The open authoring profile: GPT-6's freedom on the infra's evidence, materials and gate

[ZheningHuang]

Why

We ran GPT-6 (Codex, gpt-6-astra) on Office_room two ways and rendered both from the same 59 capture cameras with the same Cycles setup:

• GPT-6 alone — the raw scan (frames, depth, point cloud, room.usdz) and a one-page brief, no LiteReality tooling. 37 min, 69 tool calls.
• GPT-6 on our infra — the current simulation authoring profile over the pipeline's assets.

Our side won clearly on materials (fetch_material's photo-matched PBR: carpet pile, wood grain, acoustic tile, the blind's fabric — GPT-6 alone had permission to fetch and still chose flat procedural colour), on real glazing, on reconstructed/articulated assets and on sim export. GPT-6 alone won on layout and fixtures: it measured desk and shelf heights from the depth maps, triangulated chair positions from silhouettes, built the heater with fins and dials, the AC louvres, sockets at the measured spacing, the coat on its hooks — and rebuilt until its own contact/intersection audit passed. Our scripted profile tells the model what to read, in what order, and what not to touch; none of that measuring was possible with what it was given.

This PR keeps what the infra does well and gives the model what the one-shot run had.

What

--author-profile open — a brief, not a script. It states what a finished room has and how it is judged; "how you work is yours", with two conditions (Room.py compiles after every save; save as you go). Furniture boxes may be corrected when a measurement says so; articulated and reconstructed assets must survive. Defaults to the Codex/gpt-6 harness and to budgets the brief needs (300 steps / 400 turns — the one-shot took 69 calls; our last pass was cut at 56).

Evidence kit + pack (evidence_kit/, agent/evidence_pack.py) — plain-python helpers the model imports in its own scripts, laid out beside the room as evidence/: contact sheets of every photograph, the stitches, the raw capture, and measure.probe_depth / triangulate / height_profile / pcd_slice, rectify.rectify_region, arkit_cameras. One README states the data formats and the Y-up→Z-up convention once. Verified on Office_room: the cloud's floor within 3 cm of RoomPlan's; desk and meeting table both read 0.75 m — the numbers GPT-6 measured by hand.

Validation gate (pipeline/room_qc/validate.py) — every rests_on claim measured against its support's top (+1 cm / −2 cm, over its footprint), floor-standing furniture against the floor, undeclared supports, the real-mesh contact report (python-fcl), and whether each articulated asset still reaches the build as a handle with its parts. Box overlaps are advisory (a chair tucked under a table overlaps its box legitimately). validation.json beside the layout; exit 0/2. The brief tells the model to run it; the author stage builds the room if the session didn't and runs it again, recording the verdict. Current Office_room build: PASS — 0 clashes, 131 declared supports all seated.

Trace fixes the same measurement exposed: the session record carries the model that actually ran (the Office_room gpt-6 session was filed under claude-opus-5, the role's name) and the harness; on Codex the raw rollout is copied beside the trace — the only place that says which images the model was actually shown (28, against 84 filename mentions the trace had counted).

Ported from zhening/gpt-6 (never landed on main): the simulation profile with rests_on/attached_to and the --author-profile/--author-turns flags, the Codex whole-output fix, and the chair-QC gate in reconstruct.

Verification

• 631 tests pass; the two failures are the pre-existing environment ones (test_stdio_bridge_matches_the_registry — the venv lacks the installed package; test_reconstruct_resolves_python_from_canonical_repo_root — a real scans_uploaded/Office_room shadows the test's path).
• New tests: the kit on a synthetic capture with known geometry (probe lands on the desk top, triangulation round-trips to 1e-3 m, height profile finds desk + floor, slice draws only the desk); the gate on floating / sunk / undeclared / off-support / tucked-under cases and its exit codes; profile registration, budgets, and that the scripted profiles still format.
• Smoke-tested end to end on Office_room with a 6-call Claude session: pack built, README read first, trace fields present, build + gate PASS.
• A full open-profile run on Office_room with gpt-6 is queued for when the Codex quota window reopens; the three-way render comparison (GPT-6 alone / current profile / open) will be posted here.

[ZheningHuang]

Result: the open profile on Office_room (gpt-6 via Codex)

Ran after merge on a copy of the scene (run-open/). 25.5 min, 65 tool calls, finished on its own (not cut by budget or quota). Gate from the harness: PASS — 128 objects, 115 declared supports, 0 support failures, 0 mesh clashes, articulation intact.

What it did, in order: reviewed all 59 photographs from the contact sheets → built and rendered the baseline → fetched real PBR sets for carpet, timber, ceiling and fabrics (fetch_material) → measured with the kit (desk 0.75 m, shelves 1.42 / 1.81–1.84 m) and found the scan's meeting-table box oversized, refitting it to 1.04 m Ø × 0.75 m and the desk depth to 0.87 m → authored shelves, boards, trunking, sockets, heater, AC, blinds, the coat (following the door), shelf contents and desk clutter with explicit supports → ran the gate (PASS on the first build) → kept fixing against its own photo comparisons (ceiling, one ceiling-grid/light-frame contact) → rebuilt, gate PASS again.

Four-way comparison from the same 59 capture cameras, identical Cycles setup: photo | GPT-6 alone | previous infra (simulation) | GPT-6 + new infra (open). The open column has the previous profile's materials and glazing and the one-shot run's measured layout and fixtures — the right-sized meeting table with the chairs where the photos put them, the heater with fins under the desk, the coat on its hooks, the AC louvres, the real carpet.

Renders and sheets: run/_gpt6_oneshot/compare/{compare4_mandatory.jpg, compare4_extra.jpg, compare4_all59.mp4}; the authored room: run-open/Office_room/realism_authoring/.