Official Codebase of 3DCodeVerse
Python
docs: delete the paragraphs whose subjects were deleted
gaoypeng66ddc4d06c51
diff --git a/harness/docs/ARCHITECTURE.md b/harness/docs/ARCHITECTURE.mdindex fea688e9..59a0e583 100644--- a/harness/docs/ARCHITECTURE.md+++ b/harness/docs/ARCHITECTURE.md@@ -219,16 +219,15 @@ codeverse/bench/ run_bench.py, report.py (renders through codeverse/gallery), compare_backends.py(preflights every model it needs; --wait-for-provider / --no-preflight),_infra.py (outage vs model failure: infra_failed / budget_exhausted, docs/EVAL.md §7),- ab_plan.py (the paired control/variant A/B rig, --aa calibration mode),- ab_gate_rates.py (the same run's deterministic readouts, paired per prompt),pin_plan.py (seed one plan into both arms so the paired delta stops carryingthe planner's spread — permitted only by plan_features.pin_plan_blockers),_compare_report.py (arm table incl. the `dropped` / `over budget` loss columns;dedups the append-only rows per (prompt, arm) so every reader agrees),_jsonl.py (the ONE tolerant reader/append-sealer for the resumable*.jsonl journals — a truncated last line never costs the paid rows),- ab_plan.py (paired control/variant A/B for plan + brief switches; pins- both children to the cap the §23 admission check reserved), _ab_report.py,+ ab_plan.py (paired control/variant A/B for plan + brief switches, --aa+ calibration mode; pins both children to the cap the §23 admission check+ reserved), _ab_report.py,ab_gate_rates.py (the same run's DETERMINISTIC readouts, paired perprompt: penetrating pairs, worst depth, floating parts, contract findings),_oneshot.py, _fixed_eval.py, cost_report.py,diff --git a/harness/docs/COST.md b/harness/docs/COST.mdindex 57d9681c..83c299d6 100644--- a/harness/docs/COST.md+++ b/harness/docs/COST.md@@ -398,9 +398,6 @@ Two helpers exist for the callers:`CostEstimate` before the call; `CostGuard(budget_usd).check(est)` decides.An unpriceable model is **allowed but flagged** — refusing to run because wecannot price something would be worse than running it.-* `Block(...)` + `order_blocks(...)` + `prefix_signature(...)` for cache-friendly- prompt assembly, and `prefix_report(prompts)` to measure whether a family of- prompts really shares a prefix (§4).## 11. Caveatsdiff --git a/harness/docs/DECISIONS.md b/harness/docs/DECISIONS.mdindex f4ef5eda..0d2ed206 100644--- a/harness/docs/DECISIONS.md+++ b/harness/docs/DECISIONS.md@@ -40,8 +40,7 @@ written) that were accepted because the code works that way and the tests pin it* **D4 Parallel refine only when it is safe.** Fan out only when ≥ 3 file-disjointgroups exist and every task maps to a file (threejs parts, scene zones/assets/env);single-file languages get one whole-object task. Consequence: `Workspace` itself- holds a per-root lock + index.lock retry (workspace.py); the older- `generation.workspace_lock` layer is redundant and queued for removal.+ holds a per-root lock + index.lock retry (workspace.py).* **D5 A refine round that changes no file is a plateau, not a crash.** Emits`round.no_change` and stops with the best round so far.* **D6 Finalise restores the best commit and rebuilds.** Artifacts always match the@@ -81,7 +80,6 @@ written) that were accepted because the code works that way and the tests pin it`round`, `kind`, `files_hint`, `language`, `track`, `read_only`, `mcp_command` … aretyped `AgentJob` fields; the `extra` dict and its `_lift_legacy_extra` shim (and theunused `model` field) were deleted 2026-08-29 — nothing in the tree built one.-* **D12 api-agent exposes `build` under its real name; `run_build` is an alias (Δ).*** **D13 Node ESM resolution via an import hook (Δ).** `NODE_PATH` cannot resolve bare ESMspecifiers; `run_node(three_hook=True)` adds `--import runtime_js/lib/resolve_three.mjs`;browser pages get an import map from `serve.cjs`. `ThreeJsRuntime` writesdiff --git a/harness/docs/EVAL.md b/harness/docs/EVAL.mdindex dca516ca..459a4f10 100644--- a/harness/docs/EVAL.md+++ b/harness/docs/EVAL.md@@ -179,8 +179,7 @@ Notes: the earlier criteria-first schema compressed flash to 0.6–0.7 (std 0.01"images beat gate text for visible facts" in the prompt fixed pro hallucinating"nothing moves" from contract-gate text.-* Variance on your own runs: `3dcv judge <slug> --n k` and read `score_std` /- `judges.metrics.judge_agreement`.+* Variance on your own runs: `3dcv judge <slug> --n k` and read `score_std`.* Sanity anchors: a skeleton placeholder should score ≈ 0.3–0.5; a deliberately wrongobject should trip the `intent_fidelity` floor / `wrong_object` defect; a floatingpart should cap via `connectivity` findings with `data["kind"]="floating"`.diff --git a/harness/docs/INTERFACES.md b/harness/docs/INTERFACES.mdindex 5af9e371..241aa856 100644--- a/harness/docs/INTERFACES.md+++ b/harness/docs/INTERFACES.md@@ -355,7 +355,6 @@ PairwiseJudge.compare_many(spec, candidates: list[RenderSet], *, rubric) -> Rankfrom codeverse.judges.calibration import calibrate, CalibrationTable # (run_dirs, *, model_id, n_samples=3, out_dir, geometry_mode,# rounds, …) -> rows + pearson/spearman(errors vs score), mean_std, cost; CLI: python -m codeverse.judges.calibration RUN… --n 3from codeverse.judges.vlm_judge import ReferenceJudge # image-conditioned specs-from codeverse.judges.metrics import judge_agreement, plateau, best_index```Rubrics: `static_object_v1` (0.72), `articulated_v1` (requires pose views via`missing_pose_sheet` cap), `scene_v1` (frame-gate caps dark/blown/flat/content_small),@@ -372,7 +371,7 @@ rec = get_track(spec.track, **options).run(spec, ws, resume=False) -> RunRecord# (CLI --candidates > spec.options.candidates > settings.default_candidates); StaticObject | Articulated | Scene | GraphicsTrackPipeline.run(spec, ws, *, resume=False, force=False) -> RunRecordfrom codeverse.orchestrator import RoundPolicy, StopPolicy, StopDecision, BestSelector, judge_sigma, \- best_score, last_gain, REWRITE_KIND, build_refine_instructions, compact_instructions+ best_score, last_gain, best_index, REWRITE_KIND, build_refine_instructions, compact_instructionsRoundPolicy(max_rounds=4, plateau_window=2, min_delta=0.02, target=0.8, judge_on_gate_errors=True, max_refine_tasks=6,max_instructions_per_task=6, parallel_min_tasks=2, n_candidates=1, pairwise_margin=0.03,pairwise_min_confidence=0.6, judge_samples=1,diff --git a/harness/docs/SKILLS.md b/harness/docs/SKILLS.mdindex bdf3a917..13a69c2c 100644--- a/harness/docs/SKILLS.md+++ b/harness/docs/SKILLS.md@@ -441,24 +441,16 @@ So at n = 8 with a free plan, neither readout can separate this switch from noth### What to try next, in order-0. **Give `api-agent` a first-class affordance.** This is now the top item and it is not a- measurement problem. Either a `read_skill(name)` tool beside `read_cookbook` — which the- agent *does* call — or, for the routed set, inline the one highest-priority body the way- the single-shot arm already does. A pointer in message 0 competes with everything else in- message 0; a tool in the tool list does not. Re-measure the read rate before anything- else: with 0 %, no A/B of this switch on `api-agent` can measure the library at all.-1. **Pin the plan** (`docs/EVAL.md` §8.1). The skills switch is generation-side — the+0. **Pin the plan** (`docs/EVAL.md` §8.1). The skills switch is generation-side — theplanner runs in the `plan` stage before anything is attached — so plan-once-write-both isvalid here, and it removes the dominant variance term from *both* readouts instead ofaveraging it down. It also costs one planner call *less* per pair. This is the singlehighest-value change and nothing else is worth running before it.-2. **Reconcile the prompt corpus on weld overlap** (§9). `contract.md` is fixed, but+1. **Reconcile the prompt corpus on weld overlap** (§9). `contract.md` is fixed, but`tracks/generate_static.j2` still says "overlap neighbours by ≥ 0.002 m (push a leg2-5 mm into the seat)". While that stands, `cv3d-part-contact` is arguing with theprompt inside the same session, and the pair-count readout is measuring the argument.-3. **Get one api-agent battery with the switch on**, for the exact-read calibration (§4).- Everything about the read metric on the CLI backends is an upper bound until then.-4. Only then re-run this A/B, with the gate counts as the primary readout.+2. Only then re-run this A/B, with the gate counts as the primary readout.---@@ -579,9 +571,6 @@ so; the rest are live.runs only lint and connectivity — never contract, joint_sweep, scene_frames or gl_frames.The *judged* column of any non-blender A/B is meaningless. The primary readout is safe:`bench/skill_targets.py` reads each arm's own harness record.-* **`telemetry.record_exact_read` is wired to `read_file` only** (`agents/api_agent.py`), so- `read_skill` never reaches `skill_reads.jsonl`. One line in `SkillTools.read_skill` makes- `3dcv skills report` truthful for api-agent without depending on atime at all.* **The atime probe is blind in any git workspace, and its control proves it.** The controlbundle came back "opened" in 27 of 33 sessions; the only 6 sessions where it stayed cleanwere the 6 scene sessions, where nothing was delivered. So the shipped report is honestdiff --git a/harness/docs/SKILLS_LEDGER.md b/harness/docs/SKILLS_LEDGER.mdindex ef05ea5d..1043f541 100644--- a/harness/docs/SKILLS_LEDGER.md+++ b/harness/docs/SKILLS_LEDGER.md@@ -256,8 +256,7 @@ it lands these four rows stay `mixed`, not `measured`.each arm's own harness record. Know that the *judged* column of any non-blender A/B ismeaningless.* **`telemetry.record_exact_read` is wired to `read_file` only** (`agents/api_agent.py`), so- `read_skill` — api-agent's actual read channel — never reaches `skill_reads.jsonl`. One- line in `SkillTools.read_skill` makes `3dcv skills report` truthful without atime.+ `read_skill` — api-agent's actual read channel — never reaches `skill_reads.jsonl`.*(Moot since 2026-08-28: the api-agent and its `read_skill` channel were deleted with it;vendor-CLI reads go through the filesystem and leave no exact-read channel to wire.)** **`scene_frames/content_*` has no classified kind.** Frame coverage is half of what