More tests for the negative case

tests/unit/analysis/test_hooked_sae_transformer.py

@@ -581,3 +581,68 @@ def test_run_with_cache_with_saes_use_error_term_true(
         cache_with_sae[hooked_sae.cfg.hook_name + ".hook_sae_output"],
         atol=1e-5,
     )
+
+
+def test_add_sae_with_use_error_term_false(
+    model: HookedSAETransformer,
+    hooked_sae: SAE,
+):
+    """Verifies that add_sae with use_error_term=False changes the model output."""
+    # Get output without SAE
+    output_without_sae = get_logits(model(prompt))
+
+    # Add SAE with use_error_term=False
+    model.add_sae(hooked_sae, use_error_term=False)
+    output_with_sae = get_logits(model(prompt))
+
+    # Compare outputs - they should be different
+    assert not torch.allclose(output_without_sae, output_with_sae, atol=1e-5)
+
+    # Clean up
+    model.reset_saes()
+
+
+def test_run_with_saes_use_error_term_false(
+    model: HookedSAETransformer,
+    hooked_sae: SAE,
+):
+    """Verifies that run_with_saes with use_error_term=False changes the model output."""
+    # Get output without SAE
+    output_without_sae = get_logits(model(prompt))
+
+    # Run with SAE and use_error_term=False
+    output_with_sae = get_logits(
+        model.run_with_saes(prompt, saes=[hooked_sae], use_error_term=False)
+    )
+
+    # Compare outputs - they should be different
+    assert not torch.allclose(output_without_sae, output_with_sae, atol=1e-5)
+
+
+def test_run_with_cache_with_saes_use_error_term_false(
+    model: HookedSAETransformer,
+    hooked_sae: SAE,
+):
+    """Verifies that run_with_cache_with_saes with use_error_term=False changes the model output."""
+    # Get output without SAE
+    output_without_sae, cache_without_sae = model.run_with_cache(prompt)
+    output_without_sae = get_logits(output_without_sae)
+
+    # Run with SAE and use_error_term=False
+    output_with_sae, cache_with_sae = model.run_with_cache_with_saes(
+        prompt, saes=[hooked_sae], use_error_term=False
+    )
+    output_with_sae = get_logits(output_with_sae)
+
+    # Compare outputs - they should be different
+    assert not torch.allclose(output_without_sae, output_with_sae, atol=1e-5)
+
+    # Verify that the cache contains the SAE activations
+    assert hooked_sae.cfg.hook_name + ".hook_sae_acts_post" in cache_with_sae
+
+    # Verify that the activations at the SAE hook point are different in both caches
+    assert not torch.allclose(
+        cache_without_sae[hooked_sae.cfg.hook_name],
+        cache_with_sae[hooked_sae.cfg.hook_name + ".hook_sae_output"],
+        atol=1e-5,
+    )