fix(bridge): stop_at_layer never stops on native Bridge

.github/workflows/checks.yml

@@ -102,7 +102,7 @@ jobs:
       - name: Install dependencies
         run: |
           uv lock --check
-          uv sync
+          uv sync --extra inspect
       - name: Authenticate HuggingFace
         if: env.HF_TOKEN != ''
         run: uv run python -c "import os; from huggingface_hub import login; login(token=os.environ['HF_TOKEN'])"
@@ -141,7 +141,7 @@ jobs:
       - name: Install dependencies
         run: |
           uv lock --check
-          uv sync
+          uv sync --extra inspect
       - name: MPS Availability Check
         run: |
           uv run python -c "
@@ -159,9 +159,7 @@ jobs:
           --ignore=tests/unit/model_bridge/test_optimizer_compatibility.py
           --ignore=tests/unit/model_bridge/test_gpt_oss_moe.py
           --ignore=tests/unit/model_bridge/test_component_inspection.py
-          --ignore=tests/unit/model_bridge/test_key_analysis.py
           --ignore=tests/unit/model_bridge/test_benchmark_gated_hooks_fire.py
-          --ignore=tests/unit/model_bridge/test_weight_processing_adapter_paths.py
           --ignore=tests/unit/model_bridge/test_bridge_generate_kv_cache.py
           --ignore=tests/unit/model_bridge/test_bridge_vs_hooked_transformer_patching.py
           --ignore=tests/unit/model_bridge/compatibility/
@@ -296,7 +294,7 @@ jobs:
       - name: Install dependencies
         run: |
           uv lock --check
-          uv sync
+          uv sync --extra inspect
       - name: Authenticate HuggingFace
         if: env.HF_TOKEN != ''
         run: uv run python -c "import os; from huggingface_hub import login; login(token=os.environ['HF_TOKEN'])"

.github/workflows/oracle-parity.yml

@@ -1,17 +1,11 @@
 name: Oracle Parity
 
 # Runs the @pytest.mark.slow oracle parity suite for JacobianLens.
-# Not part of the standard PR gate — requires ~4 GB VRAM and network access.
-# Triggered manually or automatically when the test file or this workflow changes.
+# Not part of the standard PR gate — requires ~4 GB VRAM (or a long CPU run)
+# and network access, so it is manual-trigger only. Run it after changing
+# tests/integration/test_jacobian_lens_oracle_parity.py or this workflow.
 on:
   workflow_dispatch:
-  push:
-    branches:
-      - main
-      - dev*
-    paths:
-      - "tests/integration/test_jacobian_lens_oracle_parity.py"
-      - ".github/workflows/oracle-parity.yml"
 
 permissions:
   contents: read
@@ -21,6 +15,8 @@ jobs:
     name: Oracle Parity Checks
     runs-on: ubuntu-latest
     timeout-minutes: 60
+    env:
+      HAS_HF_TOKEN: ${{ secrets.HF_TOKEN != '' }}
     steps:
       - uses: actions/checkout@v4
 
@@ -46,7 +42,7 @@ jobs:
           uv pip install "git+https://github.com/anthropics/jacobian-lens.git@${ORACLE_COMMIT}"
 
       - name: Authenticate HuggingFace
-        if: env.HF_TOKEN != ''
+        if: ${{ env.HAS_HF_TOKEN == 'true' }}
         run: uv run python -c "import os; from huggingface_hub import login; login(token=os.environ['HF_TOKEN'])"
         env:
           HF_TOKEN: ${{ secrets.HF_TOKEN }}

AGENTS.md

@@ -22,7 +22,7 @@ Sub-folder rules: [tests/AGENTS.md](tests/AGENTS.md) · [supported_architectures
 
 ## 1. What this repo is
 
-**TransformerLens** — mechanistic-interpretability library. Loads 9,000+ models across 50+ architecture families (see [supported_models.json](transformer_lens/tools/model_registry/data/supported_models.json)) and exposes internal activations through a hook system for caching, editing, and ablating intermediate state. Built on HuggingFace `transformers`.
+**TransformerLens** — mechanistic-interpretability library. Loads 15,000+ models across 140+ architecture families (see [supported_models.json](transformer_lens/tools/model_registry/data/supported_models.json)) and exposes internal activations through a hook system for caching, editing, and ablating intermediate state. Built on HuggingFace `transformers`.
 
 ## 2. Two systems live in this repo
 
@@ -31,9 +31,9 @@ Sub-folder rules: [tests/AGENTS.md](tests/AGENTS.md) · [supported_architectures
 | **`TransformerBridge`** | v3 — default for new work | [transformer_lens/model_bridge/](transformer_lens/model_bridge/) | Raw HF weights by default; `bridge.enable_compatibility_mode()` for HT-equivalent | [transformer_lens/tools/model_registry/data/supported_models.json](transformer_lens/tools/model_registry/data/supported_models.json) |
 | **`HookedTransformer`** | Legacy, maintenance mode, deprecated in 3.0 | [transformer_lens/HookedTransformer.py](transformer_lens/HookedTransformer.py) + [transformer_lens/components/](transformer_lens/components/) | Folds LayerNorm + centres weights → does NOT match HF | [transformer_lens/supported_models.py](transformer_lens/supported_models.py) (**HT-only**) |
 
-> ⚠ The **HookedTransformer acceptance suite is quarantined** ([test_hooked_transformer.py](tests/acceptance/test_hooked_transformer.py), [test_hooked_encoder.py](tests/acceptance/test_hooked_encoder.py), [test_hooked_encoder_decoder.py](tests/acceptance/test_hooked_encoder_decoder.py); see [QUARANTINES.md](tests/QUARANTINES.md)). HT changes land untested at the acceptance level — extra manual care required.
+> ⚠ `HookedTransformer` has no acceptance suite of its own — `test_hooked_transformer.py` was removed by the reanchoring in #1603, which moved Bridge tests onto frozen goldens. Changes to `HookedTransformer` itself land untested at the acceptance level; extra manual care required. The `HookedEncoder` (BERT) and `HookedEncoderDecoder` (T5) acceptance suites do cover those two classes. See [QUARANTINES.md](tests/QUARANTINES.md) for the remaining skips.
 
-Bridge architecture-adapter pattern: each HF architecture has one file in [supported_architectures/](transformer_lens/model_bridge/supported_architectures/) mapping HF module paths to canonical names. Bridge hooks are architecture-native (e.g. `blocks.{i}.hook_out`); HT-style aliases live in [bridge.py](transformer_lens/model_bridge/bridge.py).
+Bridge architecture-adapter pattern: each HF architecture has one file in [supported_architectures/](transformer_lens/model_bridge/supported_architectures/) mapping HF module paths to canonical names. Bridge hooks are architecture-native (e.g. `blocks.{i}.hook_out`); HT-style aliases live in [bridge_core.py](transformer_lens/model_bridge/bridge_core.py).
 
 **Mirroring rule:** if you change `HookedTransformer` behaviour that has a `TransformerBridge` counterpart, update both in the same PR. [supported_models.py](transformer_lens/supported_models.py) is HT-only — Bridge-only models go in the Bridge registry data file.
 
@@ -107,7 +107,7 @@ Python: **>=3.10, <4.0**. CI tests 3.10, 3.11, 3.12. Format/type/docstring check
 ## 5. Hook naming — HT vs Bridge
 
 - **HT canonical**: uniform across architectures — `hook_embed`, `blocks.{i}.hook_resid_pre`, `blocks.{i}.attn.hook_q`, `blocks.{i}.hook_resid_post`.
-- **Bridge-native**: architecture-shaped — `blocks.{i}.hook_out`, `blocks.{i}.attn.q.hook_out`. HT aliases registered via `build_alias_to_canonical_map()` in [bridge.py](transformer_lens/model_bridge/bridge.py).
+- **Bridge-native**: architecture-shaped — `blocks.{i}.hook_out`, `blocks.{i}.attn.q.hook_out`. HT aliases registered via `build_alias_to_canonical_map()` in [bridge_core.py](transformer_lens/model_bridge/bridge_core.py).
 
 Prefer Bridge-native names in new code. Raw-HF-forward drivers comparing against `boot_transformers` must match its load configuration (fp32, eager attention) and probe for optional features like `resid_mid` rather than assume.
 

README.md

@@ -20,7 +20,7 @@ interpretability](https://distill.pub/2020/circuits/zoom-in/) of GPT-2 Style lan
 goal of mechanistic interpretability is to take a trained model and reverse engineer the algorithms
 the model learned during training from its weights.
 
-TransformerLens lets you load in 9,000+ open source language models across 50+ architecture families,
+TransformerLens lets you load in 15,000+ open source language models across 140+ architecture families,
 and exposes the internal activations of the model to you. You can cache any internal activation in the
 model, and add in functions to edit, remove or replace these activations as the model runs.
 
@@ -52,7 +52,7 @@ logits, activations = bridge.run_with_cache("Hello World")
 
 > Gated models (Llama, Mistral, Gemma, ...) require `HF_TOKEN` in your environment. See [Environment Variables](https://TransformerLensOrg.github.io/TransformerLens/content/getting_started.html#environment-variables) for the full list.
 
-`TransformerBridge` is the recommended 3.0 path and supports 9,000+ models across 50+ architecture families (see [`supported_models.json`](transformer_lens/tools/model_registry/data/supported_models.json) for the full inventory). By default it preserves raw HuggingFace weights – logits and activations match HF, *not* legacy `HookedTransformer` (which folds LayerNorm and centers weights by default). Call `bridge.enable_compatibility_mode()` after booting for HookedTransformer-equivalent numerics. The legacy `HookedTransformer.from_pretrained` API is still available but deprecated — see the [Migrating to TransformerLens 3](https://TransformerLensOrg.github.io/TransformerLens/content/migrating_to_v3.html) guide.
+`TransformerBridge` is the recommended path and supports 15,000+ models across 140+ architecture families (see [`supported_models.json`](transformer_lens/tools/model_registry/data/supported_models.json) for the full inventory). By default it preserves raw HuggingFace weights – logits and activations match HF, *not* legacy `HookedTransformer` (which folds LayerNorm and centers weights by default). Call `bridge.enable_compatibility_mode()` after booting for HookedTransformer-equivalent numerics. The legacy `HookedTransformer.from_pretrained` API is still available but deprecated — see the [Migrating to TransformerLens 3](https://TransformerLensOrg.github.io/TransformerLens/content/migrating_to_v3.html) guide.
 
 ## Key Tutorials