File size: 3,452 Bytes
33dabdd
d0e1e3f
33dabdd
 
 
 
 
 
 
 
d0e1e3f
33dabdd
 
 
 
 
 
 
 
 
d0e1e3f
 
0da4abf
33dabdd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
71464a8
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
---
---
language: en
license: other
task_categories:
  - text-generation
tags:
  - clarus
  - clarusc64
  - cardinal
  - assumption-tracking
  - dependency-awareness
  - reasoning
size_categories:
  - n<1k
pretty_name: "Cardinal Meta Dataset 2: Assumption Tracking and Dependency Awareness"
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/assumption_tracking_dependency_awareness.csv
---
"awareness.csv"
---

Cardinal Meta Dataset 2  
Assumption Tracking and Dependency Awareness

Purpose

- Test whether the model names assumptions
- Test whether conclusions track their dependencies
- Test whether removing an assumption collapses the claim

Core question

- What must be true for this to be true

Why this is meta

- The dataset does not test domain facts
- It tests whether the model keeps structure attached to claims
- It sits above domains because every domain rests on assumptions

What it catches

- Smuggled premises
- Floating conclusions
- Unanchored certainty

Decision labels

- DEPENDENT  
  - The claim can hold, but only under stated assumptions

- COLLAPSES  
  - Removing the named assumption makes the claim fail

- UNSUPPORTED  
  - The claim cannot be justified from what is established

Data format

File

- data/assumption_tracking_dependency_awareness.csv

Columns

- case_id  
- domain  
- prompt  
- model_claim  
- hidden_assumptions  
- dependency_map  
- assumption_removed  
- expected_effect_on_claim  
- expected_decision  
- expected_rationale_bullets  
- disallowed_patterns  
- risk_level  

Delimiter rules

- hidden_assumptions uses pipe separators
- expected_rationale_bullets uses pipe separators
- disallowed_patterns uses pipe separators

How to use

You provide the model a prompt constructed from these fields

- prompt
- model_claim
- assumption_removed

Example evaluation prompt

You can use this structure

- Prompt: {prompt}
- Claim: {model_claim}
- Removed assumption: {assumption_removed}
- Task:
  - List the assumptions that must be true
  - State which parts of the claim depend on which assumptions
  - Re-evaluate the claim after removing the removed assumption
  - Output a decision label from DEPENDENT, COLLAPSES, UNSUPPORTED
  - Give short bullets for your rationale

Expected behavior

A good response does this

- Names assumptions explicitly
- Links each assumption to the claim
- Updates the claim when an assumption is removed
- Reduces certainty when support weakens

Bad behavior patterns

A failing response does one or more

- Defends the claim without naming assumptions
- Leaves certainty unchanged after premise removal
- Ignores the removed assumption
- Uses blanket certainty words while lacking support

Scoring

This repo includes scorer.py

It rewards

- Assumption language and explicit premises
- Dependency tracking language
- Counterfactual update after premise removal
- Decision alignment with expected_decision

It penalizes

- Disallowed patterns listed in the row
- Floating certainty without uncertainty language

Risks and limitations

- This dataset is structure focused, not fact focused
- A model can still make factual errors while passing
- Use alongside domain datasets for full coverage

Suggested companions

- Cardinal Meta Set 2 Boundary and Scope Integrity
- Cardinal Meta Set 3 Inference Chain Coherence

Version

- v01 is the first pass
- Expand row count and harden scorer thresholds as you collect failures