Skip to content

Official Codebase of 3DCodeVerse

Python

docs: correction — the "branching coupling" gap does not exist

ziyao445122be8fdb
diff --git a/harness/bench/data/coupled/README.md b/harness/bench/data/coupled/README.md
index 7008a8a0..7607d6f1 100644
--- a/harness/bench/data/coupled/README.md
+++ b/harness/bench/data/coupled/README.md
@@ -29,7 +29,9 @@ have exactly the shape of noise. The battery was run for the mechanical readout
**The mechanical difference is structural**: asked for the coupling, the planner declares
it every time and eight of ten mechanisms come out with one degree of freedom; not asked,
-three do.
+three do. The two that keep more than one degree of freedom keep the right ones — the
+pram's four wheel spins beside its fold hinge, the pantograph mirror's tilt beside its
+extension — so `movable − mimic` counts DOF, not missed couplings.
**The three are the interesting part.** `nomimic`'s couplings are not the planner's —
they cannot be, the field is gone — they are the AGENT's, written by hand into
diff --git a/harness/bench/data/plan_stage/README.md b/harness/bench/data/plan_stage/README.md
index 5269733a..1e1870ec 100644
--- a/harness/bench/data/plan_stage/README.md
+++ b/harness/bench/data/plan_stage/README.md
@@ -97,12 +97,12 @@ Two readings:
median of 3 followers per 6 movable joints; here it is 10 of 10 and the strict one-input
mechanisms collapse to a single degree of freedom. The difference is in the prompt, not
the planner.
-* **Branching couplings are still under-declared.** Where one input drives one chain
- (umbrella, blind, scissor lift, ladder, garage door, treadle) the plan leaves exactly one
- input. Where it drives several chains at once — the pram's fold hinge moving the handle,
- the seat back and two leg pairs — it declares 3 of 8 and leaves 5 inputs standing. That
- is the case the sweep still over-drives, and the one a build-and-judge battery has to
- look at.
+* **`inputs left` is a DOF count, not a defect count.** Read joint by joint on the ten
+ recorded plans, eight leave exactly one free joint and it IS the input; the pram's five
+ are its fold hinge plus four wheel spins (the fold itself couples handle, seat back and
+ both leg pairs), and the pantograph mirror's second is the mirror tilt. The planner
+ declared the coupling correctly in all ten — an earlier reading of this table called the
+ pram "3 of 8 declared" and was wrong.
The plan stage is therefore NOT what a full run of this battery would be measuring: it
loses nothing here.
diff --git a/harness/docs/PAPER_WRITING.md b/harness/docs/PAPER_WRITING.md
index 822d328d..2686995e 100644
--- a/harness/docs/PAPER_WRITING.md
+++ b/harness/docs/PAPER_WRITING.md
@@ -294,9 +294,12 @@ median goes **6 → 3** driven joints, and 24 URDFs are fully one-input.
**A battery built for the question** (`coupled_v1`, ten one-input linkages, 2026-09-04)
separates what the planner does from what the agent does. Plan stage alone, 200 calls:
-196 valid plans, **zero** planning losses, **193 declare a coupling**; the strict
-one-input mechanisms collapse to a single degree of freedom and the branching ones (a
-pram's fold hinge driving handle, seat back and two leg pairs) declare 3 of 8. Then two
+196 valid plans, **zero** planning losses, **193 declare a coupling**; the ten
+mechanisms each come back with the coupling declared, and the free joints left over are
+the input itself (eight of ten) or a real second degree of freedom — the pram's four wheel
+spins, the pantograph mirror's tilt. (An earlier reading of the same table treated
+`movable − mimic` as undeclared couplings and reported a branching-coupling gap; joint by
+joint there is none.) Then two
build-and-judge arms, baseline round only, differing by two edits — the `mimic` field
typed to `None` so the planner cannot fill it, and the paragraph asking for it removed
from the plan prompt (`bench/data/coupled/`):