状态:PRE_OUTPUT_LOCKED / OUTPUT_BLIND_SOURCE_SCREENING_FROZEN / NATIVE_MATERIALIZATION_IN_PROGRESS
EVAL_VALIDITY_R0 是无训练的评价有效性审计:在严格更丰富的 oracle 输入下,当前
scene-fact → event-fact → feedback 评价链是否保持单调,因而可以公平比较未来模型?允许的
claim 仅为 Development evaluator-construct evidence;不产生模型优劣、独立泛化、产品、
Android、默认 App 或安全结论。完整合同是
EVAL_VALIDITY_R0_CONTRACT_2026-08-02.md。
先建立不能重叠的 source-session 注册表:
E:\codex-tools\bin\blindassist-python.cmd -m scripts.research.eval_validity_r0.build_exclusion_registry `
--riskseg-ledger docs/research/dual-loop/RISKSEG_R0_DATA_ROLE_LEDGER_2026-08-01.json `
--consumed-truth-ledger artifacts.local/evidence/riskseg-r0/event-eval/frozen-cohort-v1/truth_ledger.jsonl `
--output artifacts.local/evidence/eval-validity-r0/exclusion-registry-v1.json将任何 source-mask-only SANPO discovery 与该注册表相交,得到可继续作 RGB 盲审的候选(仍不是 event truth):
E:\codex-tools\bin\blindassist-python.cmd -m scripts.research.eval_validity_r0.select_unseen_discovery `
--discovery artifacts.local/evidence/eval-validity-r0/sanpo-train-sparse-discovery-v2.json `
--discovery artifacts.local/evidence/eval-validity-r0/sanpo-train-sparse-discovery-v3.json `
--exclusion-registry artifacts.local/evidence/eval-validity-r0/exclusion-registry-v1.json `
--output artifacts.local/evidence/eval-validity-r0/unseen-discovery-selection-v1.json严格 normal-walkable 的 source-mask shortlist 需要独立扫描;它同样不产生负例真值:
E:\codex-tools\bin\blindassist-python.cmd -m scripts.research.eval_validity_r0.discover_normal_candidates `
--exclusion-registry artifacts.local/evidence/eval-validity-r0/exclusion-registry-v1.json `
--max-sessions 96 `
--output artifacts.local/evidence/eval-validity-r0/sanpo-train-normal-discovery-v1.json冻结 48 个 source-session 的 output-blind screening universe 后,先以 GCS 元数据确认每个 窗口有 60 个连续的 native RGB/mask frame。这个步骤不能将 source-mask profile 写成 event bucket;它只修正缺帧/越界窗口,并保留原 source-mask reference frame。
E:\codex-tools\bin\blindassist-python.cmd -m scripts.research.eval_validity_r0.freeze_screening_cohort `
--hazard-selection artifacts.local/evidence/eval-validity-r0/unseen-discovery-selection-v3.json `
--normal-discovery artifacts.local/evidence/eval-validity-r0/sanpo-train-normal-discovery-v1.json `
--exclusion-registry artifacts.local/evidence/eval-validity-r0/exclusion-registry-v1.json `
--output artifacts.local/evidence/eval-validity-r0/screening-cohort-v2.json
E:\codex-tools\bin\blindassist-python.cmd -m scripts.research.eval_validity_r0.reconcile_screening_windows `
--screening-cohort artifacts.local/evidence/eval-validity-r0/screening-cohort-v2.json `
--output artifacts.local/evidence/eval-validity-r0/screening-cohort-v4.json
E:\codex-tools\bin\blindassist-python.cmd -m scripts.research.eval_validity_r0.materialize_screening_inputs `
--mode plan `
--screening-cohort artifacts.local/evidence/eval-validity-r0/screening-cohort-v4.json `
--output artifacts.local/evidence/eval-validity-r0/continuous-native-input-plan-v2.json
E:\codex-tools\bin\blindassist-python.cmd -m scripts.research.eval_validity_r0.materialize_screening_inputs `
--mode fetch `
--screening-cohort artifacts.local/evidence/eval-validity-r0/screening-cohort-v4.json `
--plan artifacts.local/evidence/eval-validity-r0/continuous-native-input-plan-v2.json `
--output artifacts.local/evidence/eval-validity-r0/continuous-native-inputs-v1materialization 完成后必须先做全量准入审计。它核验原生 payload hash、旧 truth ledger、
41 个排除 session、decoded RGB hash 和 pHash/crop/mirror 候选;清空候选时全量检查,发现
候选后仅保留至多 200 条证据并报告计数下界,仍立即输出 HOLD_EVAL_VALIDITY_DATA。
E:\codex-tools\bin\blindassist-python.cmd -m scripts.research.eval_validity_r0.audit_data_admission `
--screening-cohort artifacts.local/evidence/eval-validity-r0/screening-cohort-v4.json `
--materialized-root artifacts.local/evidence/eval-validity-r0/continuous-native-inputs-v1 `
--exclusion-registry artifacts.local/evidence/eval-validity-r0/exclusion-registry-v1.json `
--old-truth-ledger artifacts.local/evidence/riskseg-r0/event-eval/frozen-cohort-v1/truth_ledger.jsonl `
--prior-image-feature-cache artifacts.local/evidence/data-contamination-audit-r0/image_features_cache.jsonl `
--output artifacts.local/evidence/eval-validity-r0/data-admission-v1.json若 HOLD 的唯一原因是完整枚举的 pHash 候选(其余 exact/session/decoded/mask 检查均为
PASS),先用下面的 RGB-only 双人独立包判定每个固定候选是同一自然采集、不同自然采集或未知。
packet 不显示 session/event 身份、pHash 阈值、mask、action fact 或任何模型/oracle 输出;private
map 不得交给 reviewer。两位 reviewer 必须对全部 opaque pair 独立提交,且每一个 pair 都必须
exact agreement 为 DISTINCT_CAPTURE。SAME_CAPTURE、UNKNOWN、不一致、缺项或候选枚举不完整
都维持 HOLD_EVAL_VALIDITY_DATA;绝不修改候选、样本、阈值或旧 30-event 队列。
E:\codex-tools\bin\blindassist-python.cmd -m scripts.research.eval_validity_r0.prepare_phash_manual_review `
--admission-receipt artifacts.local/evidence/eval-validity-r0/data-admission-v1.json `
--materialized-root artifacts.local/evidence/eval-validity-r0/continuous-native-inputs-v1 `
--workspace-root E:\linnan\linnan `
--prior-image-feature-cache artifacts.local/evidence/data-contamination-audit-r0/image_features_cache.jsonl `
--reviewer-a-root artifacts.local/evidence/eval-validity-r0/phash-reviewer-a-v2 `
--reviewer-b-root artifacts.local/evidence/eval-validity-r0/phash-reviewer-b-v2 `
--private-root artifacts.local/evidence/eval-validity-r0/phash-private-map-v2
E:\codex-tools\bin\blindassist-python.cmd -m scripts.research.eval_validity_r0.finalize_phash_manual_review `
--admission-receipt artifacts.local/evidence/eval-validity-r0/data-admission-v1.json `
--private-map artifacts.local/evidence/eval-validity-r0/phash-private-map-v2/private-review-map.json `
--packet-a artifacts.local/evidence/eval-validity-r0/phash-reviewer-a-v2/packet.json `
--packet-b artifacts.local/evidence/eval-validity-r0/phash-reviewer-b-v2/packet.json `
--review-a artifacts.local/evidence/eval-validity-r0/phash-review-a-v2.json `
--review-b artifacts.local/evidence/eval-validity-r0/phash-review-b-v2.json `
--output artifacts.local/evidence/eval-validity-r0/phash-resolution-v2.json
E:\codex-tools\bin\blindassist-python.cmd -m scripts.research.eval_validity_r0.reconcile_phash_data_admission `
--held-admission-receipt artifacts.local/evidence/eval-validity-r0/data-admission-v1.json `
--phash-resolution artifacts.local/evidence/eval-validity-r0/phash-resolution-v2.json `
--output artifacts.local/evidence/eval-validity-r0/data-admission-after-phash-v2.json只有直接 PASS,或上述固定双人 pHash 审阅生成
EVAL_VALIDITY_DATA_ADMISSION_PASSED_AFTER_PHASH_MANUAL_REVIEW,才可产生两份物理分离的 P0 packet。每个 packet
有 192 个随机排序的 opaque items(48 event × 4 anchor),每项仅含结束于该 anchor 的 RGB;
private map 绝不能交给任何 reviewer。
E:\codex-tools\bin\blindassist-python.cmd -m scripts.research.eval_validity_r0.prepare_p0_review_packets `
--screening-cohort artifacts.local/evidence/eval-validity-r0/screening-cohort-v4.json `
--admission-receipt artifacts.local/evidence/eval-validity-r0/data-admission-v1.json `
--materialized-root artifacts.local/evidence/eval-validity-r0/continuous-native-inputs-v1 `
--reviewer-a-root artifacts.local/evidence/eval-validity-r0/p0-reviewer-a-v1 `
--reviewer-b-root artifacts.local/evidence/eval-validity-r0/p0-reviewer-b-v1 `
--private-root artifacts.local/evidence/eval-validity-r0/p0-private-map-v1两位 reviewer 独立提交后,custodian 只能使用 private map 与原 packet 做下面的 JSON-only
归档;它不读取 RGB/mask,更不读取任何模型、truth 或 oracle trace。任一 UNKNOWN 或不一致会
生成 STOP_EVENT_FACT_CONSISTENCY_NOT_ESTABLISHED;漏项、packet 替换或隔离声明失效会被拒绝归档。
两种情形都不允许第三人仲裁或抽换 event:
E:\codex-tools\bin\blindassist-python.cmd -m scripts.research.eval_validity_r0.finalize_p0_anchor_agreement `
--screening-cohort artifacts.local/evidence/eval-validity-r0/screening-cohort-v4.json `
--admission-receipt artifacts.local/evidence/eval-validity-r0/data-admission-v1.json `
--private-map artifacts.local/evidence/eval-validity-r0/p0-private-map-v1/private-review-map.json `
--packet-a artifacts.local/evidence/eval-validity-r0/p0-reviewer-a-v1/packet.json `
--packet-b artifacts.local/evidence/eval-validity-r0/p0-reviewer-b-v1/packet.json `
--review-a artifacts.local/evidence/eval-validity-r0/p0-action-review-a-v1.json `
--review-b artifacts.local/evidence/eval-validity-r0/p0-action-review-b-v1.json `
--output artifacts.local/evidence/eval-validity-r0/p0-anchor-agreement-v1.json只有 P0_ANCHOR_CONSISTENCY_PASSED 才能生成两份新的、与 P0 reviewer 不同的全事件 RGB
packet;每份只包含 48 个随机排序的 opaque event,并包含全窗口的因果 RGB 序列:
E:\codex-tools\bin\blindassist-python.cmd -m scripts.research.eval_validity_r0.prepare_p1_review_packets `
--screening-cohort artifacts.local/evidence/eval-validity-r0/screening-cohort-v4.json `
--admission-receipt artifacts.local/evidence/eval-validity-r0/data-admission-v1.json `
--p0-agreement artifacts.local/evidence/eval-validity-r0/p0-anchor-agreement-v1.json `
--materialized-root artifacts.local/evidence/eval-validity-r0/continuous-native-inputs-v1 `
--reviewer-a-root artifacts.local/evidence/eval-validity-r0/p1-reviewer-a-v1 `
--reviewer-b-root artifacts.local/evidence/eval-validity-r0/p1-reviewer-b-v1 `
--private-root artifacts.local/evidence/eval-validity-r0/p1-private-map-v1两位新的 reviewer 独立提交后,custodian 运行同样 JSON-only 的 P1 归档。两人须对 knownness 与
两类区间完全一致,而且每个 P1 区间必须逐点包含/排除 P0 已冻结 anchor;任一不一致、UNKNOWN 或
与 P0 冲突都会 STOP_FULL_EVENT_FACT_CONSISTENCY_NOT_ESTABLISHED:
E:\codex-tools\bin\blindassist-python.cmd -m scripts.research.eval_validity_r0.finalize_p1_action_facts `
--screening-cohort artifacts.local/evidence/eval-validity-r0/screening-cohort-v4.json `
--admission-receipt artifacts.local/evidence/eval-validity-r0/data-admission-v1.json `
--p0-agreement artifacts.local/evidence/eval-validity-r0/p0-anchor-agreement-v1.json `
--private-map artifacts.local/evidence/eval-validity-r0/p1-private-map-v1/private-review-map.json `
--packet-a artifacts.local/evidence/eval-validity-r0/p1-reviewer-a-v1/packet.json `
--packet-b artifacts.local/evidence/eval-validity-r0/p1-reviewer-b-v1/packet.json `
--review-a artifacts.local/evidence/eval-validity-r0/p1-action-review-a-v1.json `
--review-b artifacts.local/evidence/eval-validity-r0/p1-action-review-b-v1.json `
--output artifacts.local/evidence/eval-validity-r0/p1-action-facts-v1.json通过后,P1 action facts 才能与独立冻结的 source-mask scene facts 合并为
full-event-facts;它含 P0 agreement hash、两份 full-review hash、零 unresolved units,以及
每个正例完整的 alertable/passed interval。只有它、scene-frame ledger、trace manifest 和全帧
feedback traces 都冻结后,运行:
E:\codex-tools\bin\blindassist-python.cmd -m scripts.research.eval_validity_r0.finalize_scene_facts `
--screening-cohort artifacts.local/evidence/eval-validity-r0/screening-cohort-v4.json `
--admission-receipt artifacts.local/evidence/eval-validity-r0/data-admission-v1.json `
--p0-agreement artifacts.local/evidence/eval-validity-r0/p0-anchor-agreement-v1.json `
--p1-action-facts artifacts.local/evidence/eval-validity-r0/p1-action-facts-v1.json `
--materialized-root artifacts.local/evidence/eval-validity-r0/continuous-native-inputs-v1 `
--output-root artifacts.local/evidence/eval-validity-r0/frozen-scene-event-facts-v1该步骤把原生 SANPO panoptic mask 的固定 R class-ID、G/B instance-ID 映射为可复算的
source scene mask/box(不是“应提醒”标签),再按合同中已冻结的 source-mask stratum × P1 action
interval 规则形成 12-12-12-12 bucket。任一 profile/action 不匹配或桶数量不符,都只产生
HOLD_EVAL_VALIDITY_DATA receipt;不得重抽事件、重标或调阈值。它输出的 cohort-v1.json、
full-event-facts-v1.json 和 scene-facts.jsonl 才是之后 trace runner 的输入。
E:\codex-tools\bin\blindassist-python.cmd -m scripts.research.eval_validity_r0.evaluate `
--contract docs/research/dual-loop/EVAL_VALIDITY_R0_CONTRACT_2026-08-02.json `
--exclusion-registry artifacts.local/evidence/eval-validity-r0/exclusion-registry-v1.json `
--cohort artifacts.local/evidence/eval-validity-r0/frozen-scene-event-facts-v1/cohort-v1.json `
--p0-agreement artifacts.local/evidence/eval-validity-r0/p0-anchor-agreement-v1.json `
--p1-action-facts artifacts.local/evidence/eval-validity-r0/p1-action-facts-v1.json `
--full-event-facts artifacts.local/evidence/eval-validity-r0/frozen-scene-event-facts-v1/full-event-facts-v1.json `
--scene-frames artifacts.local/evidence/eval-validity-r0/scene-frames-v1.jsonl `
--trace-manifest artifacts.local/evidence/eval-validity-r0/trace-manifest-v1.json `
--feedback-traces artifacts.local/evidence/eval-validity-r0/feedback-traces-v1.jsonl `
--output artifacts.local/evidence/eval-validity-r0/result-v1.json输入不变量:cohort 必须与 registry 的每一个 native source session 不重叠;review A/B 需要
独立、opaque 且在 trace access 前提交;P1 的 full-event facts 同样必须先冻结;scene-frames
只允许 current_yolo、truth_box、truth_mask,且每行必须绑定冻结的 source scene-fact manifest;
feedback-traces 必须有四臂,并覆盖每个冻结 event 的每一帧 delivered-feedback 布尔值。输出目录不可覆盖。
P0 review 的顶层字段为 schema_version、protocol_id、reviewer_role、screening_cohort_sha256、
isolated_context=true、other_review_visible_before_submission=false、
model_or_oracle_output_visible=false 与 items;每个 item 只含一个 opaque review_item_id
及其唯一的 anchor.frame_index / reminder_now / cleared / knownness,不能把同一 event 的多个
anchor 放进一个 item。P1 full-event-facts
还必须绑定 P0 agreement SHA、两份独立 full-review SHA,以及每个正例的完整 interval;审计器
不会接受只带汇总结论或只带 anchor interval 的替代文件。
在审计器产生 result-v1.json 后,可以渲染面向研究决策的双层 Markdown report:
E:\codex-tools\bin\blindassist-python.cmd -m scripts.research.eval_validity_r0.render_report `
--result artifacts.local/evidence/eval-validity-r0/result-v1.json `
--output artifacts.local/evidence/eval-validity-r0/result-v1.md只写入 artifacts.local/evidence/eval-validity-r0/ 的显式新目录。registry 记录历史角色的
hash 和 41 个当前排除 session;最终 result-v1.json 分开报告 representation 与 event
quality,并保存全部输入 SHA-256。输入 materializer 只下载冻结的原生 RGB/mask,绝不下载或
运行模型/oracle/Android;evaluator 本身不训练模型或运行 Android。
四个裁判测试的独立合同见
JUDGE_AUDIT_R0_CONTRACT_2026-08-02.md。裁判员只提交原子观察,不提交
ACTIONABLE/NON_ACTIONABLE;审计器按冻结规则派生动作标签。review map 使用不含语义的
opaque ID,且 causal packet 不暴露模型输出、候选类别或选择理由。每个 event 还绑定一个仅供
审计分析的 discovery_arm;正式 cohort 至少包含 source-mask 与一个独立 arm。先只做候选
preflight:
E:\codex-tools\bin\blindassist-python.cmd -m scripts.research.eval_validity_r0.judge_audit_preflight `
--discovery artifacts.local/evidence/eval-validity-r0/unseen-discovery-selection-v3.json `
--discovery-arm source_mask `
--discovery artifacts.local/evidence/eval-validity-r0/sanpo-train-normal-discovery-v1.json `
--discovery-arm source_mask `
--exclusion-registry artifacts.local/evidence/eval-validity-r0/exclusion-registry-v1.json `
--output artifacts.local/evidence/eval-validity-r0/judge-audit-preflight-v6.json该输出只回答“是否有足够独立候选开始下一阶段”,不会把 source-mask profile 变成 event truth;
当前 discovery mix 仍未建立,正式分母关闭。pilot 的时序固定为:先 output-blind 冻结 8–12
个事件,完成两份 causal 与一份 retrospective primitive review 并封存;之后才让 YOLO 以
SELECTION_ONLY 读取框、尺度、位置和标签无关的 selection_time_slot,deterministic 枚举
并按冻结排序取前 3–4 对;pair-builder 明确不能读取 reviewed_event_phase 或
reviewed_motion_relation。最后生成 native/system-chain oracle trace,再运行
judge_audit。若 pair 不足,保持 NOT_EVALUABLE/HOLD,不得回头按标签挑样本。
系统链仍沿用统一 decision kernel;同时必须提供不经过 YOLO-shaped kernel 的 native
information-ceiling 路径。缺失 depth/geometry/trajectory 或出现 UNKNOWN 时保持
NOT_EVALUABLE,primitive 与 derived stability 分开报告;不得用 retrospective 结果替代
causal 裁决。
当前第一轮 burned pilot 的失败结果保留在
artifacts.local/evidence/eval-validity-r0/judge-burned-pilot-v1/,不可与新边界下的 review
混用。修订后的 v2 packet 位于
artifacts.local/evidence/eval-validity-r0/judge-burned-pilot-v2/:选取未消费的 009–016 八个
事件、两份 causal 前缀可见 RGB packet 和一份 retrospective 全序列 RGB packet。v2 将
visibility 固定为 CURRENT_RGB_FRAME_ONLY,每个 frame card 同时提供
current_rgb_frame、visibility_rgb_frames 与字段专用 temporal_rgb_frames;模糊、旋转和
关系不确定仍由 evidence_quality/对应 primitive 记录。packet 只要求六个 primitive,不含
mask、模型输出、YOLO、oracle 或 action label。三个 reviewer 提交后,先用
seal_judge_review_bundle 生成只暴露 bundle hash 的 seal,再用
select_judge_counterfactual_pairs 读取完整的 selection-only candidate universe,按冻结排序
生成 pair manifest;若 pre-label eligible 数不足,保留 NOT_EVALUABLE,不回挑样本。
v2 的新 plan binding 与 RGB-only staging 分别由
rebind_judge_native_plan 和 stage_judge_rgb_subset 生成,并各自写入 calibration receipt;
它们不改变既有 RGB frame receipt,不带 source mask,也不打开 formal denominator。
v4 在 v2 之后进一步冻结 primitive_observability_v4:所有几何 primitive 共享
route_anchor(当前承载相机的行人支撑面及其正前方连续延伸),path_relation、
route_certainty 和 evidence_quality 只能看当前 RGB 帧;只有 motion_relation 与
phase 才按 causal past prefix / retrospective full event 使用 temporal frames。phase 明确定义为
当前 path 占用相对于允许前缀的时相:当前 non-blocking 且此前没有 blocking 为
BEFORE_INTRUSION,当前 blocking 为 CURRENT_INTRUSION,当前 non-blocking 且此前出现
blocking 为 PASSED_CLEAR,只有 path/锚点/证据/允许前缀不足才为 UNKNOWN;causal 不看未来,
retrospective 不改写 current-only 几何字段。几何字段的
AMBIGUOUS、UNKNOWN 和 INSUFFICIENT 语义不再由“是否提醒”反推,review packet 给出
独立 current-only frame cards 与字段级版本/窗口。v3 曾计划使用 017–028 十二个事件,但因
reviewer 未提交而不形成审阅证据;这些事件不再复用。v4 使用未消费的 029–040 十二个事件重新
烧录;v2/v3 review 与 v4 review、seal、pair manifest 不得混用。
- 旧 RISKSEG 30-event cohort 是 consumed Development evidence,禁止作为任何输入或调参来源。
- scene mask/box 仅是 source-grounded scene fact,不能成为“应提醒”或用户安全真值。
synthetic_oracle是 host-only evaluator-integrity control,永不成为模型 target 或运行时输入。- 任何 action review 的
UNKNOWN或 disagreement 都 fail closed,绝不能转为 no-alert/safe。
STOP_EVENT_FACT_CONSISTENCY_NOT_ESTABLISHED 时不读取 trace;
STOP_EVALUATOR_INTEGRITY_NOT_ESTABLISHED 时不解释真实 arm;
STOP_ORACLE_MONOTONICITY_NOT_ESTABLISHED 时修正 evaluator/adapter,而不是训练、挑 seed、
调阈值或修改风险规则。新 cohort 未达到 48 events / 48 sessions / 12-12-12-12 或污染审计
有未解决候选时,终态为 HOLD_EVAL_VALIDITY_DATA。
可证伪假设是:固定同一 decision chain 后,YOLO → truth box → truth mask → synthetic oracle
不会降低 hit、clearance 或时效性,也不会提高 false-alert event。任一逆向即表明评价器或
fact-to-evidence adapter 不能支撑模型排序。若 synthetic oracle 自身失败,优先怀疑评分/反馈
链而不是感知算法。
未通过 action-review 一致性的 RGB 包只能作为 ontology counterexample;未通过单调性的
trace 只能作为 evaluator regression fixture;未通过数据隔离或 pHash 人工判定的候选保持
HOLD,不得改名为新 event-eval。任何产物都不得包装为模型训练、选模或 App 晋级证据。