_summary |
✅ Fixed |
| Imp.2 | Added T_mine_transition_patterns/P_frequent_sequences (frequent sequential-pattern mining) feeding into V_transitions |
✅ Fixed |
| Imp.3 | K_findings now references P_trajectory_groups, P_associations, D_cohorts, K_cohort_quality, matching T_synthesise's actual inputs |
✅ Fixed |
| Minor | Removed unused f_num_visits from D_subjects |
✅ Applied |
Template updated to reflect the separated model-selection phase (build-model → assess → generate-knowledge → build-model), the explicit pattern-mining step, the Static Variables View, and the retained multi-cohort management loop.
SOMFlow
Note on extraction approach: SOMFlow is a tool-focused paper (not a single empirical case study). Per extraction guidelines for tool papers, I synthesized a representative composite workflow primarily from the Study 2 (exploratory analysis) usage scenario and Figure 1, which demonstrates the tool's fullest analytical cycle: iteratively training self-organizing maps on successive data subsets, analyzing cell quality/interestingness, and deciding to either adapt computations (e.g., normalization) or partition data further (e.g., filtering noise, splitting by metadata), drilling down progressively. The paper's four abstract exploration tasks (analyze, adapt, partition, reflect) structure the extraction directly.
Key extraction decisions
| Decision | Rationale |
|---|---|
SOM training split into build-model (M_som) + define-unit (D_cells/F_cell_assignment) + contextualise (A_cell_layout) |
The SOM simultaneously performs vector quantization (model), cluster/unit creation (cell assignment), and topology-preserving embedding (arrangement); ATWL requires one intent per transform, so these three facets are decomposed into separate transforms consuming the same M_som |
D_cells with internal structure: group/cluster and embedment: relational structure |
Each cell is itself an unordered group of assigned items (internal structure), while cells relate to one another through the SOM's intrinsic neighborhood topology (embedment) — both properties coexist at different levels, per the internal-structure-vs-embedment distinction |
Unified S_analysis_strategy specification covering preprocessing, metric, parameters, and partition criteria |
The paper explicitly groups "adapt" interactions (data, transformations, metric, parameter tuning) and "partition" interactions (cell/cluster/metadata-based selection) as two sides of one analyst decision at each drill-down step; a single configuration-schema specification avoids an artificial three-way branch while still controlling both the model-training and the next-subset-definition transforms |
| Loop follows Pattern A (conditional with gated refinement) | Matches the paper's recommendation-driven workflow: the analyst's assessment of quality/interestingness measures explicitly gates whether a new strategy and next subset are produced, or the drill-down branch concludes |
D_next may represent either a retained full subset (adapt-only) or a genuinely narrowed partition (partition) |
Reflects that, in the usage scenario, a single iteration sometimes only changes preprocessing (e.g., semitone normalization) and sometimes narrows the dataset (e.g., noise filtering, pitch-label split, speaker-group split); modeling both as define-unit outputs keeps the loop structure uniform across iteration types |
T_reflect/P_overall_relations after the loop |
Represents the paper's distinct "reflect" exploration task (flow-graph comparison, 2D-colormap linking across SOMs); simplified to operate on the final iteration's accumulated P_patterns/K_assessment since this is a linearised composite workflow representing one analysis branch, not the full multi-branch provenance graph |
| No explicit flow-graph/provenance entities artifact | Modeling the full branching SOMFlow graph (nodes = SOMs, edges = data flow) would require a persistent formation-structured entity updated every iteration; this was omitted to keep the composite workflow tractable, as the paper's case studies are best represented as a single representative drill-down chain |
After the review
Summary of changes applied
| # | Fix | Status |
|---|---|---|
| 1 | Added D_som_runs/F_run_lineage (retained branching provenance graph, embedment: relational structure), T_record_analysis_step (accumulates runs via assign), T_visualise_flow/V_som_flow, and revised T_reflect to compare across the retained graph rather than only the current iteration |
✅ Fixed |
| 2 | T_compute_cell_measures now takes M_som as input; T_visualise_som/V_som now include M_som and F_cell_assignment |
✅ Fixed |
| 3 | Moved define-unit (record run) → visualise (provenance flow) → abstract (reflect) into every loop iteration (before the gated branch), so reflection is continuous rather than a post-loop-only operation limited to final state |
✅ Fixed |
| Imp.1 | Added F_partition_recommendations (produced by T_compute_cell_measures); included in V_som and T_decide_strategy |
✅ Fixed |
| Imp.2 | Added T_annotate_cells/F_user_annotations; used in T_interpret and T_define_next (noise filtering) |
✅ Fixed |
| Imp.3 | D_next now declares f_pitch_contour and f_metadata features matching D_utterances' schema |
✅ Fixed |
| Imp.4 | K_findings now references P_overall_relations, K_assessment, D_som_runs, F_run_lineage, matching T_synthesise's actual inputs and including the provenance graph |
✅ Fixed |
| Minor | Added brief clarifying note in D_som_runs description that a cluster may span several similar neighbouring cells |
✅ Applied |
All other elements (SOM training/cell-assignment/grid-layout decomposition, quality/interestingness computation, strategy specification, and the Pattern-A gated conditional for adaptation/partitioning) remain structurally unchanged, as they were confirmed correct in the review.