@machinespirits/eval 0.1.2 → 0.2.1

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (102) hide show
  1. package/LICENSE +21 -0
  2. package/README.md +161 -0
  3. package/config/eval-settings.yaml +18 -0
  4. package/config/evaluation-rubric-learner.yaml +277 -0
  5. package/config/evaluation-rubric.yaml +613 -0
  6. package/config/interaction-eval-scenarios.yaml +93 -50
  7. package/config/learner-agents.yaml +124 -193
  8. package/config/machinespirits-eval.code-workspace +11 -0
  9. package/config/providers.yaml +60 -0
  10. package/config/suggestion-scenarios.yaml +1399 -0
  11. package/config/tutor-agents.yaml +716 -0
  12. package/docs/EVALUATION-VARIABLES.md +589 -0
  13. package/docs/REPLICATION-PLAN.md +577 -0
  14. package/index.js +15 -6
  15. package/package.json +16 -22
  16. package/routes/evalRoutes.js +88 -36
  17. package/scripts/analyze-judge-reliability.js +401 -0
  18. package/scripts/analyze-run.js +97 -0
  19. package/scripts/analyze-run.mjs +282 -0
  20. package/scripts/analyze-validation-failures.js +141 -0
  21. package/scripts/check-run.mjs +17 -0
  22. package/scripts/code-impasse-strategies.js +1132 -0
  23. package/scripts/compare-runs.js +44 -0
  24. package/scripts/compare-suggestions.js +80 -0
  25. package/scripts/compare-transformation.js +116 -0
  26. package/scripts/dig-into-run.js +158 -0
  27. package/scripts/eval-cli.js +2626 -0
  28. package/scripts/generate-paper-figures.py +452 -0
  29. package/scripts/qualitative-analysis-ai.js +1313 -0
  30. package/scripts/qualitative-analysis.js +688 -0
  31. package/scripts/seed-db.js +87 -0
  32. package/scripts/show-failed-suggestions.js +64 -0
  33. package/scripts/validate-content.js +192 -0
  34. package/server.js +3 -2
  35. package/services/__tests__/evalConfigLoader.test.js +338 -0
  36. package/services/anovaStats.js +499 -0
  37. package/services/contentResolver.js +407 -0
  38. package/services/dialogueTraceAnalyzer.js +454 -0
  39. package/services/evalConfigLoader.js +625 -0
  40. package/services/evaluationRunner.js +2171 -270
  41. package/services/evaluationStore.js +564 -29
  42. package/services/learnerConfigLoader.js +75 -5
  43. package/services/learnerRubricEvaluator.js +284 -0
  44. package/services/learnerTutorInteractionEngine.js +375 -0
  45. package/services/processUtils.js +18 -0
  46. package/services/progressLogger.js +98 -0
  47. package/services/promptRecommendationService.js +31 -26
  48. package/services/promptRewriter.js +427 -0
  49. package/services/rubricEvaluator.js +543 -70
  50. package/services/streamingReporter.js +104 -0
  51. package/services/turnComparisonAnalyzer.js +494 -0
  52. package/components/MobileEvalDashboard.tsx +0 -267
  53. package/components/comparison/DeltaAnalysisTable.tsx +0 -137
  54. package/components/comparison/ProfileComparisonCard.tsx +0 -176
  55. package/components/comparison/RecognitionABMode.tsx +0 -385
  56. package/components/comparison/RecognitionMetricsPanel.tsx +0 -135
  57. package/components/comparison/WinnerIndicator.tsx +0 -64
  58. package/components/comparison/index.ts +0 -5
  59. package/components/mobile/BottomSheet.tsx +0 -233
  60. package/components/mobile/DimensionBreakdown.tsx +0 -210
  61. package/components/mobile/DocsView.tsx +0 -363
  62. package/components/mobile/LogsView.tsx +0 -481
  63. package/components/mobile/PsychodynamicQuadrant.tsx +0 -261
  64. package/components/mobile/QuickTestView.tsx +0 -1098
  65. package/components/mobile/RecognitionTypeChart.tsx +0 -124
  66. package/components/mobile/RecognitionView.tsx +0 -809
  67. package/components/mobile/RunDetailView.tsx +0 -261
  68. package/components/mobile/RunHistoryView.tsx +0 -367
  69. package/components/mobile/ScoreRadial.tsx +0 -211
  70. package/components/mobile/StreamingLogPanel.tsx +0 -230
  71. package/components/mobile/SynthesisStrategyChart.tsx +0 -140
  72. package/docs/research/ABLATION-DIALOGUE-ROUNDS.md +0 -52
  73. package/docs/research/ABLATION-MODEL-SELECTION.md +0 -53
  74. package/docs/research/ADVANCED-EVAL-ANALYSIS.md +0 -60
  75. package/docs/research/ANOVA-RESULTS-2026-01-14.md +0 -257
  76. package/docs/research/COMPREHENSIVE-EVALUATION-PLAN.md +0 -586
  77. package/docs/research/COST-ANALYSIS.md +0 -56
  78. package/docs/research/CRITICAL-REVIEW-RECOGNITION-TUTORING.md +0 -340
  79. package/docs/research/DYNAMIC-VS-SCRIPTED-ANALYSIS.md +0 -291
  80. package/docs/research/EVAL-SYSTEM-ANALYSIS.md +0 -306
  81. package/docs/research/FACTORIAL-RESULTS-2026-01-14.md +0 -301
  82. package/docs/research/IMPLEMENTATION-PLAN-CRITIQUE-RESPONSE.md +0 -1988
  83. package/docs/research/LONGITUDINAL-DYADIC-EVALUATION.md +0 -282
  84. package/docs/research/MULTI-JUDGE-VALIDATION-2026-01-14.md +0 -147
  85. package/docs/research/PAPER-EXTENSION-DYADIC.md +0 -204
  86. package/docs/research/PAPER-UNIFIED.md +0 -659
  87. package/docs/research/PAPER-UNIFIED.pdf +0 -0
  88. package/docs/research/PROMPT-IMPROVEMENTS-2026-01-14.md +0 -356
  89. package/docs/research/SESSION-NOTES-2026-01-11-RECOGNITION-EVAL.md +0 -419
  90. package/docs/research/apa.csl +0 -2133
  91. package/docs/research/archive/PAPER-DRAFT-RECOGNITION-TUTORING.md +0 -1637
  92. package/docs/research/archive/paper-multiagent-tutor.tex +0 -978
  93. package/docs/research/paper-draft/full-paper.md +0 -136
  94. package/docs/research/paper-draft/images/pasted-image-2026-01-24T03-47-47-846Z-d76a7ae2.png +0 -0
  95. package/docs/research/paper-draft/references.bib +0 -515
  96. package/docs/research/transcript-baseline.md +0 -139
  97. package/docs/research/transcript-recognition-multiagent.md +0 -187
  98. package/hooks/useEvalData.ts +0 -625
  99. package/server-init.js +0 -45
  100. package/services/benchmarkService.js +0 -1892
  101. package/types.ts +0 -165
  102. package/utils/haptics.ts +0 -45
@@ -1,124 +0,0 @@
1
- /**
2
- * RecognitionTypeChart Component
3
- *
4
- * Horizontal stacked bar showing distribution of recognition types:
5
- * - Pedagogical (green) - Learning-focused recognition moments
6
- * - Metacognitive (blue) - Reflection on learning process
7
- * - Existential (purple) - Deep identity/meaning moments
8
- */
9
-
10
- import React from 'react';
11
-
12
- interface RecognitionTypeCounts {
13
- pedagogical: number;
14
- metacognitive: number;
15
- existential: number;
16
- }
17
-
18
- interface RecognitionTypeChartProps {
19
- counts: RecognitionTypeCounts;
20
- showLegend?: boolean;
21
- }
22
-
23
- export const RecognitionTypeChart: React.FC<RecognitionTypeChartProps> = ({
24
- counts,
25
- showLegend = true,
26
- }) => {
27
- const total = counts.pedagogical + counts.metacognitive + counts.existential;
28
-
29
- if (total === 0) {
30
- return (
31
- <div className="bg-gray-900/60 backdrop-blur-sm border border-white/5 rounded-xl p-4">
32
- <div className="text-xs text-gray-400 mb-3">Recognition Types</div>
33
- <div className="text-sm text-gray-500 text-center py-4">
34
- No recognition moments recorded
35
- </div>
36
- </div>
37
- );
38
- }
39
-
40
- const percentages = {
41
- pedagogical: (counts.pedagogical / total) * 100,
42
- metacognitive: (counts.metacognitive / total) * 100,
43
- existential: (counts.existential / total) * 100,
44
- };
45
-
46
- const types = [
47
- {
48
- key: 'pedagogical',
49
- label: 'Pedagogical',
50
- count: counts.pedagogical,
51
- percentage: percentages.pedagogical,
52
- color: 'bg-green-500',
53
- textColor: 'text-green-400',
54
- description: 'Learning-focused',
55
- },
56
- {
57
- key: 'metacognitive',
58
- label: 'Metacognitive',
59
- count: counts.metacognitive,
60
- percentage: percentages.metacognitive,
61
- color: 'bg-blue-500',
62
- textColor: 'text-blue-400',
63
- description: 'Process reflection',
64
- },
65
- {
66
- key: 'existential',
67
- label: 'Existential',
68
- count: counts.existential,
69
- percentage: percentages.existential,
70
- color: 'bg-purple-500',
71
- textColor: 'text-purple-400',
72
- description: 'Identity/meaning',
73
- },
74
- ];
75
-
76
- return (
77
- <div className="bg-gray-900/60 backdrop-blur-sm border border-white/5 rounded-xl p-4">
78
- <div className="flex items-center justify-between mb-3">
79
- <div className="text-xs text-gray-400">Recognition Types</div>
80
- <div className="text-xs text-gray-500">{total} total</div>
81
- </div>
82
-
83
- {/* Stacked bar */}
84
- <div className="h-6 rounded-full overflow-hidden flex bg-gray-800">
85
- {types.map(
86
- (type) =>
87
- type.percentage > 0 && (
88
- <div
89
- key={type.key}
90
- className={`${type.color} transition-all duration-500`}
91
- style={{ width: `${type.percentage}%` }}
92
- title={`${type.label}: ${type.count} (${type.percentage.toFixed(1)}%)`}
93
- />
94
- )
95
- )}
96
- </div>
97
-
98
- {/* Legend */}
99
- {showLegend && (
100
- <div className="mt-4 space-y-2">
101
- {types.map((type) => (
102
- <div key={type.key} className="flex items-center justify-between">
103
- <div className="flex items-center gap-2">
104
- <div className={`w-3 h-3 rounded-full ${type.color}`} />
105
- <span className="text-xs text-gray-300">{type.label}</span>
106
- <span className="text-xs text-gray-600">{type.description}</span>
107
- </div>
108
- <div className="flex items-center gap-2">
109
- <span className={`text-xs ${type.textColor} font-medium`}>
110
- {type.count}
111
- </span>
112
- <span className="text-xs text-gray-500">
113
- ({type.percentage.toFixed(0)}%)
114
- </span>
115
- </div>
116
- </div>
117
- ))}
118
- </div>
119
- )}
120
- </div>
121
- );
122
- };
123
-
124
- export default RecognitionTypeChart;