Skip to content

Official Codebase of 3DCodeVerse

Python

scene: the one-world switch stays OFF — its in-loop gain did not survive the fixed judge (D70)

gaoypengfd71defdca89
diff --git a/harness/codeverse/config.py b/harness/codeverse/config.py
index 01afbfa0..61a13570 100644
--- a/harness/codeverse/config.py
+++ b/harness/codeverse/config.py
@@ -179,8 +179,8 @@ SEED_RECIPES_ENV = "CV3D_SEED_RECIPES"
#: run, and whether it earns that is a measurement nobody has made yet.
SCENE_TEXTURES_ENV = "CV3D_SCENE_TEXTURES"
#: ONE session authors every zone of a scene (the whole world, one author) instead of the
-#: two-small-zones-per-session fan-out. ON by default since the loop-17 measurement (D70);
-#: `off` is the fan-out control.
+#: two-small-zones-per-session fan-out. OFF: the loop-17 gain did not survive the fixed-judge
+#: confirmation (D70); an A/B knob.
ONE_WORLD_SESSION_ENV = "CV3D_ONE_WORLD_SESSION"
_TRUE_WORDS = frozenset({"1", "on", "true", "yes", "y"})
_FALSE_WORDS = frozenset({"0", "off", "false", "no", "n"})
@@ -233,7 +233,7 @@ def scene_textures_enabled() -> bool:
def one_world_session_enabled() -> bool:
- """Do all of a scene's zones go to ONE session? ``$CV3D_ONE_WORLD_SESSION``, else ON.
+ """Do all of a scene's zones go to ONE session? ``$CV3D_ONE_WORLD_SESSION``, else off.
Measured 2026-09-08 under one fixed judge (gemini-3.1-pro-preview, n=2, the brief's
must_have list), both sides ``codex:gpt-6-astra@low``: a bare one-file scene — one author,
@@ -247,10 +247,12 @@ def one_world_session_enabled() -> bool:
goes to one session that owns every zone file (the env, the assets and the heroes keep
their own stages), with a session window scaled to the zone count. Measured on the
same three briefs (loop 17, in-loop judge): round 0 0.17 → 0.69, 0.51 → 0.58, 0.44 →
- 0.43; best round 0.75 → 0.82 PASS, 0.73 → 0.60, 0.60 → 0.77 PASS — so ON by default,
- and ``off`` is the fan-out control while the fixed-judge confirmation runs.
+ 0.43; best round 0.75 → 0.82 PASS, 0.73 → 0.60, 0.60 → 0.77 PASS. The fixed-judge
+ confirmation (cmp_scene4, both arms astra-low) did NOT reproduce it: harness 0.60 / 0.60 /
+ 0.54 vs the bare one-shot 0.81 / 0.83 / 0.60, the single session finishing four zones in
+ five minutes as a block-out. Off by default; a knob for the next measurement.
"""
- return env_flag(ONE_WORLD_SESSION_ENV, True)
+ return env_flag(ONE_WORLD_SESSION_ENV, False)
def seed_recipes_enabled() -> bool:
diff --git a/harness/docs/DECISIONS.md b/harness/docs/DECISIONS.md
index 89c17a45..afcd7f84 100644
--- a/harness/docs/DECISIONS.md
+++ b/harness/docs/DECISIONS.md
@@ -1113,8 +1113,8 @@ written) that were accepted because the code works that way and the tests pin it
grounds anything (L7): harness-owned starter geometry, like `worldShell` (D51), that the
agent edits.
-* **D70 One author for the whole world — `CV3D_ONE_WORLD_SESSION`, ON by default since its
- measurement (2026-09-08).** Under one fixed judge (gemini-3.1-pro-preview, n=2, the brief's
+* **D70 One author for the whole world — `CV3D_ONE_WORLD_SESSION`, an A/B knob, OFF: its
+ in-loop gain did not survive the fixed judge (2026-09-08).** Under one fixed judge (gemini-3.1-pro-preview, n=2, the brief's
must_have list) and one model on both sides (`codex:gpt-6-astra@low`), a bare one-file scene
— one author, no plan, no lib, no assets, no tools — scored 0.894 on the clockmaker's
workshop and 0.82 on the boat workshop; the harness's round 0 of the same briefs, its zones
@@ -1132,9 +1132,17 @@ written) that were accepted because the code works that way and the tests pin it
`CV3D_SCENE_TEXTURES` does). **Measured (loop 17, the same three briefs, in-loop judge):**
round 0 clockmaker 0.17 → 0.69, boat 0.51 → 0.58, NYC 0.44 → 0.43; best round 0.75 → 0.82
PASS, 0.73 → 0.60, 0.60 → 0.77 PASS — two passes where the fan-out had none, and the
- interior's round 0 four times higher. The default is ON; `off` keeps the fan-out as the
- control until the fixed-judge confirmation (harness-with-switch vs the bare model on the
- same briefs) is in, after which the two batching constants go.
+ interior's round 0 four times higher. **The fixed-judge confirmation did not reproduce
+ it** (`cmp_scene4`, both arms astra-low, gemini-3.1-pro-preview n=2): harness with the
+ switch 0.60 / 0.60 / 0.54 against the bare one-shot 0.81 / 0.83 / 0.60; that draw's single
+ session finished four zones in five minutes as a block-out ("missing stove", "shelves
+ missing", "primitive tools", 0.32 at round 0) where loop 17's draw had reached 0.69 — one
+ author is coherent but thin, the fan-out dense but incoherent, and n=1 per arm per brief is
+ noise-sized. The knob stays OFF; the batch header now states the session's window and asks
+ for each zone's full density (measured next as loop 18). What the whole evening says
+ instead: the bare one-file scene wins round 0 on interiors every time (0.81–0.89), so the
+ harness's next experiment is to CONTAIN it — a whole-scene single-shot draft as round 0,
+ then its gates, heroes and refine rounds on top (the bench's `oneshot+refine` shape).
## Rejected / deferred
diff --git a/harness/tests/orchestrator_tracks/test_one_world_session.py b/harness/tests/orchestrator_tracks/test_one_world_session.py
index 6269a8f3..3286410e 100644
--- a/harness/tests/orchestrator_tracks/test_one_world_session.py
+++ b/harness/tests/orchestrator_tracks/test_one_world_session.py
@@ -19,11 +19,11 @@ from tests.orchestrator_tracks.conftest import make_spec
from tests.orchestrator_tracks.fakes import FakeRuntime, FakeServices
-def test_the_switch_is_read_at_call_time_and_on_by_default(monkeypatch):
+def test_the_switch_is_read_at_call_time_and_off_by_default(monkeypatch):
monkeypatch.delenv("CV3D_ONE_WORLD_SESSION", raising=False)
- assert one_world_session_enabled() is True # since loop 17 (D70): the measured default
- monkeypatch.setenv("CV3D_ONE_WORLD_SESSION", "off")
- assert one_world_session_enabled() is False # the fan-out control
+ assert one_world_session_enabled() is False # D70: the loop-17 gain did not survive the fixed judge
+ monkeypatch.setenv("CV3D_ONE_WORLD_SESSION", "on")
+ assert one_world_session_enabled() is True
monkeypatch.setenv("CV3D_ONE_WORLD_SESSION", "garbage")
assert one_world_session_enabled() is False # a typo in a bench command is the control, never a crash
diff --git a/harness/tests/skills/test_delivery_reaches_the_session.py b/harness/tests/skills/test_delivery_reaches_the_session.py
index fb3c09e4..a85fc708 100644
--- a/harness/tests/skills/test_delivery_reaches_the_session.py
+++ b/harness/tests/skills/test_delivery_reaches_the_session.py
@@ -38,8 +38,8 @@ def scene_run(tmp_path, monkeypatch):
from tests.orchestrator_tracks.test_tracks import _planner, _scene_writer
monkeypatch.setenv("CV3D_SKILLS", "1")
- # this run PROVES the fan-out is skill-safe (one attach serving two zone sessions), so it
- # runs the fan-out — the default since D70 (2026-09-08) is one session for every zone
+ # this run PROVES the fan-out is skill-safe (one attach serving two zone sessions): pin the
+ # fan-out whatever D70's knob defaults to
monkeypatch.setenv("CV3D_ONE_WORLD_SESSION", "off")
plan = ScenePlan.model_validate(plan_example(Track.SCENE))
plan.assets = [a for a in plan.assets if a.kind == "threejs"]