#!/usr/bin/env python3 """Report saved V4 plane-sweep DEV evidence without opening fresh data or models.""" # ruff: noqa: E501 -- Chinese scientific report prose import argparse import json import math from pathlib import Path LABELS = { "C0": "DENSE_SURFACE", "C1": "SWEEP_SURFACE", "C2": "SWEEP_NO_SURFACE", } def read(root, name, default=None): path = root / (name + ".json") if not path.exists() and default is not None: return default return json.loads(path.read_text()) def number(value): return "UNDEFINED" if value is None else f"{value:.6f}" def stat(value): if value is None or value.get("mean") is None: return "UNDEFINED（空区域未填零）" ci = value["ci95"] return f"{value['mean']:.6f}，95% CI [{ci[0]:.6f}, {ci[1]:.6f}]" def negated(value): """Exact sign flip of a paired scene contrast: mean, CI and counts.""" if value is None: return None lo, hi = value["ci95"] return { **value, "mean": -value["mean"], "ci95": [-hi, -lo], "positive": value["negative"], "negative": value["positive"], } def train_fit(curves, steps): rows = curves.get("training", []) fit = {} for variant in sorted({r["variant"] for r in rows}): per_seed = [] for seed in sorted({r["seed"] for r in rows if r["variant"] == variant}): tail = [ (r["terms"]["A"]["loss_depth"] + r["terms"]["B"]["loss_depth"]) / 2 for r in rows if r["variant"] == variant and r["seed"] == seed and r["step"] > steps - 100 ] if tail: per_seed.append(sum(tail) / len(tail)) fit[variant] = sum(per_seed) / len(per_seed) if per_seed else None return fit def learned_tau(root): taus = {} for variant, name in (("C1", "sweep_training"), ("C2", "sweep_no_surface_training")): rows = read(root, name, []) values = [math.exp(r["sweep_log_tau"]) for r in rows if r.get("sweep_log_tau") is not None] taus[variant] = [round(v, 6) for v in values] return taus def run(root, docs=None): root, docs = Path(root), Path(docs) if docs else Path(root) results = read(root, "static_results") qualification = read(root, "qualification_results") state = read(root, "state_use_results") controls = read(root, "wrong_scene_results")["contrasts"] obs = read(root, "observability_analysis") split = read(root, "scene_split") contract = read(root, "training_contract", {"steps": 2000}) curves = read(root, "training_curves", {"training": [], "dev": []}) selected = read(root, "selected_checkpoints", {"status": "NOT_RECORDED"}) roster = read(root, "evaluated_variants", {"secondary_status": {}}) correlation = read(root, "state_correlation", {"per_variant_mean_over_seeds": {}}) incidents = read(root, "audit/incidents", []) tests, integrity = (read(root, n, {"status": "PENDING"}) for n in ("tests", "integrity")) if results["primary"] != "C1_SURFACE16": raise PermissionError("Frozen V2 estimator primary label expected (C1 = V4 primary)") if ( qualification["FRESH_QUALIFICATION_OPENED"] or qualification["DYNAMIC_TTT_NEXT_STAGE_ALLOWED"] ): raise PermissionError("V4 is a DEV mechanism round; fresh and Dynamic TTT stay closed") groups = results["groups"] sweep = results["surface_gain"] secondary = "C2|direct" in groups surface_with_sweep = negated(results.get("free_space_extra_gain")) c2_status = roster["secondary_status"].get("C2", "NOT_RUN_SECONDARY_OPTIONAL") wrong, shuffle = controls["C1|wrong_scene"], controls["C1|spatial_shuffle"] specificity = ( "SUPPORTED" if wrong["ci95"][0] > 0 and shuffle["ci95"][0] > 0 else "NOT_ESTABLISHED" ) sweep_status = qualification["SURFACE_TRAINING_STATUS"] if sweep_status == "SUPPORTED" and specificity == "SUPPORTED": branch = "A" elif sweep["ci95"][0] > 0 or specificity == "SUPPORTED": branch = "B" else: branch = "C" next_step = { "A": "显式几何推理能让learned carrier形成场景专属几何。下一轮建立经过独立性审计的fresh队列做资格验证，并检验训练规模；仍不开Dynamic TTT。", "B": "几何线索有部分作用。下一轮检验更多上下文视角或训练规模；仍不开Dynamic TTT。", "C": "在3个上下文视角、32×40分辨率下，最小plane-sweep不足以让carrier泛化。下一轮另写协议，考虑预训练几何先验或更多视角；仍不开Dynamic TTT。", }[branch] fit = train_fit(curves, contract["steps"]) taus = learned_tau(root) corr = correlation["per_variant_mean_over_seeds"] variants = ("C0", "C1", "C2") if secondary else ("C0", "C1") fields = { "N_TRAIN_SCENES": sum(r.get("split") == "TRAIN" for r in split["scenes"]), "N_DEV_SCENES": len(results["scene_ids"]), "DEV_REUSED_FROM_V2_V3": True, "DENSE_SURFACE_CHECKPOINT": selected.get("C0", "NOT_RECORDED"), "SWEEP_SURFACE_CHECKPOINT": selected.get("C1", "NOT_RECORDED"), "SWEEP_NO_SURFACE_CHECKPOINT": selected.get("C2", c2_status), "SECONDARY_C2_STATUS": c2_status, "DENSE_QUERY_ABSREL": number(groups["C0|direct"]["query_metrics"]["depth_absrel"]["mean"]), "SWEEP_QUERY_ABSREL": number(groups["C1|direct"]["query_metrics"]["depth_absrel"]["mean"]), "SWEEP_NO_SURFACE_QUERY_ABSREL": number( groups["C2|direct"]["query_metrics"]["depth_absrel"]["mean"] ) if secondary else "NOT_RUN_SECONDARY_OPTIONAL", "SWEEP_GAIN": number(sweep["mean"]), "SWEEP_GAIN_CI": sweep["ci95"], "SURFACE_WITH_SWEEP_GAIN": number(surface_with_sweep["mean"]) if surface_with_sweep else "NOT_RUN_SECONDARY_OPTIONAL", "SURFACE_WITH_SWEEP_GAIN_CI": surface_with_sweep["ci95"] if surface_with_sweep else "NOT_RUN_SECONDARY_OPTIONAL", "DENSE_FULL_CONTEXT_GAIN": number(results["full_context_gain"]["C0"]["mean"]), "SWEEP_FULL_CONTEXT_GAIN": number(results["full_context_gain"]["C1"]["mean"]), "SWEEP_FULL_CONTEXT_GAIN_CI": results["full_context_gain"]["C1"]["ci95"], "DENSE_WRONG_SCENE_DAMAGE": number(controls["C0|wrong_scene"]["mean"]), "SWEEP_WRONG_SCENE_DAMAGE": number(wrong["mean"]), "SWEEP_WRONG_SCENE_DAMAGE_CI": wrong["ci95"], "DENSE_SHUFFLE_DAMAGE": number(controls["C0|spatial_shuffle"]["mean"]), "SWEEP_SHUFFLE_DAMAGE": number(shuffle["mean"]), "SWEEP_SHUFFLE_DAMAGE_CI": shuffle["ci95"], "TRAIN_DEPTH_ABSREL_LAST100": {v: number(fit.get(v)) for v in variants}, "LEARNED_TAU_BY_SEED": taus, "WITHIN_VS_ACROSS_SCENE_STATE_CORR": { v: [ number(corr.get(v, {}).get("within_scene_A_B")), number(corr.get(v, {}).get("across_scene_A")), ] for v in variants }, "OBS2PLUS_DENSE": number(obs["groups"]["C0"]["OBS2PLUS"]["conditional_absrel"]["mean"]), "OBS2PLUS_SWEEP": number(obs["groups"]["C1"]["OBS2PLUS"]["conditional_absrel"]["mean"]), "OBS2PLUS_SWEEP_GAIN": number(obs["gains"]["OBS2PLUS"]["mean"]), "SWEEP_STATUS": sweep_status, "SCENE_SPECIFICITY_STATUS": specificity, "STATIC_DEV_STATUS": qualification["STATIC_DEV_STATUS"], "SEED_ROBUSTNESS": qualification["SEED_ROBUSTNESS"], "INTERPRETATION_BRANCH": branch, "FRESH_QUALIFICATION_INCLUDED": False, "FINAL_STATIC_STATUS": "NOT_ESTABLISHED", "DYNAMIC_TTT_NEXT_STAGE_ALLOWED": False, "TEST_TIME_INPUT": "RGB+CAMERA", "TEST_TIME_DEPTH_USED": False, "FINAL_HOLDOUT_TOUCHED": False, "DYNAMIC_TTT_RUN": False, "INFRASTRUCTURE_RETRIES": len(incidents), "TESTS": tests["status"], "FINAL_INTEGRITY": integrity["status"], "REPORT_DIR": str(docs), } lines = [ "# EXP-3D-PLANE-SWEEP-CARRIER-V4", "", "单因素机制诊断：V3的稠密16³ learned carrier，C1只多一条零初始化的plane-sweep旁路（不需要学习匹配的光度代价，沿参考射线做深度softmax，得到每个体素的表面与自由空间似然）。" + "SWEEP_GAIN=AbsRel(C0 DENSE_SURFACE)−AbsRel(C1 SWEEP_SURFACE)=" + stat(sweep) + f"；SWEEP_STATUS={sweep_status}；SCENE_SPECIFICITY_STATUS={specificity}；STATIC_DEV_STATUS={qualification['STATIC_DEV_STATUS']}；解读分支={branch}。", "", "Test-time input = RGB + calibrated cameras。训练是RGB+D utility track；depth只用于TRAIN loss和seal后的DEV评价。plane-sweep只读上下文RGB、上下文相机与训练集先验。所有场景都是历史曝光场景，8个DEV场景已在V2、V3中被评价：本轮是DEV机制诊断，不是资格轮，没有fresh队列。", "", "共享参数（5125个）、同seed初始值、4096个稠密证据候选、lifting、融合、refinement、heads、renderer、bounds、数据、预算与checkpoint规则全部相同；旁路共49个参数，零初始化保证训练开始时C1与C0完全相同。C2只去掉表面监督，是次要变体，不进入任何gate。统计复用V2冻结的估计器，JSON中的SURFACE_*字段在本轮表示主对比C0−C1（plane-sweep）。", "", "| Variant | Query AbsRel (95% CI) | Context AbsRel | Query−context gap | 训练末100步深度AbsRel | Opacity |", "|---|---|---:|---:|---:|---:|", ] for v in variants: g = groups[v + "|direct"] lines.append( f"| {v} {LABELS[v]} | {stat(g['query_metrics']['depth_absrel'])} | {number(g['context_metrics']['depth_absrel']['mean'])} | {number(g['generalization_gap']['mean'])} | {number(fit.get(v))} | {number(g['query_metrics']['opacity']['mean'])} |" ) answers = [ ( "plane-sweep几何线索是否改善DEV query几何？", f"SWEEP_GAIN={stat(sweep)}；SWEEP_STATUS={sweep_status}。gate沿用V2主对比：均值、CI、75%场景不差、所有LOSO>0、正向top1≤0.5、C1 wrong-scene与shuffle损伤CI>0、opacity与状态审计。", ), ( "是否跨多个场景和seed？", f"improved/tied/worse={sweep['positive']}/{sweep['tied']}/{sweep['negative']}；LOSO范围={sweep['loso_range']}；positive top1/top3={sweep['positive_top1_contribution']}/{sweep['positive_top3_contribution']}；seeds={results['seeds']}；SEED_ROBUSTNESS={qualification['SEED_ROBUSTNESS']}。每个seed的结果见static_results.json。", ), ( "状态是否变得场景专属？", f"C1 wrong-scene损伤={stat(wrong)}；C1 spatial shuffle损伤={stat(shuffle)}；SCENE_SPECIFICITY_STATUS={specificity}。对照C0：wrong-scene={stat(controls['C0|wrong_scene'])}，shuffle={stat(controls['C0|spatial_shuffle'])}。wrong-scene同时替换donor bounds，shuffle保留recipient bounds与相机。", ), ( "full context是否稳定优于anchor？", "C0=" + stat(results["full_context_gain"]["C0"]) + "；C1=" + stat(results["full_context_gain"]["C1"]) + "。anchor只有一个视角，既达不到≥2视角支持，也没有plane-sweep源视角，因此anchor状态只含学到的先验。", ), ( "训练集拟合与泛化差距如何？", "最后100步训练深度AbsRel（A/B平均、seed平均）=" + json.dumps({v: number(fit.get(v)) for v in variants}, ensure_ascii=False) + "；query−context gap：" + "；".join( f"{v}={stat(groups[v + '|direct']['generalization_gap'])}" for v in variants ) + "。只是描述，不是gate。", ), ( "模型学到了怎样的plane-sweep温度？", "exp(log_tau)（按seed）=" + json.dumps(taus, ensure_ascii=False) + "；初值0.01。温度越小表示越依赖尖锐的深度分布。", ), ( "有了几何线索后，表面监督是否还有作用？", ( "SURFACE_WITH_SWEEP_GAIN=AbsRel(C2)−AbsRel(C1)=" + stat(surface_with_sweep) + f"（正值表示表面监督有益；improved/tied/worse={surface_with_sweep['positive']}/{surface_with_sweep['tied']}/{surface_with_sweep['negative']}）。仅描述，不参与gate，不改变PRIMARY_METHOD=C1。" ) if surface_with_sweep else f"C2未纳入评价（SECONDARY_C2_STATUS={c2_status}）。", ), ( "改善是否出现在多视角可观测区域？", "C0−C1条件AbsRel：" + "；".join(f"{r}={stat(obs['gains'][r])}" for r in ("OBS0", "OBS1", "OBS2PLUS")) + "。几何推理应当主要改善OBS1/OBS2PLUS；这是seal后的区域诊断，不替代全图主指标。", ), ( "状态相关性诊断说明了什么？", "同场景A/B与跨场景A状态的密度相关（seed平均）=" + json.dumps( { v: { "within": number(corr.get(v, {}).get("within_scene_A_B")), "across": number(corr.get(v, {}).get("across_scene_A")), "direct_vs_anchor": number(corr.get(v, {}).get("direct_A_vs_anchor_A")), } for v in variants }, ensure_ascii=False, ) + "。仅描述，不参与gate。", ), ( "DEV是否达到static qualification？", qualification["STATIC_DEV_STATUS"] + "；检查=" + json.dumps(qualification["static_checks"], ensure_ascii=False) + "；主对比检查=" + json.dumps(qualification["surface_checks"], ensure_ascii=False) + "。即使通过，本轮DEV已曝光，也不构成最终资格。", ), ( "结论属于哪个预注册解读分支，下一步是什么？现在能否重开Dynamic TTT？", f"分支{branch}：{next_step} DYNAMIC_TTT_NEXT_STAGE_ALLOWED=false；DYNAMIC_TTT_RUN=false；FINAL_STATIC_STATUS=NOT_ESTABLISHED。", ), ] for i, (question, answer) in enumerate(answers, 1): lines.extend(["", f"**{i}. {question}**", "", answer]) lines.extend( [ "", "**opacity与状态审计**", "", "C0/C1 opacity checks=" + json.dumps(state["opacity_checks"], ensure_ascii=False) + "；C1 static opacity checks=" + json.dumps(state["static_opacity_checks"], ensure_ascii=False) + "；NO_HIT scenes=" + str(state["no_hit_scenes"]) + "；shared-state/seal审计=" + json.dumps(state["input_audit"], ensure_ascii=False) + "。", "", "**选择、统计与成本**", "", "每个variant×seed用同一DEV规则独立选择checkpoint（checkpoint在前期加密）。selected=" + json.dumps(selected, ensure_ascii=False) + "。先对每个query/role等权平均所有seed，再平均query/role得到scene值；10,000次paired scene bootstrap，seed 20260928，95% percentile CI。成本见cost_analysis.json；plane-sweep只增加构建状态时的计算，renderer与推理参数（除旁路49个）相同。", "", f"基础设施故障重跑次数={len(incidents)}（失败尝试归档在audit/failed_attempts且从不使用，见audit/incidents.json）；SECONDARY_C2_STATUS={c2_status}；CPU亲和性固定为8–23号核。协议见PROTOCOL.md。", "", "TEST_TIME_INPUT=RGB+CAMERA；TEST_TIME_DEPTH_USED=false；FINAL_HOLDOUT_TOUCHED=false；DYNAMIC_TTT_RUN=false。", ] ) summary = ( "PLANE SWEEP CARRIER V4 FINAL\n" + "\n".join( f"{k}={str(v).lower() if isinstance(v, bool) else v}" for k, v in fields.items() ) + "\n" ) (root / "README.md").write_text("\n".join(lines) + "\n") (root / "STATUS.md").write_text("```text\n" + summary + "```\n") (root / "terminal_summary.txt").write_text(summary) print(summary) return fields if __name__ == "__main__": p = argparse.ArgumentParser(description=__doc__) p.add_argument("--root", type=Path, required=True) p.add_argument("--docs", type=Path) a = p.parse_args() run(a.root, a.docs)