{
  "bootstrap_note": "95% CI on mean CER per engine (1000 bootstrap iterations).",
  "btn_customize": "⚙ Advanced",
  "btn_present": "⊞ Presentation",
  "cdd_help_axis": "The horizontal axis shows the mean rank of each engine (1 = best, k = worst).",
  "cdd_help_bars": "Thick horizontal bars connect engines that are statistically indistinguishable at the α = 0.05 level (Nemenyi post-hoc test).",
  "cdd_help_cd": "The red CD bar in the upper-left area shows the critical distance: two engines whose mean ranks differ by less than CD cannot be distinguished.",
  "cdd_help_ref": "Reference: Demšar (2006), Statistical Comparisons of Classifiers over Multiple Data Sets, JMLR 7:1-30.",
  "cdd_help_title": "How to read this diagram?",
  "cdd_title": "Multi-engine test — Friedman & Nemenyi",
  "char_engine_label": "Engine:",
  "class_col": "Class",
  "col_anchor": "Anchor",
  "col_anchor_title": "Anchor score: proportion of output trigrams found in the GT — low score = probable hallucinations (LLM/VLM)",
  "col_cer": "Exact CER",
  "col_cer_diplo": "Diplo. CER",
  "col_cer_diplo_title": "CER after diplomatic normalisation (ſ=s, u=v, i=j…) — measures substantial errors ignoring codified graphical variants",
  "col_cer_max": "Max CER",
  "col_cer_median": "Median CER",
  "col_cer_min": "Min CER",
  "col_diacritics": "Diacritics",
  "col_diacritics_title": "Diacritic preservation rate (accents, cedillas, umlauts…)",
  "col_docs": "Docs",
  "col_engine": "Engine",
  "col_gini": "Gini",
  "col_gini_title": "Gini coefficient of per-line CER errors — 0 = uniform errors, 1 = concentrated errors. A good engine has low CER AND low Gini.",
  "col_ligatures": "Ligatures",
  "col_ligatures_title": "Ligature recognition rate (fi, fl, œ, æ, ff…)",
  "col_mer": "MER",
  "col_overnorm": "Over-norm.",
  "col_overnorm_title": "Class 10 — LLM over-normalisation: rate of correct words degraded by the LLM",
  "col_rank": "#",
  "col_wer": "WER",
  "col_wil": "WIL",
  "corr_engine_label": "Engine:",
  "corr_note": "Pearson coefficient between CER, WER, image quality, ligatures, diacritics. Green = positive correlation, Red = negative.",
  "customize_columns": "Visible columns",
  "customize_composite_col": "Score",
  "customize_filters": "Stratum filters",
  "customize_filters_empty": "No stratum detected in corpus metadata.",
  "customize_reset": "Reset view",
  "customize_title": "Advanced mode — customization",
  "customize_weights": "Personal composite score",
  "customize_weights_disable": "Disable",
  "customize_weights_enable": "Enable",
  "customize_weights_none": "No non-zero weight — composite score inactive.",
  "customize_weights_warning": "These weights reflect your use case. There is no universally valid weighting — Picarones suggests none by default.",
  "date_locale": "en-GB",
  "doc_sidebar_header": "Documents",
  "doc_title_default": "Select a document",
  "errors_classified": "classified errors.",
  "footer_by": "by Picarones",
  "footer_generated": "Report generated on",
  "gallery_empty": "No documents match the filters.",
  "gallery_filter_all": "All",
  "gallery_filter_cer_label": "Filter CER >",
  "gallery_filter_engine_label": "Engine:",
  "gallery_sort_best": "Best engine",
  "gallery_sort_cer": "Mean CER",
  "gallery_sort_difficulty": "Difficulty",
  "gallery_sort_id": "Identifier",
  "gallery_sort_label": "Sort by:",
  "gini_cer_ideal": "— ideal: bottom-left",
  "gini_cer_note": "X-axis = mean CER, Y-axis = Gini coefficient. An ideal engine has low CER AND low Gini (rare, uniform errors).",
  "glossary_definition": "Definition",
  "glossary_empty": "No entry for this term.",
  "glossary_limits": "Limits",
  "glossary_measures": "What the metric measures",
  "glossary_reference": "Reference",
  "glossary_tooltip": "Definition",
  "glossary_usage": "Usage",
  "h_bootstrap": "95% Bootstrap Confidence Intervals",
  "h_cer_dist": "CER Distribution by Engine",
  "h_cer_doc": "CER by Document (all engines)",
  "h_characters": "Character Analysis",
  "h_clusters": "Frequent Error Clusters",
  "h_correlation": "Metric Correlation Matrix",
  "stratification_caption": "Ranking by stratum (script_type)",
  "stratification_description": "The global table ranks engines across the whole corpus. When the corpus is heterogeneous, some engines dominate on one document type and fail on another — the stratified view reveals this.",
  "stratification_median_label": "Median CER",
  "stratification_mean_label": "Mean CER",
  "stratification_docs_label": "Documents",
  "stratification_no_data_label": "—",
  "stratification_n_docs_label": "documents",
  "stratification_gap_summary": "Leader {leader} inter-stratum gap: {gap_pct} median CER points (between \"{min_stratum}\" and \"{max_stratum}\").",
  "h_calibration": "Engine calibration",
  "calibration_note": "ECE (Expected Calibration Error): weighted mean of |confidence − accuracy| gaps per bin. The lower the ECE, the more honest the engine is about its reliability — the diagonal in the diagram is perfect calibration. A high ECE means you cannot rely on confidence scores to focus human proofreading.",
  "calibration_summary_caption": "Engine calibration (ECE, MCE)",
  "calibration_engine_label": "Engine",
  "calibration_ece_label": "ECE",
  "calibration_mce_label": "MCE",
  "calibration_n_label": "Predictions",
  "calibration_acc_label": "Mean accuracy",
  "calibration_conf_label": "Mean confidence",
  "calibration_docs_label": "Docs evaluated",
  "reliability_diagram_title": "Reliability diagram",
  "reliability_x_axis": "Confidence",
  "reliability_y_axis": "Accuracy",
  "h_ner": "Named-entity precision",
  "ner_note": "F1 computed by IoU ≥ 0.5 span alignment (labels case-insensitive). Higher F1 means more faithful preservation of named entities (people, places, dates) — predicting downstream usefulness for prosopographic indexing. This metric measures OCR + NER extractor jointly; the extraction model itself may hallucinate.",
  "ner_summary_caption": "Named-entity precision",
  "ner_per_category_caption": "F1 by entity category (heatmap)",
  "ner_engine_label": "Engine",
  "ner_f1_label": "Global F1",
  "ner_precision_label": "Precision",
  "ner_recall_label": "Recall",
  "ner_doc_count_label": "Docs evaluated",
  "ner_hallucinated_label": "Hallucinations",
  "ner_missed_label": "Missed entities",
  "ner_no_data_label": "—",
  "h_inter_engine": "Cross-engine analysis",
  "inter_engine_note": "Higher divergence means two engines fail on different error classes — making them candidates for a voting ensemble. The oracle is the upper bound on token-level recall achievable by such a vote (bag-of-words proxy).",
  "divergence_caption": "Taxonomic divergence between engines",
  "divergence_metric_label": "Metric",
  "divergence_max_pair_label": "Most divergent pair",
  "divergence_diagonal_label": "(identity)",
  "oracle_caption": "Complementarity — potential gain from a majority vote",
  "oracle_best_engine": "Best single engine",
  "oracle_best_recall": "Tokens preserved",
  "oracle_recall": "Oracle (at least one engine)",
  "oracle_gap": "Potential ensemble gain",
  "oracle_doc_count": "Documents evaluated",
  "oracle_recoverable": "recoverable",
  "oracle_explanation": "The oracle is the upper bound on token-level recall achievable by a majority vote across engines (bag-of-words proxy).",
  "h_diff": "OCR Output — diff by engine",
  "h_duration": "Average Execution Time (seconds/document)",
  "h_gallery": "Document Gallery",
  "h_gini_cer": "Gini vs Mean CER",
  "h_gt": "Ground Truth (GT)",
  "h_hallucination": "Hallucination Analysis",
  "h_image": "Original Image",
  "h_line_metrics": "Error Distribution by Line",
  "h_pairwise": "Wilcoxon Tests — pairwise comparisons",
  "h_pareto": "Quality / cost trade-off",
  "h_quality_cer": "Image Quality ↔ CER (scatter plot)",
  "h_radar": "Engine Profile (radar)",
  "h_ranking": "Engine Ranking",
  "h_ratio_anchor": "Length Ratio vs Anchor Score",
  "h_reliability": "Reliability Curves",
  "h_taxonomy": "Error Taxonomy by Engine",
  "h_venn": "Shared / Exclusive Errors (Venn)",
  "hall_block_label": "Hallucinated block",
  "hall_blocks_title": "Blocks with no anchor in GT:",
  "hall_detected": "⚠️ Hallucinations detected",
  "hall_more_blocks": "additional block(s)",
  "hall_ok": "✓ Satisfactory anchoring",
  "heatmap_end": "End",
  "heatmap_mid": "Middle",
  "heatmap_start": "Start",
  "heatmap_title": "HEATMAP (position)",
  "html_lang": "en",
  "lines": "lines",
  "nav_report": "OCR report",
  "no_gini": "Gini data not available.",
  "no_hall_blocks": "No hallucinated blocks detected.",
  "no_hall_metrics": "No hallucination metrics available.",
  "no_line_metrics": "No line metrics available.",
  "no_scatter": "Data not available.",
  "pairwise_note": "Wilcoxon signed-rank test (non-parametric). Threshold α = 0.05.",
  "pareto_assumptions_summary": "Detailed assumptions per engine",
  "pareto_axis_co2": "Carbon (g CO₂)",
  "pareto_axis_cost": "Cost € / 1000 pages",
  "pareto_axis_speed": "Speed (s / page)",
  "pareto_dominated_label": "Dominated",
  "pareto_empty": "Not enough data for this view.",
  "pareto_front_label": "Pareto front",
  "pareto_note": "Engines on the Pareto front (highlighted) are those for which no other engine offers simultaneously a better CER AND a better cost. Prices are indicative (internal table, dated). Carbon mode is experimental.",
  "percentile_title": "CER PERCENTILES",
  "proportion_col": "Proportion",
  "quality_cer_note": "Each point = one document. X-axis = image quality score [0–1]. Y-axis = CER. Negative correlation expected.",
  "radar_note": "Radar axes: CER, WER, MER, WIL — inverted values (higher = better engine).",
  "ratio_anchor_note": "X-axis = trigram anchor score [0–1]. Y-axis = output/GT length ratio. ⚠️ Zone: anchor < 0.5 or ratio > 1.2 → probable hallucinations.",
  "ratio_anchor_subtitle": "— VLM hallucinations",
  "reliability_note": "For the X% easiest documents (sorted by ascending CER), what is the cumulative mean CER? A low curve = engine performing well even on easy documents.",
  "synth_cases_link": "To see how other teams reasoned through similar problems, see the case studies.",
  "synth_hint": "Generated mechanically from results — no LLM, reproducible.",
  "synth_title": "Factual summary",
  "tab_analyses": "Analyses",
  "tab_characters": "Characters",
  "tab_document": "Document",
  "tab_gallery": "Gallery",
  "tab_ranking": "Ranking",
  "taxonomy_engine_label": "Engine:",
  "taxonomy_note": "Distribution of error classes (classes 1–9 of the Picarones taxonomy).",
  "total_errors": "Total:",
  "venn_note": "Intersection of error sets between the 2 or 3 top engines. Shared errors = overlapping segments.",
  "philo_profile_title": "Philological profile",
  "philo_profile_note": "Raw counts per category of philological marker. The tool does not classify each engine's editorial convention — it is up to the researcher to read the figures and conclude according to their own editorial criteria.",
  "philo_engine_label": "Engine",
  "philo_global_label": "Global",
  "philo_strict_label": "Strict",
  "philo_expansion_label": "Expansion",
  "philo_n_total_label": "n total",
  "philo_unicode_blocks_title": "Accuracy by Unicode block",
  "philo_unicode_blocks_note": "Percentage of characters correctly restored per Unicode block found in the GT (excluding Basic Latin).",
  "philo_abbreviations_title": "Medieval abbreviations (Capelli)",
  "philo_abbreviations_note": "Strict = abbreviated form (ꝑ, ꝓ, ⁊…) preserved as-is; Expansion = abbreviated OR expanded form (per, pro, et…) present. The strict/expansion ratio per engine indicates the adopted convention (diplomatic / modernizing).",
  "philo_mufi_title": "MUFI coverage (Medieval Unicode Font Initiative)",
  "philo_mufi_note": "Rate of MUFI characters from the GT (þ, ð, ƿ, ſ, æ, PUA letters…) correctly restored by the OCR. Central editorial criterion for medievalists.",
  "philo_mufi_coverage_label": "Coverage",
  "philo_early_modern_title": "Early modern typographic markers (16th-18th c.)",
  "philo_early_modern_note": "Preservation of ligatures (fi fl ff), long s (ſ), dotless i (ı), ampersand (&) and nasal tildes (ã õ ñ). One row per engine, one column per category.",
  "philo_modern_archives_title": "Modern archive abbreviations (19th-20th c.)",
  "philo_modern_archives_note": "Strict = abbreviation preserved (Mme, S.A.R., bd, vol., …); Expansion = abbreviated OR expanded form. Displayed per category: civility, ordinals, currency, administrative, civil status, typographic punctuation, Latin, bibliographic, address.",
  "philo_roman_numerals_title": "Roman numerals: restitution by status",
  "philo_roman_numerals_note": "For each Roman numeral in the GT, restitution status: strict (exact form), case_changed (case modified), j_dropped (medieval j normalized), converted_to_arabic, lost. The breakdown indicates the convention: majority strict → diplomatic; majority arabic → deep modernization.",
  "philo_roman_status_strict_preserved": "Strict",
  "philo_roman_status_case_changed": "Case",
  "philo_roman_status_j_dropped": "j dropped",
  "philo_roman_status_converted_to_arabic": "→ arabic",
  "philo_roman_status_lost": "Lost",
  "pipeline_report_title": "Composed pipeline report",
  "pipeline_report_note": "Raw counts per step. The tool measures and aggregates — it does not classify the pipeline as 'good' or 'bad'. It is up to the researcher to judge the figures according to their own criteria.",
  "pipeline_summary_title": "Benchmark summary",
  "pipeline_name_label": "Pipeline",
  "pipeline_corpus_label": "Corpus",
  "pipeline_n_docs_label": "Documents",
  "pipeline_succeeded_label": "Pipelines succeeded",
  "pipeline_failed_label": "Pipelines failed",
  "pipeline_duration_label": "Total duration",
  "pipeline_steps_title": "Per-step details",
  "pipeline_step_name_label": "Step",
  "pipeline_success_rate_label": "Success rate",
  "pipeline_duration_mean_label": "Mean duration",
  "pipeline_duration_median_label": "Median duration",
  "pipeline_junction_metrics_label": "Junction metrics",
  "pipeline_error_breakdown_label": "Errors",
  "pipeline_docs_short": "docs",
  "pipeline_comparison_report_title": "Pipeline comparison report",
  "pipeline_comparison_report_note": "Raw comparative data. The tool measures and ranks — it does not settle the editorial debate. It is up to the researcher to read the figures and conclude according to their own criteria.",
  "pipeline_comparison_summary_title": "Comparison summary",
  "pipeline_n_pipelines_label": "Pipelines compared",
  "pipeline_n_pipelines_short": "pipelines",
  "pipeline_per_pipeline_label": "Per pipeline",
  "pipeline_ranking_title": "Ranking by {label}",
  "pipeline_rank_label": "Rank",
  "pipeline_value_label": "Value",
  "pipeline_gain_title": "Gain vs {baseline} on {label}",
  "pipeline_gain_absolute_label": "Absolute gain",
  "pipeline_gain_relative_label": "Relative gain",
  "pipeline_baseline_marker": "(baseline)",
  "baseline_corpus_title": "Is this corpus typical?",
  "baseline_corpus_harder": "Difficulty observed {current:.2f} — at the {percentile:.0f}th percentile of the {n_runs} previous corpora of your institution. This corpus is harder than usual.",
  "baseline_corpus_easier": "Difficulty observed {current:.2f} — at the {percentile:.0f}th percentile of the {n_runs} previous corpora. This corpus is easier than usual.",
  "baseline_corpus_usual": "Difficulty observed {current:.2f} — at the {percentile:.0f}th percentile of the {n_runs} previous corpora. This corpus is in the average.",
  "taxocooc_title": "Error class co-occurrence",
  "taxocooc_note": "Jaccard index at the document level: 1.00 = these two classes always appear together; 0.00 = never. Read the most co-occurring pairs below.",
  "taxocooc_n_docs": "Computed on {n_docs} documents.",
  "taxocooc_pair_label": "Pair",
  "taxocooc_jaccard_label": "Jaccard",
  "intradoc_title": "Intra-document evolution of error classes",
  "intradoc_note": "Heatmap class × position: relative density per class (darker = concentrated). A class concentrated in the first column suggests a margin error; a uniform distribution suggests a scribe error.",
  "intradoc_n_words": "Computed on {n_words_gt} GT words, split into {n_bins} bins.",
  "taxocomp_title": "Taxonomic profile: {engine_a} vs {engine_b}",
  "taxocomp_note": "Mirror chart of error proportions per class. Color by editorial recoverability (green = correctable, red = irrecoverable). At equal global CER, an engine whose errors are mostly green is preferable for a critical edition.",
  "taxocomp_level_label": "Category",
  "taxocomp_recoverable": "Recoverable",
  "taxocomp_difficult": "Difficult",
  "taxocomp_irrecoverable": "Irrecoverable",
  "lexmod_title": "Lexical modernization (top tokens)",
  "lexmod_note": "GT tokens that the engine rewrites most often. Reading: « maistre → maître modernized in 85 % of cases » tells you what to fix in the prompt to preserve historical spelling.",
  "lexmod_gt_label": "Historical GT form",
  "lexmod_hyp_label": "OCR variants",
  "lexmod_n_label": "n GT",
  "lexmod_rate_label": "% modernized",
  "levers_title": "Improvement leverages",
  "levers_note": "Factual observations synthesized from the analysis modules. No prescription imposed — the researcher decides what is actionable for their workflow.",
  "levers_top_classes": "Main classes:",
  "levers_importance_high": "Important",
  "levers_importance_medium": "Notable",
  "levers_importance_low": "Minor",
  "levers_label_dominant_recoverable_class": "Mostly recoverable error classes",
  "levers_label_pareto_concentration": "CER Pareto concentration",
  "levers_label_complementarity_observation": "Inter-engine complementarity",
  "levers_label_lexical_modernization_observation": "Systematic lexical modernization",
  "levers_label_robustness_projection_observation": "Projected deficit on the real corpus",
  "levers_dominant_recoverable_phrase": "{pct}% of {engine}'s errors ({n_recov}/{n_total}) are classified as recoverable (case_error, ligature_error, abbreviation_error).",
  "levers_pareto_phrase": "On {engine}, {n_top} documents ({top_pct}% of the corpus) concentrate {cer_pct}% of the cumulative CER (out of {n_total} documents).",
  "levers_complementarity_phrase": "The bag-of-words oracle achieves a recall {abs_pct} points higher (+{rel_pct}% relative) than the best single engine.",
  "levers_complementarity_phrase_with_engine": "The bag-of-words oracle achieves a recall {abs_pct} points higher (+{rel_pct}% relative) than the best single engine ({best_engine}).",
  "levers_lexical_phrase": "Top GT tokens systematically modernized by {engine}: {items}.",
  "levers_robustness_phrase": "Projected deficit of {engine} on the real corpus: {deficit_pct} CER points cumulated over {n_types} degradations.",
  "levers_robustness_phrase_with_worst": "Projected deficit of {engine} on the real corpus: {deficit_pct} CER points cumulated over {n_types} degradations — worst degradation: {worst_type} ({worst_pct} points).",
  "search_title": "Fuzzy searchability",
  "search_note": "Fraction of GT tokens recovered in the OCR output within Levenshtein distance ≤ 2 — direct proxy of full-text search quality (Elastic, Solr, Gallica).",
  "search_engine": "Engine",
  "search_recall": "Recall",
  "search_count": "Recovered tokens / total",
  "search_docs": "Docs",
  "numseq_title": "Numerical-sequence precision",
  "numseq_note": "Strict score (form preserved) — the value in parentheses is the score on the value (XIV ↔ 14 accepted). Foliation: recto/verso are not interchangeable.",
  "numseq_engine": "Engine",
  "numseq_global": "Global",
  "numseq_cat_year": "Year",
  "numseq_cat_roman": "Roman",
  "numseq_cat_foliation": "Foliation",
  "numseq_cat_currency": "Amount",
  "numseq_cat_regnal": "Regnal",
  "readability_title": "Readability (Flesch delta)",
  "readability_note": "Flesch readability difference between OCR output and GT. Δ > +5: over-normalisation (typical of LLMs that modernize archaic text). Δ < -5: heavy OCR degradation. Δ ≈ 0: faithful to the linguistic register.",
  "readability_engine": "Engine",
  "readability_delta_mean": "Mean Δ",
  "readability_delta_median": "Median Δ",
  "readability_over_norm_rate": "% over-normalised",
  "readability_under_norm_count": "Under-normalised docs",
  "readability_docs": "Docs",
  "robproj_title": "Projected robustness deficit on the real corpus",
  "robproj_note": "Projection of synthetic degradation curves onto real image characteristics. The total deficit assumes independence of degradations — a useful diagnostic approximation, not a verdict.",
  "robproj_summary": "Per-engine summary",
  "robproj_detail": "Detail per (engine × degradation) pair",
  "robproj_engine": "Engine",
  "robproj_total": "Total deficit (CER pts)",
  "robproj_n_types": "Types evaluated",
  "robproj_worst": "Worst degradation",
  "robproj_deg_type": "Degradation",
  "robproj_n_docs": "Docs",
  "robproj_n_with_data": "Docs with data",
  "robproj_deficit": "Projected ΔCER (pts)",
  "robproj_above": "Docs ≥ critical threshold",
  "specialization_title": "Inter-engine specialisation",
  "specialization_note": "Jensen-Shannon divergence between the taxonomic profiles of each pair of engines (0 = identical profiles, 1 = fully disjoint). A highly specialised pair signals error categories of different natures — it is for the researcher to act on it, not for the tool to prescribe an ensemble.",
  "specialization_engine_a": "Engine A",
  "specialization_engine_b": "Engine B",
  "specialization_score": "Score",
  "specialization_category": "Reading",
  "specialization_cat_similar": "Similar profiles",
  "specialization_cat_distinct": "Distinct profiles",
  "specialization_cat_highly_specialized": "Highly specialised",
  "stability_title": "Multi-run stability",
  "stability_note": "When an LLM/VLM engine is non-deterministic, the variance between successive runs on the same documents is a proxy for scientific reliability. A high CV or low identical-run rate undermines the interpretation of the average CER.",
  "stability_engine": "Engine",
  "stability_n_runs": "Runs",
  "stability_cer": "Mean CER ± σ",
  "stability_cv": "CV (%)",
  "stability_identical": "% identical runs",
  "stability_distinct": "Distinct outputs",
  "throughput_title": "Effective throughput",
  "throughput_note": "Pages processable per hour including post-OCR human correction time. Discriminates between a fast but imprecise cloud and a slow but reliable local engine. Correction constant: {time_per_error}s per error (HTR-United default, overridable).",
  "throughput_engine": "Engine",
  "throughput_raw": "Pages/h raw",
  "throughput_effective": "Pages/h usable",
  "throughput_drag": "% correction",
  "throughput_pages": "Pages",
  "throughput_errors": "Errors",
  "longitudinal_title": "Evolution over time",
  "longitudinal_note": "Trend and change-points on the SQLite history of previous runs. A positive change signals cumulative degradation — useful to link a regression to a pipeline or model change.",
  "longitudinal_engine": "Engine",
  "longitudinal_n_runs": "Runs",
  "longitudinal_first": "First CER",
  "longitudinal_last": "Last CER",
  "longitudinal_delta": "Cumulative Δ (pts)",
  "longitudinal_slope": "Annual slope (pts/yr)",
  "longitudinal_r2": "R²",
  "longitudinal_change": "Change-point",
  "imgpred_title": "Corpus image profile",
  "imgpred_note": "Palaeographic complexity score combining noise, blur, low contrast and rotation. The homogeneity score signals whether the global average is reliable (uniform corpus) or misleading (heterogeneous corpus — then see the stratified view).",
  "imgpred_complexity": "Palaeographic complexity",
  "imgpred_homogeneity": "Corpus homogeneity",
  "imgpred_score": "Score",
  "imgpred_mean": "Mean",
  "imgpred_median": "Median",
  "imgpred_min": "Min",
  "imgpred_max": "Max",
  "imgpred_stdev": "Stdev",
  "imgpred_docs": "Docs",
  "imgpred_feature": "Feature",
  "imgpred_feat_mean": "Mean",
  "imgpred_feat_stdev": "Stdev",
  "imgpred_feat_norm": "Normalised contribution",
  "imgpred_feat_noise": "Noise level",
  "imgpred_feat_sharpness": "Sharpness",
  "imgpred_feat_contrast": "Contrast",
  "imgpred_feat_rotation": "Rotation",
  "absorption_title": "Error absorption per junction",
  "absorption_note": "At each junction of the pipeline, two flows are measured independently: how many errors are corrected vs. how many are introduced. A junction that corrects a lot but also introduces a lot is absorbing upstream differences rather than improving them.",
  "absorption_junction": "Junction",
  "absorption_errors_before": "Errors before",
  "absorption_errors_after": "Errors after",
  "absorption_corrected": "Corrected",
  "absorption_introduced": "Introduced",
  "absorption_corr_rate": "% corrected",
  "absorption_intro_rate": "% introduced",
  "absorption_net": "Net improvement",
  "absorption_sample": "Sample (intro)",
  "dag_title": "Pipeline DAG",
  "dag_note": "Directed graph of the composed pipeline. Each edge shows the artifact type transmitted and the metric computed at the junction. Green/orange/red colour code by threshold. Inspection tool — YAML remains the source of truth.",
  "dag_legend": "Reading",
  "dag_legend_green": "high quality",
  "dag_legend_yellow": "moderate quality",
  "dag_legend_red": "low quality",
  "incr_title": "Controlled comparison by slot",
  "incr_note": "Isolated effect of the varying module on pipelines while controlling for the other slots. For each value of the slot: mean ± stdev, average rank across fixed groups, number of observations. A design-of-experiment approach for honest comparisons.",
  "incr_slot_label": "Varying slot",
  "incr_value": "Value",
  "incr_mean": "Mean score",
  "incr_stdev": "± σ",
  "incr_rank": "Mean rank",
  "incr_n_obs": "Observations",
  "incr_groups": "Fixed groups",
  "audit_title": "Audited modules",
  "audit_note": "Summary of the modules used in the composed pipeline. A module that does not pass the audit is not executable. Metadata comes from the manifest provided by the contributor (author, licence, citation).",
  "audit_pass": "audit OK",
  "audit_fail": "failed checks",
  "audit_module": "Module",
  "audit_status": "Audit",
  "audit_version": "Version",
  "audit_author": "Author",
  "audit_license": "Licence",
  "audit_io": "Input → output",
  "audit_citation": "Citation",
  "audit_homepage": "Project page"
}