joenathan commited on
Commit
90fbf65
·
verified ·
1 Parent(s): f8a26f0

Soften unvalidated-vs-disproven language to match the measured confound

Browse files
Files changed (1) hide show
  1. app.py +6 -5
app.py CHANGED
@@ -121,8 +121,8 @@ with gr.Blocks(title="Explainable & Auditable Alloy Defect Detection",
121
 
122
  **Read the tabs in order** — they are an argument, not a menu.
123
  The model works and explains itself; that explanation can be hijacked by a
124
- meaningless artifact; and once you test the explanation directly it does not
125
- hold up. What survives is a statistical guarantee, and that is what we would
126
  actually hand an auditor.
127
  """
128
  )
@@ -169,7 +169,8 @@ with gr.Blocks(title="Explainable & Auditable Alloy Defect Detection",
169
  of **random** pixels.
170
 
171
  A faithful explanation should collapse faster when you remove the evidence
172
- it pointed at. This one did the opposite: deleting the top 30% of
 
173
  Grad-CAM-ranked pixels cost only **0.546** of predicted probability, while
174
  deleting a random 30% cost **0.718**.
175
  """
@@ -181,8 +182,8 @@ with gr.Blocks(title="Explainable & Auditable Alloy Defect Detection",
181
  )
182
  gr.Markdown(
183
  """
184
- **Be precise about what failed.** Grad-CAM produced the heatmap; the course's
185
- deletion test then showed that heatmap could not be relied on for this model.
186
  The *attribution* fell short — the *audit* worked exactly as designed,
187
  catching something that 98% accuracy, the per-class breakdown and a
188
  convincing-looking overlay all missed. A diagnostic that only ever confirms
 
121
 
122
  **Read the tabs in order** — they are an argument, not a menu.
123
  The model works and explains itself; that explanation can be hijacked by a
124
+ meaningless artifact; and once you test the explanation directly it cannot be
125
+ confirmed. What survives is a statistical guarantee, and that is what we would
126
  actually hand an auditor.
127
  """
128
  )
 
169
  of **random** pixels.
170
 
171
  A faithful explanation should collapse faster when you remove the evidence
172
+ it pointed at. This one did the opposite — which leaves it **unvalidated,
173
+ not disproven**: deleting the top 30% of
174
  Grad-CAM-ranked pixels cost only **0.546** of predicted probability, while
175
  deleting a random 30% cost **0.718**.
176
  """
 
182
  )
183
  gr.Markdown(
184
  """
185
+ **Be precise about what happened.** Grad-CAM produced the heatmap; the course's
186
+ deletion test then failed to confirm it for this model.
187
  The *attribution* fell short — the *audit* worked exactly as designed,
188
  catching something that 98% accuracy, the per-class breakdown and a
189
  convincing-looking overlay all missed. A diagnostic that only ever confirms