From e82bf13f07076da7c67e7407d3d503cab2bf63d6 Mon Sep 17 00:00:00 2001
From: Luca Belli <129434630+sator-labs@users.noreply.github.com>
Date: Thu, 1 Oct 2026 16:12:51 -0700
Subject: [PATCH 1/5] refactor: move run_for_user_models into the generation
 package

`vera generate` imported `run_for_user_models` from the legacy root
`generate.py`, so the new CLI depended on a deprecated script. Move it and
its `_legacy_model_config` helper into `generate_conversations/main.py`,
where it calls `run_generation` directly instead of going through the
legacy `main` wrapper, and export it from the package.

Tests that patched `generate.main` for the `vera` path now patch
`generate_conversations.main.run_generation`; the keyword arguments are the
same, since `main` only forwarded them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---
 generate.py                          | 80 ----------------------------
 generate_conversations/__init__.py   |  4 +-
 generate_conversations/main.py       | 79 +++++++++++++++++++++++++++
 tests/integration/test_cli_parity.py |  9 ++--
 tests/unit/test_vera_cli.py          | 14 ++---
 vera_cli/generate.py                 |  4 +-
 6 files changed, 96 insertions(+), 94 deletions(-)

diff --git a/generate.py b/generate.py
index 1a224337..596bf7f2 100644
--- a/generate.py
+++ b/generate.py
@@ -10,7 +10,6 @@
 
 from generate_conversations import run_generation
 from llm_clients.llm_interface import DEFAULT_START_PROMPT
-from utils.config_schema import GenerationConfig, ModelSpec
 from utils.debug import set_debug
 from utils.rubric_manifest import (
     load_manifest_persona_context_template,
@@ -63,85 +62,6 @@ async def main(
     )
 
 
-async def run_for_user_models(
-    generation: GenerationConfig,
-    *,
-    max_personas: Optional[int],
-    into: Optional[str] = None,
-) -> List[str]:
-    """Run one generation per user-side model, sequentially.
-
-    STOPGAP. This exists so the CLI hands the generation domain a resolved
-    `GenerationConfig` and lets the domain decide how a multi-model request
-    expands into runs, instead of the CLI looping over models and flattening
-    each one into the legacy dict shape itself.
-
-    Deliberately thin: sequential, with no shared run identity across models.
-    Sequential is not incidental — `max_concurrent` is applied *within* one
-    generation, so running models concurrently would silently multiply the
-    caller's concurrency cap against the same provider.
-
-    Placed beside `main` so it moves into the permanent `generate/` package in
-    the same atomic change (see docs/architecture.md). The real fix is for the
-    generation domain to accept these types natively, at which point
-    `_legacy_model_config` below and this wrapper both disappear.
-
-    Returns the run folder each user model wrote to, in the order the models
-    were given. `main` has always returned its output folder and this wrapper
-    used to drop it; `vera pipeline` needs it, because chaining generation into
-    judging means knowing where the transcripts landed without re-deriving it
-    from the filesystem or scraping it out of a log. `vera generate` ignores
-    the return.
-    """
-    run_folders: List[str] = []
-    for user in generation.user:
-        _, run_folder = await main(
-            persona_model_config=_legacy_model_config(user),
-            # Only the agent side may carry `name`; see `_legacy_model_config`.
-            agent_model_config=_legacy_model_config(generation.chatbot)
-            | {"name": generation.chatbot.name},
-            persona_files=list(generation.personas),
-            persona_extra_run_params=dict(user.extra_params),
-            agent_extra_run_params=dict(generation.chatbot.extra_params),
-            max_turns=generation.turns,
-            runs_per_prompt=user.repeats,
-            persona_names=None,
-            verbose=True,
-            # `into` names an existing run folder to continue; without it,
-            # `output` is a parent to mint a new run under.
-            output_folder=into or generation.output,
-            run_id=None,
-            max_concurrent=generation.max_concurrent,
-            max_total_words=generation.max_total_words,
-            max_personas=max_personas,
-            persona_speaks_first=generation.persona_speaks_first,
-            session_types=generation.sessions,
-            resume=into is not None,
-            persona_context_template_path=generation.persona_context_template,
-        )
-        run_folders.append(run_folder)
-    return run_folders
-
-
-def _legacy_model_config(model: ModelSpec) -> Dict[str, Any]:
-    """Flatten a `ModelSpec` into the dict shape `main` still expects.
-
-    STOPGAP translation shim, deleted once the generation domain takes
-    `ModelSpec` directly. `repeats` is dropped because it is passed separately
-    as `runs_per_prompt`.
-
-    Callers add `name` for the agent side only. That asymmetry is not this
-    function's to decide, and it is not about output naming — run folders are
-    built from `model` (`utils/naming.py:model_token_for_run_folder`). `name` is
-    the provider's display name, defaulted to `"Provider"` in
-    `generate_conversations/runner.py`. It is legal on the agent config only
-    because the runner filters reserved keys out of that one before splatting
-    it into `LLMFactory.create_llm`, while splatting the persona config raw —
-    so a `name` key there would collide. Fix belongs in the runner.
-    """
-    return {**model.extra_params, "model": model.name}
-
-
 async def resolve_persona_inputs(manifest: str) -> tuple[List[str], str]:
     """Resolve generation inputs for legacy callers of ``generate.py``."""
     persona_files = await load_manifest_personas(manifest)
diff --git a/generate_conversations/__init__.py b/generate_conversations/__init__.py
index 73b7f5f1..13032b88 100644
--- a/generate_conversations/__init__.py
+++ b/generate_conversations/__init__.py
@@ -1,6 +1,6 @@
 """Generate Conversations Package - LLM Conversation Simulation"""
 
-from .main import run_generation
+from .main import run_for_user_models, run_generation
 from .runner import ConversationRunner
 
-__all__ = ["ConversationRunner", "run_generation"]
+__all__ = ["ConversationRunner", "run_for_user_models", "run_generation"]
diff --git a/generate_conversations/main.py b/generate_conversations/main.py
index 5561e40e..66c4fa4b 100644
--- a/generate_conversations/main.py
+++ b/generate_conversations/main.py
@@ -12,6 +12,7 @@
 from datetime import datetime
 from typing import Any, Dict, List, Optional
 
+from utils.config_schema import GenerationConfig, ModelSpec
 from utils.naming import (
     build_generation_run_folder_name,
     model_token_for_run_folder,
@@ -147,3 +148,81 @@ async def run_generation(
         print(message)
 
     return results, output_folder
+
+
+async def run_for_user_models(
+    generation: GenerationConfig,
+    *,
+    max_personas: Optional[int],
+    into: Optional[str] = None,
+) -> List[str]:
+    """Run one generation per user-side model, sequentially.
+
+    STOPGAP. This exists so the CLI hands the generation domain a resolved
+    `GenerationConfig` and lets the domain decide how a multi-model request
+    expands into runs, instead of the CLI looping over models and flattening
+    each one into the legacy dict shape itself.
+
+    Deliberately thin: sequential, with no shared run identity across models.
+    Sequential is not incidental — `max_concurrent` is applied *within* one
+    generation, so running models concurrently would silently multiply the
+    caller's concurrency cap against the same provider.
+
+    The real fix is for the generation domain to accept these types natively,
+    at which point `_legacy_model_config` below and this wrapper both
+    disappear.
+
+    Returns the run folder each user model wrote to, in the order the models
+    were given. `run_generation` has always returned its output folder and
+    this wrapper used to drop it; `vera pipeline` needs it, because chaining
+    generation into judging means knowing where the transcripts landed
+    without re-deriving it from the filesystem or scraping it out of a log.
+    `vera generate` ignores the return.
+    """
+    run_folders: List[str] = []
+    for user in generation.user:
+        _, run_folder = await run_generation(
+            persona_model_config=_legacy_model_config(user),
+            # Only the agent side may carry `name`; see `_legacy_model_config`.
+            agent_model_config=_legacy_model_config(generation.chatbot)
+            | {"name": generation.chatbot.name},
+            persona_files=list(generation.personas),
+            persona_extra_run_params=dict(user.extra_params),
+            agent_extra_run_params=dict(generation.chatbot.extra_params),
+            max_turns=generation.turns,
+            runs_per_prompt=user.repeats,
+            persona_names=None,
+            verbose=True,
+            # `into` names an existing run folder to continue; without it,
+            # `output` is a parent to mint a new run under.
+            output_folder=into or generation.output,
+            run_id=None,
+            max_concurrent=generation.max_concurrent,
+            max_total_words=generation.max_total_words,
+            max_personas=max_personas,
+            persona_speaks_first=generation.persona_speaks_first,
+            session_types=generation.sessions,
+            resume=into is not None,
+            persona_context_template_path=generation.persona_context_template,
+        )
+        run_folders.append(run_folder)
+    return run_folders
+
+
+def _legacy_model_config(model: ModelSpec) -> Dict[str, Any]:
+    """Flatten a `ModelSpec` into the dict shape `run_generation` expects.
+
+    STOPGAP translation shim, deleted once the generation domain takes
+    `ModelSpec` directly. `repeats` is dropped because it is passed separately
+    as `runs_per_prompt`.
+
+    Callers add `name` for the agent side only. That asymmetry is not this
+    function's to decide, and it is not about output naming — run folders are
+    built from `model` (`utils/naming.py:model_token_for_run_folder`). `name` is
+    the provider's display name, defaulted to `"Provider"` in
+    `generate_conversations/runner.py`. It is legal on the agent config only
+    because the runner filters reserved keys out of that one before splatting
+    it into `LLMFactory.create_llm`, while splatting the persona config raw —
+    so a `name` key there would collide. Fix belongs in the runner.
+    """
+    return {**model.extra_params, "model": model.name}
diff --git a/tests/integration/test_cli_parity.py b/tests/integration/test_cli_parity.py
index 33848a3a..a847f412 100644
--- a/tests/integration/test_cli_parity.py
+++ b/tests/integration/test_cli_parity.py
@@ -72,14 +72,15 @@
 def calls(monkeypatch: pytest.MonkeyPatch) -> Iterator[list[dict[str, Any]]]:
     """Stub the domain boundary and collect the kwargs each CLI sends it.
 
-    Both patches are needed: `generate` bound `run_generation` at import time,
-    while a fresh `runpy` exec of `generate.py` re-imports it from the package.
+    Both patches are needed: `run_for_user_models` (the `vera` path) calls the
+    `run_generation` bound in `generate_conversations.main`, while a fresh
+    `runpy` exec of `generate.py` re-imports it from the package.
 
     `chdir` matters because the legacy CLI resolves paths against the working
     directory, so its manifest argument only resolves from the repository root.
     """
-    import generate
     import generate_conversations
+    import generate_conversations.main
 
     recorded: list[dict[str, Any]] = []
 
@@ -89,7 +90,7 @@ async def record(**kwargs: Any) -> tuple[list[dict[str, Any]], str]:
 
     monkeypatch.chdir(REPO_ROOT)
     with (
-        patch.object(generate, "run_generation", record),
+        patch.object(generate_conversations.main, "run_generation", record),
         patch.object(generate_conversations, "run_generation", record),
     ):
         yield recorded
diff --git a/tests/unit/test_vera_cli.py b/tests/unit/test_vera_cli.py
index 3316f603..23be43a9 100644
--- a/tests/unit/test_vera_cli.py
+++ b/tests/unit/test_vera_cli.py
@@ -8,7 +8,7 @@
 
 import pytest
 
-import generate as generation_domain
+import generate_conversations.main as generation_domain
 import vera
 from utils.config_schema import InvocationConfig, ModelSpec
 from vera_cli import (
@@ -222,7 +222,7 @@ def test_cli_defaults_resolve_before_print(capsys: pytest.CaptureFixture) -> Non
 
 def test_generation_delegates_every_value_to_generate_main() -> None:
     with patch.object(
-        generation_domain, "main", new_callable=AsyncMock
+        generation_domain, "run_generation", new_callable=AsyncMock
     ) as generate_main:
         generate_main.return_value = ([], "output/generated")
         result = vera.main(
@@ -273,7 +273,7 @@ def test_generation_delegates_every_value_to_generate_main() -> None:
 def test_each_user_model_gets_one_generate_main_call(tmp_path: Path) -> None:
     manifest = _write_target(tmp_path, persona_count=2)
     with patch.object(
-        generation_domain, "main", new_callable=AsyncMock
+        generation_domain, "run_generation", new_callable=AsyncMock
     ) as generate_main:
         generate_main.return_value = ([], "output/generated")
         vera.main(
@@ -452,7 +452,9 @@ def test_incomplete_target_fails_before_dispatch(tmp_path: Path) -> None:
     manifest.write_text(json.dumps({"personas": ["personas.tsv"]}), encoding="utf-8")
 
     with (
-        patch.object(generation_domain, "main", new_callable=AsyncMock) as runner,
+        patch.object(
+            generation_domain, "run_generation", new_callable=AsyncMock
+        ) as runner,
         pytest.raises(SystemExit) as error,
     ):
         vera.main(
@@ -534,7 +536,7 @@ def test_into_continues_an_existing_run_folder(tmp_path: Path) -> None:
     run_folder = tmp_path / "p_user__a_bot__t5__r1"
     run_folder.mkdir()
     with patch.object(
-        generation_domain, "main", new_callable=AsyncMock
+        generation_domain, "run_generation", new_callable=AsyncMock
     ) as generate_main:
         generate_main.return_value = ([], "output/generated")
         result = vera.main(
@@ -610,7 +612,7 @@ def test_into_is_invocation_only_so_it_may_accompany_config(tmp_path: Path) -> N
     run_folder.mkdir()
     config = _write_config(tmp_path, _generation_config())
     with patch.object(
-        generation_domain, "main", new_callable=AsyncMock
+        generation_domain, "run_generation", new_callable=AsyncMock
     ) as generate_main:
         generate_main.return_value = ([], "output/generated")
         result = vera.main(
diff --git a/vera_cli/generate.py b/vera_cli/generate.py
index d34e375c..41b1a89c 100644
--- a/vera_cli/generate.py
+++ b/vera_cli/generate.py
@@ -25,7 +25,7 @@
 from pathlib import Path
 from typing import Any
 
-from generate import run_for_user_models
+from generate_conversations import run_for_user_models
 from utils.config_schema import GenerationConfig, InvocationConfig, ModelSpec, RunConfig
 from utils.debug import set_debug
 from utils.utils import parse_key_value_list
@@ -342,7 +342,7 @@ async def _execute(run_configs: list[RunConfig]) -> list[str]:
 
     Expanding a run's user models into individual generations belongs to the
     domain, not the CLI, so this passes the resolved `GenerationConfig` straight
-    through to `generate.run_for_user_models` — a stopgap wrapper, see its
+    through to `generate_conversations.run_for_user_models` — a stopgap wrapper, see its
     docstring. Targets stay sequential here: `--target all` runs share chatbot,
     user models, turns, and repeats, and run folder names carry only
     second-granularity timestamps, so concurrent starts would collide.

From 62671840c92d81add5bd58c482e9ebe681e794a3 Mon Sep 17 00:00:00 2001
From: Luca Belli <129434630+sator-labs@users.noreply.github.com>
Date: Thu, 1 Oct 2026 16:13:17 -0700
Subject: [PATCH 2/5] chore: move the legacy scripts into legacy/ (rename only)

Pure renames, so git history follows each file. The tree does not run
between this commit and the next, which updates every reference; skip it
when bisecting.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---
 generate.py => legacy/generate.py                    | 0
 judge.py => legacy/judge.py                          | 0
 run_pipeline.py => legacy/run_pipeline.py            | 0
 {scripts => legacy}/run_recommended_vera_pipeline.sh | 0
 4 files changed, 0 insertions(+), 0 deletions(-)
 rename generate.py => legacy/generate.py (100%)
 rename judge.py => legacy/judge.py (100%)
 rename run_pipeline.py => legacy/run_pipeline.py (100%)
 rename {scripts => legacy}/run_recommended_vera_pipeline.sh (100%)

diff --git a/generate.py b/legacy/generate.py
similarity index 100%
rename from generate.py
rename to legacy/generate.py
diff --git a/judge.py b/legacy/judge.py
similarity index 100%
rename from judge.py
rename to legacy/judge.py
diff --git a/run_pipeline.py b/legacy/run_pipeline.py
similarity index 100%
rename from run_pipeline.py
rename to legacy/run_pipeline.py
diff --git a/scripts/run_recommended_vera_pipeline.sh b/legacy/run_recommended_vera_pipeline.sh
similarity index 100%
rename from scripts/run_recommended_vera_pipeline.sh
rename to legacy/run_recommended_vera_pipeline.sh

From f39bdeb007598a151bbed03ecea6dd7e97359447 Mon Sep 17 00:00:00 2001
From: Luca Belli <129434630+sator-labs@users.noreply.github.com>
Date: Thu, 1 Oct 2026 16:17:19 -0700
Subject: [PATCH 3/5] chore: run the legacy scripts from legacy/ and update
 every reference

The legacy entry points now run as modules from the repository root:
`uv run python -m legacy.generate`, `-m legacy.judge`, `-m legacy.score`,
and `-m legacy.run_pipeline`.

- Move `judge/score.py`'s CLI (`main`) to `legacy/score.py`; the scoring
  domain (`run_scoring`) stays in `judge/score.py`.
- Add `legacy/README.md`: deprecated, may be removed at any time.
- `legacy/run_pipeline.py` imports `legacy.generate` and loads its sibling
  `judge.py` by path relative to itself, not to the working directory.
- Update tests, shell scripts, Dockerfile `CMD` (now `vera.py --help`),
  CODEOWNERS, the Claude slash commands and permissions, and the docs.

With `generate.py` out of the root, `generate_conversations/` can be
renamed to `generate/` without shadowing.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---
 .claude/commands/run-generator.md             |   2 +-
 .claude/commands/run-judge.md                 |   2 +-
 .claude/commands/verify.md                    |   2 +-
 .claude/settings.json                         |   4 +-
 .github/CODEOWNERS                            |   2 +-
 AGENTS.md                                     |   2 +-
 Dockerfile                                    |   2 +-
 README.md                                     |   2 +-
 docs/architecture.md                          |  16 +-
 docs/evaluating.md                            |   4 +-
 docs/legacy-scripts.md                        |  40 ++---
 judge/score.py                                |  65 +-------
 legacy/README.md                              |   5 +
 legacy/__init__.py                            |   1 +
 legacy/run_pipeline.py                        |  12 +-
 legacy/run_recommended_vera_pipeline.sh       |   6 +-
 legacy/score.py                               |  68 ++++++++
 llm_clients/config.py                         |   2 +-
 scripts/pool_vera_scores.py                   |   2 +-
 scripts/run_combinations.sh                   |   2 +-
 scripts/score_publication_evaluations.sh      |   2 +-
 tests/integration/test_cli_parity.py          |   5 +-
 .../test_judge_against_clinician_ratings.py   |   4 +-
 tests/integration/test_pipeline.py            | 150 +++++++++---------
 tests/integration/test_score_parity.py        |   6 +-
 tests/integration/test_scoring.py             |  22 ++-
 .../test_generate_cli.py                      |   2 +-
 tests/unit/judge/test_judge_cli.py            |   4 +-
 tests/unit/test_vera_pipeline.py              |   2 +-
 vera_cli/pipeline.py                          |   2 +-
 30 files changed, 234 insertions(+), 206 deletions(-)
 create mode 100644 legacy/README.md
 create mode 100644 legacy/__init__.py
 create mode 100644 legacy/score.py

diff --git a/.claude/commands/run-generator.md b/.claude/commands/run-generator.md
index c96b9972..138f92ff 100644
--- a/.claude/commands/run-generator.md
+++ b/.claude/commands/run-generator.md
@@ -1,4 +1,4 @@
-Help the user run the conversation generator (generate.py):
+Help the user run the conversation generator (`uv run python vera.py generate`; flags in docs/cli.md):
 
 1. Ask the user for parameters:
    - Model to use (default: claude-3-7-sonnet)
diff --git a/.claude/commands/run-judge.md b/.claude/commands/run-judge.md
index 95ba68b2..450c1b38 100644
--- a/.claude/commands/run-judge.md
+++ b/.claude/commands/run-judge.md
@@ -1,4 +1,4 @@
-Help the user run the conversation evaluator (judge.py):
+Help the user run the conversation evaluator (`uv run python vera.py judge`; flags in docs/cli.md):
 
 1. Ask the user for parameters:
    - Input directory (default: output/conversations)
diff --git a/.claude/commands/verify.md b/.claude/commands/verify.md
index 998863bc..f29c8cca 100644
--- a/.claude/commands/verify.md
+++ b/.claude/commands/verify.md
@@ -1,6 +1,6 @@
 Verify that a code change works correctly without hitting live APIs.
 
-IMPORTANT: Never run `generate.py`, `judge.py`, or `run_pipeline.py` directly — these make live LLM API calls. Verification must use the test suite only.
+IMPORTANT: Never run the `vera` commands or the scripts in `legacy/` directly — these make live LLM API calls. Verification must use the test suite only.
 
 Steps:
 1. Identify what changed (git diff --stat) to understand the scope
diff --git a/.claude/settings.json b/.claude/settings.json
index 45c78a2f..9e09a161 100644
--- a/.claude/settings.json
+++ b/.claude/settings.json
@@ -10,8 +10,8 @@
       "Bash(.venv/bin/ruff check --fix .)",
       "Bash(.venv/bin/pyright:*)",
       "Bash(.venv/bin/pytest:*)",
-      "Bash(python3 generate.py:*)",
-      "Bash(python3 judge.py:*)",
+      "Bash(python3 -m legacy.generate:*)",
+      "Bash(python3 -m legacy.judge:*)",
       "Bash(python3 -m pytest:*)",
       "Bash(pytest:*)",
       "Bash(git status)",
diff --git a/.github/CODEOWNERS b/.github/CODEOWNERS
index 8815575b..accdb1fe 100644
--- a/.github/CODEOWNERS
+++ b/.github/CODEOWNERS
@@ -4,7 +4,7 @@
 /llm_clients/ @SpringCare/vera-mh-maintainers
 /judge/ @SpringCare/vera-mh-maintainers
 /generate/ @SpringCare/vera-mh-maintainers
-/run_pipeline.py @SpringCare/vera-mh-maintainers
+/legacy/ @SpringCare/vera-mh-maintainers
 /openspec/ @SpringCare/vera-mh-maintainers
 /pyproject.toml @SpringCare/vera-mh-maintainers
 
diff --git a/AGENTS.md b/AGENTS.md
index 1540c0b8..c3576cf2 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -134,7 +134,7 @@ uv run pyright
 pre-commit run --all-files
 ```
 
-`generate.py`, `judge.py`, `run_pipeline.py`, `judge/score.py`, and `scripts/run_recommended_vera_pipeline.sh` are deprecated and will be removed; don't build on them. Their flags and `vera` equivalents are in [docs/legacy-scripts.md](docs/legacy-scripts.md).
+The scripts in [`legacy/`](legacy/) (`generate.py`, `judge.py`, `score.py`, `run_pipeline.py`, `run_recommended_vera_pipeline.sh`) are deprecated and may be removed at any time; don't build on them, and never import them from `vera_cli/` or the domain packages. Their flags and `vera` equivalents are in [docs/legacy-scripts.md](docs/legacy-scripts.md).
 
 Use dated model IDs (e.g. `claude-sonnet-4-5-20250929`) as in README; shorthand aliases may not resolve.
 
diff --git a/Dockerfile b/Dockerfile
index fd4c91ab..264b4c35 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -11,4 +11,4 @@ RUN uv sync --frozen
 
 COPY . .
 
-CMD ["python", "generate.py"]
\ No newline at end of file
+CMD ["python", "vera.py", "--help"]
\ No newline at end of file
diff --git a/README.md b/README.md
index ca0cce54..9290f51a 100644
--- a/README.md
+++ b/README.md
@@ -89,7 +89,7 @@ pre-commit install             # optional: format and lint on commit
 
 ## Legacy scripts
 
-The pre-2.0 entry points—`generate.py`, `judge.py`, `run_pipeline.py`, `judge/score.py`, and `scripts/run_recommended_vera_pipeline.sh`—still work, but they are **deprecated and will be removed soon**. Their flags differ from the `vera` CLI (for example `-c` means `--max-concurrent` in `generate.py`). Use `vera` for new work; if you still depend on a script, see [docs/legacy-scripts.md](docs/legacy-scripts.md) for its options and the `vera` equivalent.
+The pre-2.0 entry points live in [`legacy/`](legacy/) and run as modules, for example `uv run python -m legacy.generate`. They still work, but they are **deprecated and may be removed at any time**. Their flags differ from the `vera` CLI (for example `-c` means `--max-concurrent` in `legacy/generate.py`). Use `vera` for new work; if you still depend on a script, see [docs/legacy-scripts.md](docs/legacy-scripts.md) for its options and the `vera` equivalent.
 
 ## License
 
diff --git a/docs/architecture.md b/docs/architecture.md
index f97bb14e..79ed2969 100644
--- a/docs/architecture.md
+++ b/docs/architecture.md
@@ -123,14 +123,12 @@ namespaces, input config files, or target manifests. They define no CLI behavior
 defaults. Pooling likewise delegates to the function owned by the scoring domain
 rather than to a script entry point.
 
-Legacy root scripts may remain temporarily while their replacement feature is
-migrated. During that transition, `vera_cli` may import the reusable function
-from a root script, but never its argument parser. For generation, the temporary
-flow is `vera_cli.generate` → `generate.run_for_user_models` → `generate.main`.
-Removing `generate.py` and moving those functions plus the existing
-`generate_conversations/` code into the permanent `generate/` package is one
-atomic later change, so a root `generate.py` module and a top-level `generate/`
-package never coexist.
+Legacy scripts live in `legacy/` until they are removed, and nothing outside
+`legacy/` and its tests imports them. For generation, the flow is
+`vera_cli.generate` → `generate_conversations.run_for_user_models` →
+`generate_conversations.run_generation`. With `generate.py` out of the root,
+renaming `generate_conversations/` to the permanent `generate/` package no
+longer collides with a root `generate` module.
 
 `run_for_user_models` and its `_legacy_model_config` helper are explicit
 stopgaps. They put the expansion of a run's user models, and the flattening of
@@ -141,7 +139,7 @@ models identically.
 
 | Subcommand | Delegates to | Purpose |
 |------------|--------------|---------|
-| `vera generate` | generation application function (temporarily `generate.run_for_user_models`) | Simulate conversations → `<target>/c_<chatbot>/u_*/conversations/` |
+| `vera generate` | generation application function (temporarily `generate_conversations.run_for_user_models`) | Simulate conversations → `<target>/c_<chatbot>/u_*/conversations/` |
 | `vera judge` | `judge.runner` | Evaluate transcripts → `<target>/c_<chatbot>/u_*/evaluations/j_*` |
 | `vera score` | `score.score` | Aggregate `results.csv` → scores and visualizations |
 | `vera pool` | `score.pool` | Concatenate multiple evaluation folders into one pooled result → `<target>/c_<chatbot>/pooled/u_<a>+<b>_*` |
diff --git a/docs/evaluating.md b/docs/evaluating.md
index 4ae2346d..3efbe068 100644
--- a/docs/evaluating.md
+++ b/docs/evaluating.md
@@ -242,8 +242,8 @@ Add configuration to [config.py](../llm_clients/config.py) if your LLM requires
 ### 5. Use the new LLM in your simulations
 
 ```bash
-python3 generate.py -u your-model-name -p your-model-name -t 5 -r 1
-python3 judge.py -f conversations/{YOUR_FOLDER} -j your-model-name
+uv run python vera.py generate -c your-model-name -u your-model-name -t 5 --target SI
+uv run python vera.py judge --conversations output/{YOUR_RUN_FOLDER} -j your-model-name --target SI
 ```
 
 ## Prompt caching (by provider)
diff --git a/docs/legacy-scripts.md b/docs/legacy-scripts.md
index 63de80b7..74f6c668 100644
--- a/docs/legacy-scripts.md
+++ b/docs/legacy-scripts.md
@@ -1,23 +1,25 @@
 # Legacy scripts (deprecated)
 
-> **These scripts will be removed soon.** Use the [`vera` CLI](cli.md) for new work. This page exists so workflows that still depend on a script keep running while they migrate.
+> **These scripts may be removed at any time.** Use the [`vera` CLI](cli.md) for new work. This page exists so workflows that still depend on a script keep running while they migrate.
+
+The scripts live in [`legacy/`](../legacy/). Run the Python ones from the repository root as modules (`uv run python -m legacy.<script>`); running the file directly can't import the project's packages.
 
 | Legacy script | Replacement |
 |---------------|-------------|
-| `generate.py` | `vera generate` |
-| `judge.py` | `vera judge` |
-| `python -m judge.score` | `vera score` |
-| `run_pipeline.py` | `vera pipeline` |
-| `scripts/run_recommended_vera_pipeline.sh` | `vera pipeline --config` with [`configs/recommended-SI.json`](../configs/recommended-SI.json), then [pooling](scoring.md#pooling-several-evaluations) |
+| `python -m legacy.generate` | `vera generate` |
+| `python -m legacy.judge` | `vera judge` |
+| `python -m legacy.score` | `vera score` |
+| `python -m legacy.run_pipeline` | `vera pipeline` |
+| `legacy/run_recommended_vera_pipeline.sh` | `vera pipeline --config` with [`configs/recommended-SI.json`](../configs/recommended-SI.json), then [pooling](scoring.md#pooling-several-evaluations) |
 
 **The flags are not the same.** The `vera` CLI uses `-u`/`-c`/`-j` for user/chatbot/judge. The scripts use `-p` for the chatbot, and reuse `-c` and `-r` for unrelated things (`-c` is `--max-concurrent` in `generate.py` and `--conversation` in `judge.py`; `-r` is `--runs` in `generate.py` and `--rubrics` in `judge.py`). Don't copy flags between the two.
 
 The scripts also take rubric bundles only as full manifest paths (`data/SI/rubric_manifest.json`); there is no `--target SI` shorthand. Omitting the flag defaults to SI.
 
-## `generate.py`
+## `legacy/generate.py`
 
 ```bash
-uv run python generate.py -u gpt-4o -p gpt-4o -t 6 -r 1
+uv run python -m legacy.generate -u gpt-4o -p gpt-4o -t 6 -r 1
 ```
 
 | Flag | `vera generate` equivalent | Description |
@@ -43,10 +45,10 @@ uv run python generate.py -u gpt-4o -p gpt-4o -t 6 -r 1
 | `-usp`, `--user-start-prompt` | none | Prompt sent to the user model for its first turn |
 | `-d`, `--debug` | `-d` | Debug logging |
 
-## `judge.py`
+## `legacy/judge.py`
 
 ```bash
-uv run python judge.py -f output/<generation-run>/ -j gpt-5.4
+uv run python -m legacy.judge -f output/<generation-run>/ -j gpt-5.4
 ```
 
 | Flag | `vera judge` equivalent | Description |
@@ -63,20 +65,20 @@ uv run python judge.py -f output/<generation-run>/ -j gpt-5.4
 | `-pj`, `--per-judge` | `--per-judge` | Apply `--max-concurrent` per judge model |
 | `-vw`, `--verbose-workers` | none | Verbose worker logging |
 
-## `judge/score.py`
+## `legacy/score.py`
 
 ```bash
-uv run python -m judge.score -r output/<generation-run>/evaluations/<evaluation-run>/results.csv
+uv run python -m legacy.score -r output/<generation-run>/evaluations/<evaluation-run>/results.csv
 ```
 
 Same flags as `vera score` (`-r`, `-o`, `--skip-risk-analysis`), except `--personas-tsv` / `-p` defaults to `data/SI/personas.tsv`. `vera score --personas` has no default and skips the risk breakdown when omitted.
 
-## `run_pipeline.py`
+## `legacy/run_pipeline.py`
 
-Runs `generate.py`, `judge.py`, and `judge/score.py` in sequence:
+Runs the generation, judging, and scoring scripts in sequence:
 
 ```bash
-uv run python run_pipeline.py \
+uv run python -m legacy.run_pipeline \
   --user-agent claude-sonnet-4-5-20250929 \
   --provider-agent gpt-4o \
   --runs 2 \
@@ -93,14 +95,14 @@ To resume:
 - **`--resume-judge`**—set `-jo` to the existing `j_*` folder.
 - **Both**—set `-co` to the `p_*` folder; it must contain exactly one `j_*` under `evaluations/`.
 
-See `uv run python run_pipeline.py --help` for everything.
+See `uv run python -m legacy.run_pipeline --help` for everything.
 
-## `scripts/run_recommended_vera_pipeline.sh`
+## `legacy/run_recommended_vera_pipeline.sh`
 
-Runs the recommended profile through `run_pipeline.py`, once per user model, then pools the two evaluations:
+Runs the recommended profile through `legacy/run_pipeline.py`, once per user model, then pools the two evaluations:
 
 ```bash
-./scripts/run_recommended_vera_pipeline.sh <provider-agent-model> [extra run_pipeline.py args]
+./legacy/run_recommended_vera_pipeline.sh <provider-agent-model> [extra run_pipeline arguments]
 ```
 
 Its defaults, and the `VERA_*` environment variables that override them, are documented in the script. Their `vera pipeline` equivalents:
diff --git a/judge/score.py b/judge/score.py
index b6425e87..4d5f8a1e 100644
--- a/judge/score.py
+++ b/judge/score.py
@@ -1,8 +1,7 @@
-#!/usr/bin/env python3
 """
 Score evaluation results from judge/runner.py output.
 
-Run with `python -m judge.score -r evaluations/[eval_folder]/results.csv`
+Run with `vera score -r <results.csv>`; the legacy CLI is `legacy/score.py`.
 
 Reads results.csv, re-calculates the dataframe from the tsv files in the same
 folder if the results.csv is empty, calculates dimension-level and aggregate scores,
@@ -11,7 +10,6 @@
 - scores/scores_by_risk_visualization.png: Scores broken down by persona risk level
 """
 
-import argparse
 import json
 import traceback
 from pathlib import Path
@@ -490,7 +488,7 @@ def run_scoring(
     `judge.run.run_judging`: it receives final values, does the work, and
     returns. It does not parse arguments, apply defaults, or decide where
     anything lives -- every one of those is the caller's. That boundary is what
-    lets one function serve both `vera score` and the legacy `main` below.
+    lets one function serve both `vera score` and the legacy `legacy/score.py`.
 
     Visualization failures are warnings rather than errors: the scores are the
     result, and a chart that could not be drawn does not invalidate them.
@@ -568,62 +566,3 @@ def run_scoring(
         traceback.print_exc()
 
     return results, scores_dir
-
-
-def main():
-    """Legacy CLI entry point: parse arguments, then call `run_scoring`."""
-    parser = argparse.ArgumentParser(
-        description=(
-            "Score evaluation results from judge/runner.py output "
-            "and generate visualizations"
-        )
-    )
-
-    parser.add_argument(
-        "--results-csv",
-        "-r",
-        required=True,
-        help="Path to results.csv file from judge evaluation",
-    )
-    parser.add_argument(
-        "--output-json",
-        "-o",
-        default=None,
-        help="Path to save JSON output "
-        "(default: scores/scores.json next to results.csv)",
-    )
-    parser.add_argument(
-        "--personas-tsv",
-        "-p",
-        default="data/SI/personas.tsv",
-        help=(
-            "Path to personas.tsv file for risk-level analysis "
-            "(default: data/SI/personas.tsv)"
-        ),
-    )
-    parser.add_argument(
-        "--skip-risk-analysis",
-        action="store_true",
-        help="Skip risk-level analysis and visualization",
-    )
-
-    args = parser.parse_args()
-
-    try:
-        run_scoring(
-            results_csv=args.results_csv,
-            output_json=args.output_json,
-            personas_tsv=args.personas_tsv,
-            skip_risk_analysis=args.skip_risk_analysis,
-        )
-    except (FileNotFoundError, ValueError) as error:
-        print(f"Error: {error}")
-        return 1
-
-    return 0
-
-
-if __name__ == "__main__":
-    import sys
-
-    sys.exit(main())
diff --git a/legacy/README.md b/legacy/README.md
new file mode 100644
index 00000000..c55de1b4
--- /dev/null
+++ b/legacy/README.md
@@ -0,0 +1,5 @@
+# Legacy scripts
+
+These are the pre-2.0 entry points. They are deprecated and **may be removed at any time**. Use the [`vera` CLI](../docs/cli.md) instead; [docs/legacy-scripts.md](../docs/legacy-scripts.md) maps each script's flags to its `vera` equivalent.
+
+Run them from the repository root as modules, for example `uv run python -m legacy.generate`, `-m legacy.judge`, `-m legacy.score`, or `-m legacy.run_pipeline`.
diff --git a/legacy/__init__.py b/legacy/__init__.py
new file mode 100644
index 00000000..84a1be9f
--- /dev/null
+++ b/legacy/__init__.py
@@ -0,0 +1 @@
+"""Deprecated pre-2.0 entry points; see legacy/README.md."""
diff --git a/legacy/run_pipeline.py b/legacy/run_pipeline.py
index 251bf554..0cecb0ff 100644
--- a/legacy/run_pipeline.py
+++ b/legacy/run_pipeline.py
@@ -441,14 +441,16 @@ async def main():
 
     # Import generate and judge main functions
     # We import here to avoid circular dependencies and to allow --debug flag to be set
-    # Import judge.py main function
-    # (note: judge.py is a module file, judge/ is a package)
+    # Import legacy/judge.py main function by path
+    # (note: legacy/judge.py is a module file, judge/ is a package)
     import importlib.util
 
-    from generate import main as generate_main
-    from generate import resolve_persona_inputs
+    from legacy.generate import main as generate_main
+    from legacy.generate import resolve_persona_inputs
 
-    spec = importlib.util.spec_from_file_location("judge_script", "judge.py")
+    spec = importlib.util.spec_from_file_location(
+        "judge_script", Path(__file__).with_name("judge.py")
+    )
     judge_script = importlib.util.module_from_spec(spec)
     spec.loader.exec_module(judge_script)
     judge_main = judge_script.main
diff --git a/legacy/run_recommended_vera_pipeline.sh b/legacy/run_recommended_vera_pipeline.sh
index 0b8c7efd..7f09eb5b 100755
--- a/legacy/run_recommended_vera_pipeline.sh
+++ b/legacy/run_recommended_vera_pipeline.sh
@@ -24,10 +24,10 @@
 # the other, until this script is retired in favor of `vera pipeline`.
 #
 # Usage:
-#   ./scripts/run_recommended_vera_pipeline.sh <provider-agent-model> [extra run_pipeline.py args...]
+#   ./legacy/run_recommended_vera_pipeline.sh <provider-agent-model> [extra run_pipeline.py args...]
 #
 # Example:
-#   ./scripts/run_recommended_vera_pipeline.sh gpt-4o
+#   ./legacy/run_recommended_vera_pipeline.sh gpt-4o
 #
 # Optional environment (override defaults without editing this file):
 #   VERA_OUTPUT_PARENT     Where new p_* run folders go (default: output)
@@ -103,7 +103,7 @@ run_pipeline_capture_eval() {
   log="$(mktemp)"
   # tee duplicates stream to stderr so the user sees progress; stdout would
   # otherwise pollute the captured eval path from command substitution.
-  uv run python run_pipeline.py "$@" 2>&1 | tee "$log" >&2
+  uv run python -m legacy.run_pipeline "$@" 2>&1 | tee "$log" >&2
   local st="${PIPESTATUS[0]}"
   if [[ "$st" -ne 0 ]]; then
     rm -f "$log"
diff --git a/legacy/score.py b/legacy/score.py
new file mode 100644
index 00000000..adc486d2
--- /dev/null
+++ b/legacy/score.py
@@ -0,0 +1,68 @@
+"""Legacy CLI for scoring: ``uv run python -m legacy.score -r <results.csv>``.
+
+Deprecated; use ``vera score``. Its only differences from ``vera score`` are
+the flag names and the ``--personas-tsv`` default, documented in
+docs/legacy-scripts.md.
+"""
+
+import argparse
+import sys
+
+from judge.score import run_scoring
+
+
+def main():
+    """Legacy CLI entry point: parse arguments, then call `run_scoring`."""
+    parser = argparse.ArgumentParser(
+        description=(
+            "Score evaluation results from judge/runner.py output "
+            "and generate visualizations"
+        )
+    )
+
+    parser.add_argument(
+        "--results-csv",
+        "-r",
+        required=True,
+        help="Path to results.csv file from judge evaluation",
+    )
+    parser.add_argument(
+        "--output-json",
+        "-o",
+        default=None,
+        help="Path to save JSON output "
+        "(default: scores/scores.json next to results.csv)",
+    )
+    parser.add_argument(
+        "--personas-tsv",
+        "-p",
+        default="data/SI/personas.tsv",
+        help=(
+            "Path to personas.tsv file for risk-level analysis "
+            "(default: data/SI/personas.tsv)"
+        ),
+    )
+    parser.add_argument(
+        "--skip-risk-analysis",
+        action="store_true",
+        help="Skip risk-level analysis and visualization",
+    )
+
+    args = parser.parse_args()
+
+    try:
+        run_scoring(
+            results_csv=args.results_csv,
+            output_json=args.output_json,
+            personas_tsv=args.personas_tsv,
+            skip_risk_analysis=args.skip_risk_analysis,
+        )
+    except (FileNotFoundError, ValueError) as error:
+        print(f"Error: {error}")
+        return 1
+
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/llm_clients/config.py b/llm_clients/config.py
index 1368343b..21bb7aa4 100644
--- a/llm_clients/config.py
+++ b/llm_clients/config.py
@@ -99,7 +99,7 @@ def get_base_url(cls, provider: str) -> Optional[str]:
 
     # The recommended VERA-MH profile: the two user models and the judge behind
     # the published scores. configs/recommended-SI.json must match these (a unit
-    # test enforces it), and scripts/run_recommended_vera_pipeline.sh reads them
+    # test enforces it), and legacy/run_recommended_vera_pipeline.sh reads them
     # through get_recommended_models(). Change them only together with the
     # published methodology -- scores from other models are not comparable.
     RECOMMENDED_USER_MODELS = ("gpt-5.2", "claude-opus-4-5-20251101")
diff --git a/scripts/pool_vera_scores.py b/scripts/pool_vera_scores.py
index a8765c99..d0006110 100755
--- a/scripts/pool_vera_scores.py
+++ b/scripts/pool_vera_scores.py
@@ -7,7 +7,7 @@
 straight concatenation of every source's rows. Common uses include combining the
 two user-agent suites of the recommended profile (each judged with ``gpt-5.4``
 by default), or merging separate judge runs (e.g. GPT-4o and Sonnet) over the
-same conversations. ``scripts/run_recommended_vera_pipeline.sh`` calls this
+same conversations. ``legacy/run_recommended_vera_pipeline.sh`` calls this
 itself; ``vera pipeline`` does not — it prints the evaluation folders to pass
 here, because pooling is a separate command by design (``docs/pipeline.md``).
 
diff --git a/scripts/run_combinations.sh b/scripts/run_combinations.sh
index a66b204c..062285b1 100755
--- a/scripts/run_combinations.sh
+++ b/scripts/run_combinations.sh
@@ -26,7 +26,7 @@ failed=()
 for user in "${users[@]}"; do
     for provider in "${providers[@]}"; do
         cmd=(
-            uv run python generate.py
+            uv run python -m legacy.generate
             --user-agent "$user"
             --provider-agent "$provider"
             --turns "$TURNS"
diff --git a/scripts/score_publication_evaluations.sh b/scripts/score_publication_evaluations.sh
index 4cf6d614..8592329f 100755
--- a/scripts/score_publication_evaluations.sh
+++ b/scripts/score_publication_evaluations.sh
@@ -51,7 +51,7 @@ while IFS= read -r eval_folder; do
     echo "[$count/$total] 📈 Scoring: $eval_folder"
     
     # Run score.py
-    if python3 -m judge.score -r "$results_csv" 2>&1; then
+    if python3 -m legacy.score -r "$results_csv" 2>&1; then
         success=$((success + 1))
         echo "   ✅ Success"
     else
diff --git a/tests/integration/test_cli_parity.py b/tests/integration/test_cli_parity.py
index a847f412..b0c0dd9b 100644
--- a/tests/integration/test_cli_parity.py
+++ b/tests/integration/test_cli_parity.py
@@ -9,7 +9,8 @@
 `generate.py`/`judge.py`/`run_pipeline.py` and makes `vera.py` the only entry
 point. At that moment the comparison loses its second side, so delete this file
 rather than adapting it. See the transitional-boundary section of
-`docs/architecture.md` and the docstring on `generate.run_for_user_models`.
+`docs/architecture.md` and the docstring on
+`generate_conversations.run_for_user_models`.
 
 The seam is `generate_conversations.run_generation`: both CLIs converge on
 exactly one call to it, so stubbing it and diffing the recorded kwargs compares
@@ -101,7 +102,7 @@ def run_legacy_cli() -> None:
     sys.argv = list(LEGACY_ARGV)
     # Returns rather than exiting: the script only calls `sys.exit(1)` when every
     # conversation was skipped, and the stub reports no conversations at all.
-    runpy.run_path(str(REPO_ROOT / "generate.py"), run_name="__main__")
+    runpy.run_module("legacy.generate", run_name="__main__")
 
 
 def run_unified_cli() -> None:
diff --git a/tests/integration/test_judge_against_clinician_ratings.py b/tests/integration/test_judge_against_clinician_ratings.py
index 216d04d1..c996d03c 100644
--- a/tests/integration/test_judge_against_clinician_ratings.py
+++ b/tests/integration/test_judge_against_clinician_ratings.py
@@ -193,7 +193,9 @@ def test_judge_scores_match_expected_ratings(
         cmd = [
             "uv",
             "run",
-            "judge.py",
+            "python",
+            "-m",
+            "legacy.judge",
             "-f",
             str(conversations_dir),
             "-j",
diff --git a/tests/integration/test_pipeline.py b/tests/integration/test_pipeline.py
index a06bf5c4..986b405e 100644
--- a/tests/integration/test_pipeline.py
+++ b/tests/integration/test_pipeline.py
@@ -85,7 +85,7 @@ class TestPipelineArgumentParsing:
 
     def test_parse_arguments_required_only(self):
         """Test parsing with only required arguments."""
-        from run_pipeline import parse_arguments
+        from legacy.run_pipeline import parse_arguments
 
         test_args = [
             "--user-agent",
@@ -111,7 +111,7 @@ def test_parse_arguments_required_only(self):
 
     def test_parse_short_co_is_conversation_output(self):
         """-co is shorthand for --conversation-output."""
-        from run_pipeline import parse_arguments
+        from legacy.run_pipeline import parse_arguments
 
         test_args = [
             "--user-agent",
@@ -133,7 +133,7 @@ def test_parse_short_co_is_conversation_output(self):
 
     def test_parse_short_jo_is_judge_output(self):
         """-jo is shorthand for --judge-output."""
-        from run_pipeline import parse_arguments
+        from legacy.run_pipeline import parse_arguments
 
         test_args = [
             "--user-agent",
@@ -162,7 +162,7 @@ def test_parse_short_jo_is_judge_output(self):
     )
     def test_parse_arguments_with_judge_output(self, co_flag, jo_flag):
         """Parse co/jo; resolve_pipeline_resume_paths for a non-resume run."""
-        from run_pipeline import parse_arguments, resolve_pipeline_resume_paths
+        from legacy.run_pipeline import parse_arguments, resolve_pipeline_resume_paths
 
         test_args = [
             "--user-agent",
@@ -191,7 +191,7 @@ def test_parse_arguments_with_judge_output(self, co_flag, jo_flag):
 
     def test_parse_arguments_with_extra_params(self):
         """Test parsing with extra model parameters."""
-        from run_pipeline import parse_arguments
+        from legacy.run_pipeline import parse_arguments
 
         test_args = [
             "--user-agent",
@@ -224,7 +224,7 @@ def test_parse_arguments_with_extra_params(self):
 
     def test_parse_arguments_multiple_judge_models(self):
         """Test parsing with multiple judge models."""
-        from run_pipeline import parse_arguments
+        from legacy.run_pipeline import parse_arguments
 
         test_args = [
             "--user-agent",
@@ -247,7 +247,7 @@ def test_parse_arguments_multiple_judge_models(self):
 
     def test_parse_arguments_missing_required(self):
         """Test that missing required arguments raises error."""
-        from run_pipeline import parse_arguments
+        from legacy.run_pipeline import parse_arguments
 
         test_args = [
             "--user-agent",
@@ -261,7 +261,7 @@ def test_parse_arguments_missing_required(self):
 
     def test_parse_arguments_optional_flags(self):
         """Test parsing optional boolean flags."""
-        from run_pipeline import parse_arguments
+        from legacy.run_pipeline import parse_arguments
 
         test_args = [
             "--user-agent",
@@ -290,7 +290,7 @@ def test_parse_arguments_optional_flags(self):
 
     def test_parse_arguments_with_all_optional_arguments(self):
         """Test parsing with all optional arguments provided."""
-        from run_pipeline import parse_arguments
+        from legacy.run_pipeline import parse_arguments
 
         test_args = [
             "--user-agent",
@@ -556,7 +556,7 @@ def test_conversation_and_judge_output_attributes(self, pipeline_args):
 
     def test_parse_arguments_with_run_id(self):
         """Test parsing arguments with --run-id."""
-        from run_pipeline import parse_arguments
+        from legacy.run_pipeline import parse_arguments
 
         test_args = [
             "--user-agent",
@@ -580,7 +580,7 @@ def test_parse_arguments_with_run_id(self):
 
     def test_parse_arguments_with_rubrics(self):
         """Test parsing arguments with --rubrics."""
-        from run_pipeline import parse_arguments
+        from legacy.run_pipeline import parse_arguments
 
         test_args = [
             "--user-agent",
@@ -605,7 +605,7 @@ def test_parse_arguments_with_rubrics(self):
 
     def test_parse_arguments_defaults_for_new_args(self):
         """Test that new arguments have correct defaults."""
-        from run_pipeline import parse_arguments
+        from legacy.run_pipeline import parse_arguments
 
         test_args = [
             "--user-agent",
@@ -631,7 +631,7 @@ def test_parse_arguments_defaults_for_new_args(self):
 
     def test_short_flags_for_extra_params(self):
         """Test that short flags work for extra params arguments."""
-        from run_pipeline import parse_arguments
+        from legacy.run_pipeline import parse_arguments
 
         test_args = [
             "--user-agent",
@@ -664,7 +664,7 @@ def test_short_flags_for_extra_params(self):
 
     def test_short_flag_for_run_id(self):
         """Test that short flag -i works for run-id."""
-        from run_pipeline import parse_arguments
+        from legacy.run_pipeline import parse_arguments
 
         test_args = [
             "--user-agent",
@@ -720,14 +720,14 @@ async def test_step1_validation_folder_not_exists(
         import sys
         from unittest.mock import patch
 
-        from run_pipeline import main as pipeline_main
+        from legacy.run_pipeline import main as pipeline_main
 
         # Mock generate module's main to return a non-existent folder
         async def mock_generate(*args, **kwargs):
             return None, str(tmp_path / "nonexistent")
 
         # Patch generate.main at the source
-        with patch("generate.main", side_effect=mock_generate):
+        with patch("legacy.generate.main", side_effect=mock_generate):
             # Mock sys.exit to raise SystemExit instead of actually exiting
             with patch.object(sys, "exit", side_effect=SystemExit) as mock_exit:
                 # Mock importlib to avoid judge loading (not needed for step 1 test)
@@ -748,7 +748,7 @@ async def test_step1_validation_no_conversation_files(
         import sys
         from unittest.mock import patch
 
-        from run_pipeline import main as pipeline_main
+        from legacy.run_pipeline import main as pipeline_main
 
         gen_run = _make_gen_run_empty_conversations(tmp_path)
 
@@ -756,7 +756,7 @@ async def test_step1_validation_no_conversation_files(
         async def mock_generate(*args, **kwargs):
             return None, str(gen_run)
 
-        with patch("generate.main", side_effect=mock_generate):
+        with patch("legacy.generate.main", side_effect=mock_generate):
             # Mock sys.exit to raise SystemExit instead of actually exiting
             with patch.object(sys, "exit", side_effect=SystemExit) as mock_exit:
                 with patch("importlib.util.spec_from_file_location"):
@@ -774,7 +774,7 @@ async def test_step1_validation_only_log_files(self, tmp_path, valid_pipeline_ar
         import sys
         from unittest.mock import patch
 
-        from run_pipeline import main as pipeline_main
+        from legacy.run_pipeline import main as pipeline_main
 
         gen_run = _make_gen_run_with_txt(
             tmp_path,
@@ -788,7 +788,7 @@ async def test_step1_validation_only_log_files(self, tmp_path, valid_pipeline_ar
         async def mock_generate(*args, **kwargs):
             return None, str(gen_run)
 
-        with patch("generate.main", side_effect=mock_generate):
+        with patch("legacy.generate.main", side_effect=mock_generate):
             # Mock sys.exit to raise SystemExit instead of actually exiting
             with patch.object(sys, "exit", side_effect=SystemExit) as mock_exit:
                 with patch("importlib.util.spec_from_file_location"):
@@ -808,7 +808,7 @@ async def test_step2_validation_no_evaluation_folder(
         import sys
         from unittest.mock import MagicMock, patch
 
-        from run_pipeline import main as pipeline_main
+        from legacy.run_pipeline import main as pipeline_main
 
         gen_run = _make_gen_run_with_txt(
             tmp_path, {"conv1.txt": "User: Hi\nAssistant: Hello"}
@@ -827,7 +827,7 @@ async def mock_judge(args):
         mock_judge_module.main = mock_judge
 
         with (
-            patch("generate.main", side_effect=mock_generate),
+            patch("legacy.generate.main", side_effect=mock_generate),
             patch("importlib.util.module_from_spec", return_value=mock_judge_module),
             patch("importlib.util.spec_from_file_location"),
         ):
@@ -849,7 +849,7 @@ async def test_step2_validation_folder_not_exists(
         import sys
         from unittest.mock import MagicMock, patch
 
-        from run_pipeline import main as pipeline_main
+        from legacy.run_pipeline import main as pipeline_main
 
         gen_run = _make_gen_run_with_txt(
             tmp_path, {"conv1.txt": "User: Hi\nAssistant: Hello"}
@@ -868,7 +868,7 @@ async def mock_judge(args):
         mock_judge_module.main = mock_judge
 
         with (
-            patch("generate.main", side_effect=mock_generate),
+            patch("legacy.generate.main", side_effect=mock_generate),
             patch("importlib.util.module_from_spec", return_value=mock_judge_module),
             patch("importlib.util.spec_from_file_location"),
         ):
@@ -888,7 +888,7 @@ async def test_step2_validation_no_results_csv(self, tmp_path, valid_pipeline_ar
         import sys
         from unittest.mock import MagicMock, patch
 
-        from run_pipeline import main as pipeline_main
+        from legacy.run_pipeline import main as pipeline_main
 
         gen_run = _make_gen_run_with_txt(
             tmp_path, {"conv1.txt": "User: Hi\nAssistant: Hello"}
@@ -912,7 +912,7 @@ async def mock_judge(args):
         mock_judge_module.main = mock_judge
 
         with (
-            patch("generate.main", side_effect=mock_generate),
+            patch("legacy.generate.main", side_effect=mock_generate),
             patch("importlib.util.module_from_spec", return_value=mock_judge_module),
             patch("importlib.util.spec_from_file_location"),
         ):
@@ -934,7 +934,7 @@ async def test_step2_validation_empty_folder_error_message(
         import sys
         from unittest.mock import MagicMock, patch
 
-        from run_pipeline import main as pipeline_main
+        from legacy.run_pipeline import main as pipeline_main
 
         gen_run = _make_gen_run_with_txt(
             tmp_path, {"conv1.txt": "User: Hi\nAssistant: Hello"}
@@ -959,7 +959,7 @@ async def mock_judge(args):
         mock_judge_module.main = mock_judge
 
         with (
-            patch("generate.main", side_effect=mock_generate),
+            patch("legacy.generate.main", side_effect=mock_generate),
             patch("importlib.util.module_from_spec", return_value=mock_judge_module),
             patch("importlib.util.spec_from_file_location"),
         ):
@@ -988,7 +988,7 @@ async def test_results_csv_has_expected_columns(
         import csv
         from unittest.mock import MagicMock, patch
 
-        from run_pipeline import main as pipeline_main
+        from legacy.run_pipeline import main as pipeline_main
 
         gen_run = _make_gen_run_with_txt(
             tmp_path, {"conv1.txt": "User: Hi\nAssistant: Hello"}
@@ -1030,12 +1030,12 @@ def mock_score(*args, **kwargs):
         mock_judge_module.main = mock_judge
 
         with (
-            patch("generate.main", side_effect=mock_generate),
+            patch("legacy.generate.main", side_effect=mock_generate),
             patch("importlib.util.module_from_spec", return_value=mock_judge_module),
             patch("importlib.util.spec_from_file_location"),
-            patch("run_pipeline.score_results", new=mock_score),
-            patch("run_pipeline.print_scores"),
-            patch("run_pipeline.create_visualizations"),
+            patch("legacy.run_pipeline.score_results", new=mock_score),
+            patch("legacy.run_pipeline.print_scores"),
+            patch("legacy.run_pipeline.create_visualizations"),
         ):
             with patch("sys.argv", valid_pipeline_args + ["--skip-risk-analysis"]):
                 await pipeline_main()
@@ -1092,7 +1092,7 @@ async def test_results_csv_missing_required_columns(
         """Test pipeline handling when results.csv is missing required columns."""
         from unittest.mock import MagicMock, patch
 
-        from run_pipeline import main as pipeline_main
+        from legacy.run_pipeline import main as pipeline_main
 
         gen_run = _make_gen_run_with_txt(
             tmp_path, {"conv1.txt": "User: Hi\nAssistant: Hello"}
@@ -1120,12 +1120,12 @@ def mock_score(*args, **kwargs):
         mock_judge_module.main = mock_judge
 
         with (
-            patch("generate.main", side_effect=mock_generate),
+            patch("legacy.generate.main", side_effect=mock_generate),
             patch("importlib.util.module_from_spec", return_value=mock_judge_module),
             patch("importlib.util.spec_from_file_location"),
-            patch("run_pipeline.score_results", new=mock_score),
-            patch("run_pipeline.print_scores"),
-            patch("run_pipeline.create_visualizations"),
+            patch("legacy.run_pipeline.score_results", new=mock_score),
+            patch("legacy.run_pipeline.print_scores"),
+            patch("legacy.run_pipeline.create_visualizations"),
         ):
             with patch("sys.argv", valid_pipeline_args + ["--skip-risk-analysis"]):
                 # Pipeline should complete but the scoring function will handle
@@ -1139,7 +1139,7 @@ async def test_validation_success_messages(
         """Test that validation success messages are displayed."""
         from unittest.mock import MagicMock, patch
 
-        from run_pipeline import main as pipeline_main
+        from legacy.run_pipeline import main as pipeline_main
 
         gen_run = _make_gen_run_with_txt(
             tmp_path,
@@ -1171,12 +1171,12 @@ def mock_score(*args, **kwargs):
         mock_judge_module.main = mock_judge
 
         with (
-            patch("generate.main", side_effect=mock_generate),
+            patch("legacy.generate.main", side_effect=mock_generate),
             patch("importlib.util.module_from_spec", return_value=mock_judge_module),
             patch("importlib.util.spec_from_file_location"),
-            patch("run_pipeline.score_results", new=mock_score),
-            patch("run_pipeline.print_scores"),
-            patch("run_pipeline.create_visualizations"),
+            patch("legacy.run_pipeline.score_results", new=mock_score),
+            patch("legacy.run_pipeline.print_scores"),
+            patch("legacy.run_pipeline.create_visualizations"),
         ):
             with patch("sys.argv", valid_pipeline_args + ["--skip-risk-analysis"]):
                 await pipeline_main()
@@ -1197,7 +1197,7 @@ class TestPipelineResumeParsing:
     """CLI parsing for --resume-generate / --resume-judge."""
 
     def test_parse_resume_flags_default_false(self):
-        from run_pipeline import parse_arguments
+        from legacy.run_pipeline import parse_arguments
 
         test_args = [
             "--user-agent",
@@ -1217,7 +1217,7 @@ def test_parse_resume_flags_default_false(self):
             assert args.resume_judge is False
 
     def test_parse_resume_flags_can_be_true(self):
-        from run_pipeline import parse_arguments
+        from legacy.run_pipeline import parse_arguments
 
         test_args = [
             "--user-agent",
@@ -1244,7 +1244,7 @@ class TestPipelineResumeValidation:
     """resolve_pipeline_resume_paths() path checks."""
 
     def test_fresh_run_uses_conversation_output(self, tmp_path):
-        from run_pipeline import resolve_pipeline_resume_paths
+        from legacy.run_pipeline import resolve_pipeline_resume_paths
 
         gen_parent = tmp_path / "gen_parent"
         args = argparse.Namespace(
@@ -1256,7 +1256,7 @@ def test_fresh_run_uses_conversation_output(self, tmp_path):
         assert args._pipeline_gen_folder == str(gen_parent.resolve())
 
     def test_resume_generate_rejects_non_p_output(self, tmp_path):
-        from run_pipeline import resolve_pipeline_resume_paths
+        from legacy.run_pipeline import resolve_pipeline_resume_paths
 
         bad = tmp_path / "not_a_p_folder"
         bad.mkdir()
@@ -1270,7 +1270,7 @@ def test_resume_generate_rejects_non_p_output(self, tmp_path):
         assert exc_info.value.code == 2
 
     def test_resume_generate_requires_existing_directory(self, tmp_path):
-        from run_pipeline import resolve_pipeline_resume_paths
+        from legacy.run_pipeline import resolve_pipeline_resume_paths
 
         missing = tmp_path / "p_x__a_y__t1__r1__nope"
         args = argparse.Namespace(
@@ -1283,7 +1283,7 @@ def test_resume_generate_requires_existing_directory(self, tmp_path):
         assert exc_info.value.code == 2
 
     def test_resume_generate_accepts_valid_run_folder(self, tmp_path):
-        from run_pipeline import resolve_pipeline_resume_paths
+        from legacy.run_pipeline import resolve_pipeline_resume_paths
 
         run_dir = tmp_path / "p_persona__a_agent__t4__r1__20260101_120000"
         run_dir.mkdir()
@@ -1295,7 +1295,7 @@ def test_resume_generate_accepts_valid_run_folder(self, tmp_path):
         resolve_pipeline_resume_paths(args)
 
     def test_resume_judge_requires_existing_directory(self, tmp_path):
-        from run_pipeline import resolve_pipeline_resume_paths
+        from legacy.run_pipeline import resolve_pipeline_resume_paths
 
         p_run = tmp_path / "p_x__a_y__t1__r1__ts"
         p_run.mkdir()
@@ -1312,7 +1312,7 @@ def test_resume_judge_requires_existing_directory(self, tmp_path):
         assert exc_info.value.code == 2
 
     def test_resume_judge_rejects_parent_evaluations_directory(self, tmp_path):
-        from run_pipeline import resolve_pipeline_resume_paths
+        from legacy.run_pipeline import resolve_pipeline_resume_paths
 
         p_run = tmp_path / "p_x__a_y__t1__r1__ts"
         p_run.mkdir()
@@ -1328,7 +1328,7 @@ def test_resume_judge_rejects_parent_evaluations_directory(self, tmp_path):
         assert exc_info.value.code == 2
 
     def test_resume_judge_requires_j_evaluation_folder_basename(self, tmp_path):
-        from run_pipeline import resolve_pipeline_resume_paths
+        from legacy.run_pipeline import resolve_pipeline_resume_paths
 
         p_run = tmp_path / "p_x__a_y__t1__r1__ts"
         p_run.mkdir()
@@ -1346,7 +1346,7 @@ def test_resume_judge_requires_j_evaluation_folder_basename(self, tmp_path):
         assert exc_info.value.code == 2
 
     def test_resume_judge_accepts_valid_evaluation_folder(self, tmp_path):
-        from run_pipeline import resolve_pipeline_resume_paths
+        from legacy.run_pipeline import resolve_pipeline_resume_paths
 
         p_run = tmp_path / "p_gpt4o__a_x__t1__r1__20260101_120000"
         p_run.mkdir()
@@ -1362,7 +1362,7 @@ def test_resume_judge_accepts_valid_evaluation_folder(self, tmp_path):
         resolve_pipeline_resume_paths(args)
 
     def test_resume_judge_requires_judge_output(self):
-        from run_pipeline import resolve_pipeline_resume_paths
+        from legacy.run_pipeline import resolve_pipeline_resume_paths
 
         args = argparse.Namespace(
             resume_generate=False,
@@ -1374,7 +1374,7 @@ def test_resume_judge_requires_judge_output(self):
         assert exc_info.value.code == 2
 
     def test_both_resume_requires_single_j_under_evaluations(self, tmp_path):
-        from run_pipeline import resolve_pipeline_resume_paths
+        from legacy.run_pipeline import resolve_pipeline_resume_paths
 
         run_dir = tmp_path / "p_u__a_p__t1__r1__20260101_120000"
         run_dir.mkdir()
@@ -1421,7 +1421,7 @@ async def test_generate_and_judge_receive_resume_false_by_default(
     ):
         from unittest.mock import AsyncMock, MagicMock, patch
 
-        from run_pipeline import main as pipeline_main
+        from legacy.run_pipeline import main as pipeline_main
 
         gen_run = _make_gen_run_with_txt(
             tmp_path, {"c.txt": "User: hi\nAssistant: hey"}
@@ -1438,12 +1438,12 @@ async def test_generate_and_judge_receive_resume_false_by_default(
         mock_judge_module.main = judge_mock
 
         with (
-            patch("generate.main", side_effect=gen_mock),
+            patch("legacy.generate.main", side_effect=gen_mock),
             patch("importlib.util.module_from_spec", return_value=mock_judge_module),
             patch("importlib.util.spec_from_file_location"),
-            patch("run_pipeline.score_results", return_value={}),
-            patch("run_pipeline.print_scores"),
-            patch("run_pipeline.create_visualizations"),
+            patch("legacy.run_pipeline.score_results", return_value={}),
+            patch("legacy.run_pipeline.print_scores"),
+            patch("legacy.run_pipeline.create_visualizations"),
             patch("sys.argv", valid_pipeline_args + ["--skip-risk-analysis"]),
         ):
             await pipeline_main()
@@ -1459,7 +1459,7 @@ async def test_generate_receives_resume_true_with_resume_generate(
     ):
         from unittest.mock import AsyncMock, MagicMock, patch
 
-        from run_pipeline import main as pipeline_main
+        from legacy.run_pipeline import main as pipeline_main
 
         run_dir = _make_gen_run_with_txt(
             tmp_path,
@@ -1487,12 +1487,12 @@ async def test_generate_receives_resume_true_with_resume_generate(
         ]
 
         with (
-            patch("generate.main", side_effect=gen_mock),
+            patch("legacy.generate.main", side_effect=gen_mock),
             patch("importlib.util.module_from_spec", return_value=mock_judge_module),
             patch("importlib.util.spec_from_file_location"),
-            patch("run_pipeline.score_results", return_value={}),
-            patch("run_pipeline.print_scores"),
-            patch("run_pipeline.create_visualizations"),
+            patch("legacy.run_pipeline.score_results", return_value={}),
+            patch("legacy.run_pipeline.print_scores"),
+            patch("legacy.run_pipeline.create_visualizations"),
             patch("sys.argv", argv),
         ):
             await pipeline_main()
@@ -1506,7 +1506,7 @@ async def test_judge_receives_resume_true_with_resume_judge(
     ):
         from unittest.mock import AsyncMock, MagicMock, patch
 
-        from run_pipeline import main as pipeline_main
+        from legacy.run_pipeline import main as pipeline_main
 
         gen_run = _make_gen_run_with_txt(
             tmp_path,
@@ -1533,12 +1533,12 @@ async def test_judge_receives_resume_true_with_resume_judge(
         ]
 
         with (
-            patch("generate.main", side_effect=gen_mock),
+            patch("legacy.generate.main", side_effect=gen_mock),
             patch("importlib.util.module_from_spec", return_value=mock_judge_module),
             patch("importlib.util.spec_from_file_location"),
-            patch("run_pipeline.score_results", return_value={}),
-            patch("run_pipeline.print_scores"),
-            patch("run_pipeline.create_visualizations"),
+            patch("legacy.run_pipeline.score_results", return_value={}),
+            patch("legacy.run_pipeline.print_scores"),
+            patch("legacy.run_pipeline.create_visualizations"),
             patch("sys.argv", argv),
         ):
             await pipeline_main()
@@ -1552,7 +1552,7 @@ async def test_both_resume_flags_wired_independently(
     ):
         from unittest.mock import AsyncMock, MagicMock, patch
 
-        from run_pipeline import main as pipeline_main
+        from legacy.run_pipeline import main as pipeline_main
 
         run_dir = _make_gen_run_with_txt(
             tmp_path,
@@ -1580,12 +1580,12 @@ async def test_both_resume_flags_wired_independently(
         ]
 
         with (
-            patch("generate.main", side_effect=gen_mock),
+            patch("legacy.generate.main", side_effect=gen_mock),
             patch("importlib.util.module_from_spec", return_value=mock_judge_module),
             patch("importlib.util.spec_from_file_location"),
-            patch("run_pipeline.score_results", return_value={}),
-            patch("run_pipeline.print_scores"),
-            patch("run_pipeline.create_visualizations"),
+            patch("legacy.run_pipeline.score_results", return_value={}),
+            patch("legacy.run_pipeline.print_scores"),
+            patch("legacy.run_pipeline.create_visualizations"),
             patch("sys.argv", argv),
         ):
             await pipeline_main()
diff --git a/tests/integration/test_score_parity.py b/tests/integration/test_score_parity.py
index b9eb6fc5..b4aad37b 100644
--- a/tests/integration/test_score_parity.py
+++ b/tests/integration/test_score_parity.py
@@ -25,7 +25,7 @@
 import pytest
 
 import vera
-from judge import score as score_domain
+from legacy import score as legacy_score
 
 REPO_ROOT = Path(__file__).resolve().parents[2]
 EVAL_FIXTURE = REPO_ROOT / "tests/fixtures/eval_tsv_rebuild__20260415_140037"
@@ -73,7 +73,7 @@ def test_both_entry_points_produce_the_same_scores(
         *extra_legacy,
     ]
     with patch.object(sys, "argv", legacy_argv):
-        assert score_domain.main() == 0
+        assert legacy_score.main() == 0
 
     parser = vera.build_parser()
     args = parser.parse_args(
@@ -91,4 +91,4 @@ def test_legacy_main_still_reports_a_missing_results_csv(tmp_path: Path) -> None
     unchanged for callers that check the status rather than the exception.
     """
     with patch.object(sys, "argv", ["score.py", "-r", str(tmp_path / "absent.csv")]):
-        assert score_domain.main() == 1
+        assert legacy_score.main() == 1
diff --git a/tests/integration/test_scoring.py b/tests/integration/test_scoring.py
index 5235e842..1abbe3a8 100644
--- a/tests/integration/test_scoring.py
+++ b/tests/integration/test_scoring.py
@@ -69,7 +69,7 @@ def validate_test_environment():
         pytest.skip(f"Missing required environment variables: {missing}")
 
     # Validate repo structure
-    required_files = ["generate.py", "judge.py", "data/SI/personas.tsv"]
+    required_files = ["legacy/generate.py", "legacy/judge.py", "data/SI/personas.tsv"]
     repo_root = Path.cwd()
     missing_files = [f for f in required_files if not (repo_root / f).exists()]
     if missing_files:
@@ -117,7 +117,9 @@ async def run_generate_cli(
             "uv",
             "run",
             "python3",
-            "generate.py",
+            "python",
+            "-m",
+            "legacy.generate",
             "--user-agent",
             user_model,
             "--provider-agent",
@@ -215,7 +217,9 @@ async def run_judge_cli(
             "uv",
             "run",
             "python3",
-            "judge.py",
+            "python",
+            "-m",
+            "legacy.judge",
             "--folder",
             str(conversations_dir),
             "--judge-model",
@@ -326,7 +330,9 @@ def run_pipeline_cli(
             "uv",
             "run",
             "python3",
-            "run_pipeline.py",
+            "python",
+            "-m",
+            "legacy.run_pipeline",
             "--user-agent",
             config["USER_MODEL"],
             "--provider-agent",
@@ -751,7 +757,9 @@ async def test_run_pipeline_integration(self, test_workspace, repo_root):
         # Create test arguments for first persona (Omar) with minimal configuration
         timestamp = int(time.time())
         test_args = [
-            "run_pipeline.py",
+            "python",
+            "-m",
+            "legacy.run_pipeline",
             "--user-agent",
             TEST_CONFIG["USER_MODEL"],
             "--provider-agent",
@@ -934,7 +942,9 @@ async def test_run_pipeline_vs_individual_calls(self, test_workspace, repo_root)
         # Test 2: Run integrated pipeline
         timestamp = int(time.time())
         test_args = [
-            "run_pipeline.py",
+            "python",
+            "-m",
+            "legacy.run_pipeline",
             "--user-agent",
             TEST_CONFIG["USER_MODEL"],
             "--provider-agent",
diff --git a/tests/unit/generate_conversations/test_generate_cli.py b/tests/unit/generate_conversations/test_generate_cli.py
index 8c703f5a..64555072 100644
--- a/tests/unit/generate_conversations/test_generate_cli.py
+++ b/tests/unit/generate_conversations/test_generate_cli.py
@@ -7,7 +7,7 @@
 
 import pytest
 
-import generate
+from legacy import generate
 
 
 def _generation_kwargs(output_folder: str, **overrides: object) -> dict:
diff --git a/tests/unit/judge/test_judge_cli.py b/tests/unit/judge/test_judge_cli.py
index 73ab1dbb..691c835d 100644
--- a/tests/unit/judge/test_judge_cli.py
+++ b/tests/unit/judge/test_judge_cli.py
@@ -9,9 +9,9 @@
 import judge.run as _judge_run
 from utils.conversation_layout import resolve_conversation_input
 
-# Load judge.py script (project root) so we can test get_parser and main
+# Load legacy/judge.py by path so we can test get_parser and main
 _PROJECT_ROOT = Path(__file__).resolve().parents[3]
-_JUDGE_SCRIPT = _PROJECT_ROOT / "judge.py"
+_JUDGE_SCRIPT = _PROJECT_ROOT / "legacy" / "judge.py"
 _spec = importlib.util.spec_from_file_location("judge_script", _JUDGE_SCRIPT)
 assert _spec is not None and _spec.loader is not None
 _judge_script = importlib.util.module_from_spec(_spec)
diff --git a/tests/unit/test_vera_pipeline.py b/tests/unit/test_vera_pipeline.py
index 9a170f1c..acd4be67 100644
--- a/tests/unit/test_vera_pipeline.py
+++ b/tests/unit/test_vera_pipeline.py
@@ -258,7 +258,7 @@ def test_pipeline_declares_no_stage_specific_knobs() -> None:
 def test_shipped_recommended_config_resolves() -> None:
     """The checked-in profile is an artifact that would otherwise rot silently.
 
-    It publishes the same profile `scripts/run_recommended_vera_pipeline.sh`
+    It publishes the same profile `legacy/run_recommended_vera_pipeline.sh`
     runs, so a field going stale has to fail here rather than at the start of
     an expensive run.
     """
diff --git a/vera_cli/pipeline.py b/vera_cli/pipeline.py
index 57276347..1e3da562 100644
--- a/vera_cli/pipeline.py
+++ b/vera_cli/pipeline.py
@@ -20,7 +20,7 @@
 That chaining is the entire reason the command exists. Running the three
 commands by hand works, but the caller has to copy the generation run folder
 into the judge invocation and the `results.csv` path into the score
-invocation — which is what `scripts/run_recommended_vera_pipeline.sh` does by
+invocation — which is what `legacy/run_recommended_vera_pipeline.sh` does by
 scraping them out of a log.
 
 `docs/pipeline.md` covers what makes this command's input different from the

From ae8770399457cb1cbb959afc82b478657126c651 Mon Sep 17 00:00:00 2001
From: Luca Belli <129434630+sator-labs@users.noreply.github.com>
Date: Thu, 1 Oct 2026 16:27:23 -0700
Subject: [PATCH 4/5] docs: add the v2.0.1 changelog entry for the legacy/ move

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---
 CHANGELOG.md | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index 5aa3b28b..8d16a768 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,6 +4,12 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 
 ## [Unreleased]
 
+## [v2.0.1] \- Unreleased
+
+### Breaking / migration
+
+- **Legacy scripts moved to [`legacy/`](legacy/)**—Run them from the repository root as modules: `uv run python -m legacy.generate`, `-m legacy.judge`, `-m legacy.score` (was `python -m judge.score`), and `-m legacy.run_pipeline`. They may be removed at any time; [`docs/legacy-scripts.md`](docs/legacy-scripts.md) maps their flags to `vera`.
+
 ### Breaking / migration
 
 - **`vera score` skips the risk-level breakdown when no personas file is given** — Legacy [`judge/score.py`](judge/score.py) defaults `--personas-tsv` to `data/SI/personas.tsv`. `vera score`'s `--personas` has no default: omit it and the risk-level breakdown is skipped with a message, so `scores_by_risk.json` is not written at all. The default was dropped because `load_personas_risk_levels` ([`judge/score_utils.py`](judge/score_utils.py)) joins on a column literally named `"Short Current Suicide Risk Level"`; against any other target's personas file every row silently became `"Unknown"` and the resulting `scores_by_risk.json` was empty — output that looked like a result but was not one. Pass `--personas data/SI/personas.tsv` for exactly the legacy behavior. Legacy [`judge/score.py`](judge/score.py) keeps its default until it is removed.

From 19592235247fbf81935d4cfdec0fc77c9dd9b0bb Mon Sep 17 00:00:00 2001
From: Luca Belli <129434630+sator-labs@users.noreply.github.com>
Date: Thu, 1 Oct 2026 17:21:32 -0700
Subject: [PATCH 5/5] docs: note where run_for_user_models came from and why it
 moved

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---
 generate_conversations/main.py | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/generate_conversations/main.py b/generate_conversations/main.py
index 66c4fa4b..6150f34a 100644
--- a/generate_conversations/main.py
+++ b/generate_conversations/main.py
@@ -150,6 +150,10 @@ async def run_generation(
     return results, output_folder
 
 
+# Moved here from the root `generate.py` when that script went to `legacy/`
+# (v2.0.1): `vera generate` calls this, and `vera_cli` must not import from
+# `legacy/`. It used to call the legacy `main()`, a pass-through with the same
+# keyword arguments, so it now calls `run_generation` directly.
 async def run_for_user_models(
     generation: GenerationConfig,
     *,