From bcddbcaba9f14b168b123d5d0983ba0af58444ae Mon Sep 17 00:00:00 2001 From: 91haoran Date: Mon, 16 Nov 2026 13:17:12 +0900 Subject: [PATCH 01/25] fix(policy): dedicated new-project H6 validation, bind option+digest+source, user-facing next action --- core/human_decision_review.py | 50 ++++++++++++++++++++++++++++++----- 1 file changed, 43 insertions(+), 7 deletions(-) diff --git a/core/human_decision_review.py b/core/human_decision_review.py index 4377447..4d7a094 100644 --- a/core/human_decision_review.py +++ b/core/human_decision_review.py @@ -7,6 +7,7 @@ import argparse import json +import re from pathlib import Path # A machine-readable mirror of the scientific review definitions in WORKFLOW.md. @@ -138,21 +139,30 @@ def review(packet: dict) -> dict: continue present.add(name) - # Opt-in new-project H6 bundle consistency. H3-H5 may be drafted by AI, + # Explicit new-project H6 bundle consistency. H3-H5 may be drafted by AI, # but after the single concrete H6 reply their *author* decision records # must cite the SAME original user message. Never infer approval from the # flag or a string pointer; only a trusted host can authenticate messages. - if packet.get("new_project_h6_bundle") is True and target in { - "H6", "H7", "H8", "H9", "H10"}: + is_new_h6 = packet.get("new_project_h6_bundle") is True + if is_new_h6 and target in {"H6", "H7", "H8", "H9", "H10"}: h6 = rows.get("H6") original = h6.get("original_user_message_ref") if isinstance(h6, dict) else None - if not isinstance(original, str) or not original.strip(): + digest = h6.get("plan_sha256") if isinstance(h6, dict) else None + if (not isinstance(original, str) or not original.strip() + or not isinstance(digest, str) + or re.fullmatch(r"[0-9a-f]{64}", digest) is None + or h6.get("kind") not in {"USER_CONFIRMED", "USER_DELEGATED"}): malformed.append("H6") for step in ("H3", "H4", "H5"): entry = rows.get(step) if (not isinstance(entry, dict) or - entry.get("kind") not in {"USER_CONFIRMED", "USER_DELEGATED"} or - entry.get("original_user_message_ref") != original): + entry.get("kind") != h6.get("kind") + or entry.get("original_user_message_ref") != original + or entry.get("plan_sha256") != digest + or (h6.get("kind") == "USER_CONFIRMED" + and entry.get("chosen_option") != h6.get("chosen_option")) + or (h6.get("kind") == "USER_DELEGATED" + and entry.get("delegated_scope") != h6.get("delegated_scope"))): malformed.append(step) present.difference_update(malformed) @@ -195,7 +205,17 @@ def review(packet: dict) -> dict: next_review = CONTRACT["new_project_sequence"][ CONTRACT["new_project_sequence"].index(target) + 1 ] + if is_new_h6 and target in {"H6", "H7", "H8", "H9", "H10"}: + # Scientific H1-H5 dependencies stay intact. User-facing next action + # is a single H6 choice, not five separate author questionnaires. + if "H6" not in effective_present: + next_review = "H6" return { + "next_user_interaction": ( + "H6" if is_new_h6 and target in {"H6", "H7", "H8", "H9", "H10"} + and "H6" not in effective_present else None + ), + "new_project_h6_bundle_checked": is_new_h6, "target": target, "target_decision_owner": CONTRACT["steps"][target]["decision_owner"], "target_requires_user_scientific_choice": CONTRACT["steps"][target]["decision_owner"] == "USER", @@ -247,12 +267,28 @@ def review(packet: dict) -> dict: ), } +def review_new_project_h6(packet: dict) -> dict: + """Canonical new-project entry; caller cannot skip H6 bundle verification. + + Caller MUST be a trusted project-router/host for new-line applicability. + Returned review remains advisory and NEVER authenticates the source message. + """ + if not isinstance(packet, dict) or packet.get("protocol_version") != CONTRACT["protocol_version"]: + raise ValueError("New-project H6 requires current PD-RW protocol") + if packet.get("target") not in {"H6", "H7", "H8", "H9", "H10"}: + raise ValueError("New-project H6 entry requires H6 or a dependent target") + scoped = {**packet, "new_project_h6_bundle": True} + return review(scoped) + def main() -> None: parser = argparse.ArgumentParser(description="Read-only declared human decision dependency check") parser.add_argument("packet", type=Path) + parser.add_argument("--new-project-h6", action="store_true", + help="Require new-project H6 bundle, independent of caller packet flag") args = parser.parse_args() - result = review(json.loads(args.packet.read_text(encoding="utf-8"))) + packet = json.loads(args.packet.read_text(encoding="utf-8")) + result = (review_new_project_h6(packet) if args.new_project_h6 else review(packet)) print(json.dumps(result, ensure_ascii=False, indent=2)) From 2d235484071ebebc5a7a7225074c4dcd432fcc33 Mon Sep 17 00:00:00 2001 From: 91haoran Date: Mon, 16 Nov 2026 13:17:25 +0900 Subject: [PATCH 02/25] fix: avoid missing H6 row dereference during bundle validation --- core/human_decision_review.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/core/human_decision_review.py b/core/human_decision_review.py index 4d7a094..17a3b3d 100644 --- a/core/human_decision_review.py +++ b/core/human_decision_review.py @@ -146,7 +146,8 @@ def review(packet: dict) -> dict: is_new_h6 = packet.get("new_project_h6_bundle") is True if is_new_h6 and target in {"H6", "H7", "H8", "H9", "H10"}: h6 = rows.get("H6") - original = h6.get("original_user_message_ref") if isinstance(h6, dict) else None + h6 = h6 if isinstance(h6, dict) else {} + original = h6.get("original_user_message_ref") digest = h6.get("plan_sha256") if isinstance(h6, dict) else None if (not isinstance(original, str) or not original.strip() or not isinstance(digest, str) From 7349f05e0b1b9a3e287a97f0f259cab9092cf3ac Mon Sep 17 00:00:00 2001 From: 91haoran Date: Mon, 16 Nov 2026 13:17:50 +0900 Subject: [PATCH 03/25] fix: scope DEV preflight to host-known new project, never accept legacy bypass --- core/execution_preflight.py | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-) diff --git a/core/execution_preflight.py b/core/execution_preflight.py index 0969f44..dcd7597 100644 --- a/core/execution_preflight.py +++ b/core/execution_preflight.py @@ -27,11 +27,14 @@ NEEDS_RESULTS = {"SIGNIFICANT_RUN", "NEGATIVE_RESULT", "H8_CLAIM", "CLAIM_PROMOTION"} -def action_preflight(packet: dict, *, independently_authorized: bool = False) -> dict: +def action_preflight(packet: dict, *, independently_authorized: bool = False, + new_project: bool = False) -> dict: """independently_authorized is a host-only kwarg, not a packet field. A host must independently authenticate the original grant and its scope; - this function cannot do it. The CLI never passes True. + this function cannot do it. The CLI never passes True. A trusted host may + pass new_project=True to reject legacy-scope substitutes for new projects; + user/Agent packet declarations cannot determine that project identity. """ if not isinstance(packet, dict): return {"decision": "BLOCK", "reason": "MALFORMED_ACTION_PACKET"} @@ -46,6 +49,10 @@ def action_preflight(packet: dict, *, independently_authorized: bool = False) -> return {"decision": "ALLOW", "reason": "NO_RESTRICTED_EFFECT_REQUESTED", "enforcement": "ADVISORY_NOT_TOOL_SANDBOX"} if action in DEV: + # New-project identity must be supplied by a trusted host/router, never + # inferred from an Agent-controlled JSON flag or a legacy scope claim. + if new_project and packet.get("h6_author_strategy_confirmed") is not True: + return {"decision": "BLOCK", "reason": "NEW_PROJECT_H6_CHOICE_REQUIRED"} conditions = ("h5_scope_confirmed", "dev_inputs_verified", "within_h5_scope", "safe_local_nonpaid", "no_project_specific_prohibition", "within_author_selected_strategy") From 4864e1e3e9f2ce4ea4a3a0dbb84d825d3cdc8318 Mon Sep 17 00:00:00 2001 From: 91haoran Date: Mon, 16 Nov 2026 13:18:20 +0900 Subject: [PATCH 04/25] test: canonical H6 bundle and new-project preflight reject choice bypass --- tests/test_policy_integrity.py | 55 ++++++++++++++++++++++++++++++++-- 1 file changed, 53 insertions(+), 2 deletions(-) diff --git a/tests/test_policy_integrity.py b/tests/test_policy_integrity.py index 2ffdb59..92cd724 100644 --- a/tests/test_policy_integrity.py +++ b/tests/test_policy_integrity.py @@ -33,8 +33,8 @@ def user(ref): "original_user_message_ref": ref, "chosen_option": "A"} entries = {"H1": user("launch"), "H2": user("launch")} - entries.update({node: user("h6-reply") for node in - ("H3", "H4", "H5", "H6")}) + entries.update({node: {**user("h6-reply"), "plan_sha256": "a" * 64} + for node in ("H3", "H4", "H5", "H6")}) packet = {"format": "pd-rw-human-review-v1", "target": "H6", "decisions": entries, "new_project_h6_bundle": True} @@ -46,6 +46,57 @@ def user(ref): self.assertFalse(bad["can_advance_declarations_only"]) self.assertIn("H4", bad["malformed_declarations"]) + def test_canonical_new_project_h6_entry_cannot_be_skipped_by_packet_omission(self): + from core.human_decision_review import CONTRACT, review_new_project_h6 + + def declared(node, reply): + spec = CONTRACT["steps"][node] + row = {"kind": "USER_CONFIRMED", "original_user_message_ref": reply, + "chosen_option": "A", + "dimension_decisions": {x: "CHOICE_A" for x in spec["required_dimensions"]}, + "evidence_refs": {x: "source/a" for x in spec["evidence_minimum"]}} + if node in {"H3", "H4", "H5", "H6"}: + row["plan_sha256"] = "f" * 64 + return row + + rows = {node: declared(node, "launch" if node in {"H1", "H2"} else "reply") + for node in ("H1", "H2", "H3", "H4", "H5", "H6")} + packet = {"format": "pd-rw-human-review-v1", + "protocol_version": CONTRACT["protocol_version"], + "target": "H6", "decisions": rows} + valid = review_new_project_h6(packet) + self.assertTrue(valid["can_advance_declarations_only"]) + self.assertIsNone(valid["next_user_interaction"]) + self.assertFalse(valid["allows_download_training_sealed_or_writing"]) + + for change in ("missing_h6", "wrong_option", "wrong_hash", "wrong_reply"): + altered = json.loads(json.dumps(packet)) + if change == "missing_h6": + del altered["decisions"]["H6"] + elif change == "wrong_option": + altered["decisions"]["H4"]["chosen_option"] = "B" + elif change == "wrong_hash": + altered["decisions"]["H5"]["plan_sha256"] = "b" * 64 + else: + altered["decisions"]["H3"]["original_user_message_ref"] = "different" + reviewed = review_new_project_h6(altered) + self.assertFalse(reviewed["can_advance_declarations_only"], change) + self.assertEqual(reviewed["next_user_interaction"], "H6") + + def test_new_project_dev_preflight_cannot_use_legacy_scope_as_h6(self): + from core.execution_preflight import action_preflight + packet = {"action": "DEV_TRAIN", + "h5_scope_confirmed": True, "dev_inputs_verified": True, + "within_h5_scope": True, "safe_local_nonpaid": True, + "no_project_specific_prohibition": True, + "within_author_selected_strategy": True, + "legacy_unchanged_scope_confirmed": True, + "h6_author_strategy_confirmed": False} + new = action_preflight(packet, independently_authorized=True, new_project=True) + self.assertEqual(new["reason"], "NEW_PROJECT_H6_CHOICE_REQUIRED") + old = action_preflight(packet, independently_authorized=True, new_project=False) + self.assertEqual(old["decision"], "ALLOW") + def test_six_activities_preserved_and_joined_review_declared(self): activities = json.loads(read("contracts/research_activity_protocol.v1.json")) self.assertEqual(list(activities["activities"]), [f"0{i}" for i in range(1, 7)]) From e0289711319caf07bee44e3d6feda8c2df5010a6 Mon Sep 17 00:00:00 2001 From: 91haoran Date: Mon, 16 Nov 2026 13:18:55 +0900 Subject: [PATCH 05/25] fix: tie new-project activity-03 handoff to canonical H6 bundled author review --- core/contracts.py | 19 +++++++++++++++---- 1 file changed, 15 insertions(+), 4 deletions(-) diff --git a/core/contracts.py b/core/contracts.py index 93c15be..14f4556 100644 --- a/core/contracts.py +++ b/core/contracts.py @@ -124,8 +124,13 @@ def validate_state(value: Any) -> dict: # Existing six activities: declared evidence-handoff check, not a permission gate. # Scientific scope belongs to WORKFLOW.md; the JSON is a tested structural mirror. -def review_activity_handoff(packet: dict) -> dict: - """Review caller-supplied handoff metadata without opening source artifacts.""" +def review_activity_handoff(packet: dict, *, new_project: bool = False) -> dict: + """Review handoff declarations. Trusted caller supplies new-project identity. + + Pass new_project=True for a newly initiated line; this requires the H6 + bundled source/choice check at the activity-03 boundary. The boolean is + not authenticated here; WorkBuddy host identity enforcement remains external. + """ import json from pathlib import Path @@ -281,7 +286,10 @@ def review_not_applicable(name: str) -> bool: if human_packet is not None: if not isinstance(human_packet, dict): raise ContractError("human_review_packet must be an object") - from core.human_decision_review import CONTRACT as HUMAN_CONTRACT, review as review_human + from core.human_decision_review import ( + CONTRACT as HUMAN_CONTRACT, review as review_human, + review_new_project_h6, + ) if human_packet.get("protocol_version") not in ( HUMAN_CONTRACT["protocol_version"], *HUMAN_CONTRACT.get("legacy_protocol_versions", ())): @@ -293,7 +301,9 @@ def review_not_applicable(name: str) -> bool: candidate = dict(human_packet) candidate["target"] = name try: - decision = review_human(candidate) + decision = (review_new_project_h6(candidate) + if new_project and activity_id == "03" and name == "H6" + else review_human(candidate)) except ValueError as exc: raise ContractError(f"invalid human_review_packet: {exc}") from exc if not decision["can_advance_declarations_only"]: @@ -405,6 +415,7 @@ def review_not_applicable(name: str) -> bool: ) return { "activity_id": activity_id, + "new_project_h6_bundle_checked": bool(new_project and activity_id == "03"), "contract_version": contract["contract_version"], "evidence_path": evidence_path, "missing_evidence_declarations": missing, From 54c5bfc77bd520125c7592196b52a336d1602194 Mon Sep 17 00:00:00 2001 From: 91haoran Date: Mon, 16 Nov 2026 13:20:07 +0900 Subject: [PATCH 06/25] fix: keep H6 user interaction pending whenever bundled evidence is inconsistent --- core/human_decision_review.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/core/human_decision_review.py b/core/human_decision_review.py index 17a3b3d..ae50808 100644 --- a/core/human_decision_review.py +++ b/core/human_decision_review.py @@ -209,12 +209,12 @@ def review(packet: dict) -> dict: if is_new_h6 and target in {"H6", "H7", "H8", "H9", "H10"}: # Scientific H1-H5 dependencies stay intact. User-facing next action # is a single H6 choice, not five separate author questionnaires. - if "H6" not in effective_present: + if not can_advance: next_review = "H6" return { "next_user_interaction": ( "H6" if is_new_h6 and target in {"H6", "H7", "H8", "H9", "H10"} - and "H6" not in effective_present else None + and not can_advance else None ), "new_project_h6_bundle_checked": is_new_h6, "target": target, From 346bb161c8da7dfdded94d8667e87e2803918043 Mon Sep 17 00:00:00 2001 From: 91haoran Date: Mon, 16 Nov 2026 13:20:42 +0900 Subject: [PATCH 07/25] test: activity-03 routes new-project H6 bundle independently of old generic review --- tests/test_research_activity_handoff.py | 19 +++++++++++++++++++ 1 file changed, 19 insertions(+) diff --git a/tests/test_research_activity_handoff.py b/tests/test_research_activity_handoff.py index 6419339..bfecdfa 100644 --- a/tests/test_research_activity_handoff.py +++ b/tests/test_research_activity_handoff.py @@ -248,6 +248,25 @@ def test_h5_reference_unbacked_by_declared_human_decision_is_insufficient(self): self.assertEqual(review_activity_handoff(p)["human_review_declarations_missing"], ["H6"]) self.assertFalse(review_activity_handoff(p)["declared_activity_ready_not_authenticated"]) + def test_new_project_activity03_requires_bound_h6_choice(self): + p = self._entered("03") + human = _human_packet("H1", "H2", "H3", "H4", "H5", "H6") + p["human_review_packet"] = human + generic = review_activity_handoff(p) + self.assertNotIn("H6", generic["human_review_declarations_missing"]) + checked = review_activity_handoff(p, new_project=True) + self.assertTrue(checked["new_project_h6_bundle_checked"]) + self.assertIn("H6", checked["human_review_declarations_missing"]) + # The same post-card message, option, and digest is required at H3-H6. + for node in ("H3", "H4", "H5", "H6"): + human["decisions"][node].update({ + "original_user_message_ref": "synthetic/post-card-H6", + "chosen_option": "A", "plan_sha256": "a" * 64, + }) + bound = review_activity_handoff(p, new_project=True) + self.assertNotIn("H6", bound["human_review_declarations_missing"]) + self.assertFalse(bound["declared_activity_ready_not_authenticated"]) + def test_nonmethod_scope_has_disposition_not_fake_experiment_pass(self): p = self._entered("04", "NON_DATA") p["execution_kind"] = "NO_CONTROLLED_EXPERIMENT" From 4e4389e02e8931347bbcfac94df1de67073fb9ea Mon Sep 17 00:00:00 2001 From: 91haoran Date: Mon, 16 Nov 2026 13:21:40 +0900 Subject: [PATCH 08/25] fix: new-project H6 CLI diagnostic never returns success as an execution authorization --- core/human_decision_review.py | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) diff --git a/core/human_decision_review.py b/core/human_decision_review.py index ae50808..1cb9ea2 100644 --- a/core/human_decision_review.py +++ b/core/human_decision_review.py @@ -282,7 +282,7 @@ def review_new_project_h6(packet: dict) -> dict: return review(scoped) -def main() -> None: +def main() -> int: parser = argparse.ArgumentParser(description="Read-only declared human decision dependency check") parser.add_argument("packet", type=Path) parser.add_argument("--new-project-h6", action="store_true", @@ -291,7 +291,10 @@ def main() -> None: packet = json.loads(args.packet.read_text(encoding="utf-8")) result = (review_new_project_h6(packet) if args.new_project_h6 else review(packet)) print(json.dumps(result, ensure_ascii=False, indent=2)) + # A new-project H6 declaration alone is never a trusted execution grant. + # Fail closed for automation that only inspects the process exit code. + return 3 if args.new_project_h6 else 0 if __name__ == "__main__": - main() + raise SystemExit(main()) From d44b80be19102a46fb11e3d32e9ed7f94fc248cf Mon Sep 17 00:00:00 2001 From: 91haoran Date: Mon, 16 Nov 2026 13:22:37 +0900 Subject: [PATCH 09/25] policy: document canonical new-project H6 checks and host boundary in contracts/human_review_protocol.v1.json --- contracts/human_review_protocol.v1.json | 10 +++++++++- 1 file changed, 9 insertions(+), 1 deletion(-) diff --git a/contracts/human_review_protocol.v1.json b/contracts/human_review_protocol.v1.json index 198fefd..ce54b4f 100644 --- a/contracts/human_review_protocol.v1.json +++ b/contracts/human_review_protocol.v1.json @@ -285,6 +285,14 @@ "H5", "H6" ], - "generic_continue_is_choice": false + "generic_continue_is_choice": false, + "advisory_entry": "core.human_decision_review.review_new_project_h6", + "activity_03_handoff": "core.contracts.review_activity_handoff(new_project=True)", + "choice_binding_required": [ + "original_user_message_ref", + "plan_sha256", + "chosen_option_or_delegated_scope" + ], + "trusted_host_authentication_implemented": false } } From 049afd8ddab1821438a544af46e13180e4a052a3 Mon Sep 17 00:00:00 2001 From: 91haoran Date: Mon, 16 Nov 2026 13:22:40 +0900 Subject: [PATCH 10/25] policy: document canonical new-project H6 checks and host boundary in WORKFLOW.md --- WORKFLOW.md | 2 ++ 1 file changed, 2 insertions(+) diff --git a/WORKFLOW.md b/WORKFLOW.md index 40e7571..840bd34 100644 --- a/WORKFLOW.md +++ b/WORKFLOW.md @@ -11,6 +11,8 @@ ## 1. 三个决策循环(六项活动) +**新项目执行入口约束**:可信项目路由器应调用 `core.human_decision_review.review_new_project_h6()`(CLI 使用 `--new-project-h6`),活动03交接应调用 `core.contracts.review_activity_handoff(packet, new_project=True)`,DEV 预检需从宿主提供 `new_project=True`。新的 H3–H6 选择必须绑定相同原始作者回复、具体方案选项与 `plan_sha256`;仅持有此摘要/声明仍不是用户真实身份认证。旧项目使用原恢复路径,未核项目身份不得自动按新项目或旧项目裁定。诊断 CLI 的退出码 3 表示未取得宿主认证,不能当作要求用户重复回答。 + **新项目 H6 单次方案选择(前期交互优先规则)**:启动 Prompt 已明确的 H1 目标/资源与 H2 偏好直接作为用户原始配置,绝不再要求重复填写。Agent 在 H3–H5 自主检索、构建和反证候选科学问题、D1/PROOF_PATH 与方案边界;这些内容在用户选择前仅是 `PROPOSED`,不能伪记为 `USER_CONFIRMED`。随后向用户一次展示具体 H6 科研方案卡(首选、真正不同的备选、近邻、真值/独立分组、混杂风险、决定性最小验证、成本与不确定性),用户的一条**针对当前方案的明确回复**同时确定所选方案对应的 H3、H4、H5 和 H6。四节点仍分别记录原语义及**同一真实用户消息**,不拆成四次确认。泛泛的“继续”“按工作流执行”、启动 Prompt 中的宽泛探索委托、沉默或 Agent 自荐不能代替当前 H6 选择。未完成时只暂停依赖新方案的 Task 冻结与主线 DEV 执行;合法独立只读检查可继续。旧项目已有效的选择不重审,普通 H7 迭代不重问;H8 与下载、SEALED、费用、写稿等独立权限保持原样。 ### 新项目默认使用者深度参与模式:决策、依赖与冲突的唯一总控 From 1be485fb45a6320b14391a249a8585bb178b7dbd Mon Sep 17 00:00:00 2001 From: 91haoran Date: Mon, 16 Nov 2026 13:22:43 +0900 Subject: [PATCH 11/25] policy: document canonical new-project H6 checks and host boundary in core/README.md --- core/README.md | 2 ++ 1 file changed, 2 insertions(+) diff --git a/core/README.md b/core/README.md index 3acdafc..9662d0f 100644 --- a/core/README.md +++ b/core/README.md @@ -4,6 +4,8 @@ **人工科研决策依赖预检(只读、非授权)**:新项目按 [WORKFLOW.md](../WORKFLOW.md) 的 H1–H10 在用户选择/有限委托后记录原消息指针;可用 `python -m core.human_decision_review HUMAN_DECISIONS.json` 检查**声明的**依赖是否缺失、是否存在没有消息指针的自称批准。输入格式 `{"format":"pd-rw-human-review-v1","target":"H5","decisions":{"H1":{"kind":"USER_CONFIRMED","original_user_message_ref":"original-chat-ref","chosen_option":"scope"}, "H2":{"kind":"USER_DELEGATED","original_user_message_ref":"original-chat-ref","delegated_scope":"compare sources"}}}`,可补 H3/H4 等节点。若此前 H4 库的真值/可用性被新证据推翻,可在只读输入加 `"invalidated_decisions":["H4"]`,结构预检会把 H4 及其依赖的下游**旧声明**标记待审;若需改研究问题,同样需声明并审查 H3/H5。该列表本身是调用方声明,不能证明真实新证据或重新授权。**恢复时不得一次删除全部标记**:旧H4由用户重新选择后,应将未复核的H5(及其下游)继续列入 `invalidated_decisions`;只有其依赖关系经原消息或新的用户决定重新核实后,才清掉对应待复核标记。旧记录保留于 RESEARCH_RECORD,未授权的任何操作不得因此执行。输出即使为 `DECLARED_DEPENDENCIES_PRESENT_NOT_AUTHENTICATED` 也**不表示**原消息真实、创新性已核、选题已批准或获得下载/GPU/SEALED/写作许可;须由执行Agent核原用户消息与证据。不是新的 CLI 状态、科研阶段或第三份项目记录。 +**新项目 H6 专用程序入口**:`python -m core.human_decision_review DECISIONS.json --new-project-h6` 仅做声明/依赖检查,并固定启用同一 H6 选项、摘要和原始消息引用的一致性检查;该诊断始终返回退出码 3,不能成为操作授权。宿主应调用 `review_new_project_h6()` 并为活动03交接提供可信的 `new_project=True` 上下文;离线 JSON 不能证明真实用户身份。已获当前对话直接选择的合法 DEV 工作按真实原消息及其他权限连续开展,不在每个 Run 重新询问 H6。 + **新项目 H6 一次选择**:H1/H2 使用原始启动配置,Agent 自行调查 H3–H5 候选且不能提前标 USER_CONFIRMED;用户在完整 H6 卡片后一次明确选择,所选 H3–H6 共用同一真实消息引用。该合同的字段检查无法认证消息来源,不能替代宿主执行拦截。 **统一审核合同(新项目默认 v1.3)**:`contracts/human_review_protocol.v1.json` 的现行 `protocol_version` 为 `pd-rw-human-review-v1.3`(输入外层 `format=pd-rw-human-review-v1` 不变,二者不是同一个版本号)。使用 `python -m core.human_decision_review HUMAN_DECISIONS.json` 只读检查声明的 H1–H5 用户决定,以及 H6 核心 Task/研究策略、H8 核心 Claim/重大升级的作者选择或明确限域委托;H7 为 AI 在已批准范围内的技术决策。每个适用的已完成节点保留 `original_user_message_ref`、所需 `dimension_decisions`、`evidence_refs` 和选择/委托范围。`can_present_review_declarations_only` / `can_advance_declarations_only` 只能证明字段声明齐全,不能认证原始用户消息、科研证据或下载/训练/SEALED/绘图/写稿权限。历史 v1.1/v1.2 仅在旧项目来源与范围连续性可核实时兼容读取,不能称为新项目默认协议或把旧 `AGENT_DECIDED` 冒充 v1.3 作者批准;详见 [WORKFLOW.md](../WORKFLOW.md) 与 [AUTHOR_STRATEGY_REVIEW.md](../AUTHOR_STRATEGY_REVIEW.md)。 From c0bd6ae165e53506ef6271b92716f2a5b5657066 Mon Sep 17 00:00:00 2001 From: 91haoran Date: Mon, 16 Nov 2026 13:22:46 +0900 Subject: [PATCH 12/25] policy: document canonical new-project H6 checks and host boundary in AGENTS.md --- AGENTS.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/AGENTS.md b/AGENTS.md index c122baf..7648fde 100644 --- a/AGENTS.md +++ b/AGENTS.md @@ -11,7 +11,7 @@ 3. 决策冲突:用户原始授权优先;科研流程由 WORKFLOW 主控;科研事实以实际合法原始工件为准;专业细则仅由对应权威专题负责。既有项目直接续做,不重开选题/数据/训练。 4. **适用规则重核(不是全库重读)**:新会话/Agent接手或上下文压缩后,先核 `RULE_REPO | RULE_COMMIT | LOADED_FILES` 与当前任务的可恢复证据;切换项目、Task/活动、版本或受影响协议时,只重核新增适用的权威专题;新数据下载、收费/危险安装、超范围CPU/GPU、SEALED、绘图、写稿、费用、投稿等受限操作**执行前**再核对应规则和原始许可;H5/H6已选定或原旧项目范围有效时的安全非付费本地DEV安装和CPU/GPU拟合按 [WORKFLOW.md](WORKFLOW.md) 的 `AI_DEV_AUTONOMY` 自主执行,无须逐次确认。已知版本、任务、权限与证据未变时复用;不能确认受限操作的许可时停止该操作并定位原指示,允许无依赖只读工作。Markdown被阅读过不等于工具实际具备权限拦截或Agent会永久记忆。 -**人工科学决策冷启动**:新项目 H1/H2 直接复用已有启动配置、H3–H5 由 Agent 先作未批准的候选分析,只在 H6 提交一张具体科研方案卡供作者一次选择(同时确定 H3–H6);**H6核心Task/研究组合与H8核心Claim/重大升级需先由AI给带依据的首选和可行备选,再由作者选择或明确范围委托**。H7技术方案、既定策略内DEV算法/训练/修复仍由AI自主,不逐次询问。旧项目H6/H8原记录保留但旧AGENT_DECIDED不能伪装为新作者确认,只有实质战略变动时按新版审核。详见 [AUTHOR_STRATEGY_REVIEW.md](AUTHOR_STRATEGY_REVIEW.md)、[WORKFLOW.md](WORKFLOW.md);原H9/H10、SEALED与写稿门禁不变。 +**H6 可信入口提示**:新项目 H3–H6 一次用户选择的声明检查调用 `review_new_project_h6`,活动03交接使用 `review_activity_handoff(..., new_project=True)`;这两个接口本身并不认证原始聊天消息,未连接宿主可信身份时不能声称不可绕过。Agent 看到真实当前对话的具体 H6 回复后不重复询问,继续已获合法 DEV 范围内的工作。\n\n**人工科学决策冷启动**:新项目 H1/H2 直接复用已有启动配置、H3–H5 由 Agent 先作未批准的候选分析,只在 H6 提交一张具体科研方案卡供作者一次选择(同时确定 H3–H6);**H6核心Task/研究组合与H8核心Claim/重大升级需先由AI给带依据的首选和可行备选,再由作者选择或明确范围委托**。H7技术方案、既定策略内DEV算法/训练/修复仍由AI自主,不逐次询问。旧项目H6/H8原记录保留但旧AGENT_DECIDED不能伪装为新作者确认,只有实质战略变动时按新版审核。详见 [AUTHOR_STRATEGY_REVIEW.md](AUTHOR_STRATEGY_REVIEW.md)、[WORKFLOW.md](WORKFLOW.md);原H9/H10、SEALED与写稿门禁不变。 **论文蓝图检查闭环**:科研方向或Task明确即自主创建 `PAPER_BLUEPRINT.md`(默认六章、H1/H2、逐节拟写内容和Claim–Evidence);重要Run、关键负结果、核心Claim/路线变化、阶段交接时,Agent在已有 `RESEARCH_RECORD.md` 第5节事件填写 `EVENT_KIND`、`BLUEPRINT_CHECK: UPDATED/NO_CHANGE_NEEDED`、`BLUEPRINT_SECTIONS`、`BLUEPRINT_EVIDENCE` 和 `BLUEPRINT_REASON`,按原证据核受影响SEC_ID。普通调参、重复测试、无实质Debug不触发。不新增用户确认;已批准重大结构变动仍须真实作者确认;`WRITING_LOCKED`/绘图门禁不受影响。细节见 [DUAL_DOCUMENT_WORKFLOW.md](DUAL_DOCUMENT_WORKFLOW.md),`core.doc_integrity` 的静态通过不代表科学正确。 **最小防漏检查**:核当前活动、适用 H 节点、原证据和实际操作许可;按 [research_activity_protocol.v1.json](contracts/research_activity_protocol.v1.json) 将进入条件、H 决定、交接声明分别审查,禁止用 `declared_activity_ready_not_authenticated` 自授行动。H7/H9 可据证据说明不适用;`REVIEW:H1` ≠ `HEADING:H1`。 From 57bc24f6c8ed445a0e013a3242d1b4070956d2a7 Mon Sep 17 00:00:00 2001 From: 91haoran Date: Mon, 16 Nov 2026 13:23:01 +0900 Subject: [PATCH 13/25] docs: repair AGENTS new-project H6 paragraph formatting --- AGENTS.md | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/AGENTS.md b/AGENTS.md index 7648fde..b0a7728 100644 --- a/AGENTS.md +++ b/AGENTS.md @@ -11,7 +11,9 @@ 3. 决策冲突:用户原始授权优先;科研流程由 WORKFLOW 主控;科研事实以实际合法原始工件为准;专业细则仅由对应权威专题负责。既有项目直接续做,不重开选题/数据/训练。 4. **适用规则重核(不是全库重读)**:新会话/Agent接手或上下文压缩后,先核 `RULE_REPO | RULE_COMMIT | LOADED_FILES` 与当前任务的可恢复证据;切换项目、Task/活动、版本或受影响协议时,只重核新增适用的权威专题;新数据下载、收费/危险安装、超范围CPU/GPU、SEALED、绘图、写稿、费用、投稿等受限操作**执行前**再核对应规则和原始许可;H5/H6已选定或原旧项目范围有效时的安全非付费本地DEV安装和CPU/GPU拟合按 [WORKFLOW.md](WORKFLOW.md) 的 `AI_DEV_AUTONOMY` 自主执行,无须逐次确认。已知版本、任务、权限与证据未变时复用;不能确认受限操作的许可时停止该操作并定位原指示,允许无依赖只读工作。Markdown被阅读过不等于工具实际具备权限拦截或Agent会永久记忆。 -**H6 可信入口提示**:新项目 H3–H6 一次用户选择的声明检查调用 `review_new_project_h6`,活动03交接使用 `review_activity_handoff(..., new_project=True)`;这两个接口本身并不认证原始聊天消息,未连接宿主可信身份时不能声称不可绕过。Agent 看到真实当前对话的具体 H6 回复后不重复询问,继续已获合法 DEV 范围内的工作。\n\n**人工科学决策冷启动**:新项目 H1/H2 直接复用已有启动配置、H3–H5 由 Agent 先作未批准的候选分析,只在 H6 提交一张具体科研方案卡供作者一次选择(同时确定 H3–H6);**H6核心Task/研究组合与H8核心Claim/重大升级需先由AI给带依据的首选和可行备选,再由作者选择或明确范围委托**。H7技术方案、既定策略内DEV算法/训练/修复仍由AI自主,不逐次询问。旧项目H6/H8原记录保留但旧AGENT_DECIDED不能伪装为新作者确认,只有实质战略变动时按新版审核。详见 [AUTHOR_STRATEGY_REVIEW.md](AUTHOR_STRATEGY_REVIEW.md)、[WORKFLOW.md](WORKFLOW.md);原H9/H10、SEALED与写稿门禁不变。 +**H6 可信入口提示**:新项目 H3–H6 一次用户选择的声明检查调用 `review_new_project_h6`,活动03交接使用 `review_activity_handoff(..., new_project=True)`;这两个接口本身并不认证原始聊天消息,未连接宿主可信身份时不能声称不可绕过。Agent 看到真实当前对话的具体 H6 回复后不重复询问,继续已获合法 DEV 范围内的工作。 + +**人工科学决策冷启动**:新项目 H1/H2 直接复用已有启动配置、H3–H5 由 Agent 先作未批准的候选分析,只在 H6 提交一张具体科研方案卡供作者一次选择(同时确定 H3–H6);**H6核心Task/研究组合与H8核心Claim/重大升级需先由AI给带依据的首选和可行备选,再由作者选择或明确范围委托**。H7技术方案、既定策略内DEV算法/训练/修复仍由AI自主,不逐次询问。旧项目H6/H8原记录保留但旧AGENT_DECIDED不能伪装为新作者确认,只有实质战略变动时按新版审核。详见 [AUTHOR_STRATEGY_REVIEW.md](AUTHOR_STRATEGY_REVIEW.md)、[WORKFLOW.md](WORKFLOW.md);原H9/H10、SEALED与写稿门禁不变。 **论文蓝图检查闭环**:科研方向或Task明确即自主创建 `PAPER_BLUEPRINT.md`(默认六章、H1/H2、逐节拟写内容和Claim–Evidence);重要Run、关键负结果、核心Claim/路线变化、阶段交接时,Agent在已有 `RESEARCH_RECORD.md` 第5节事件填写 `EVENT_KIND`、`BLUEPRINT_CHECK: UPDATED/NO_CHANGE_NEEDED`、`BLUEPRINT_SECTIONS`、`BLUEPRINT_EVIDENCE` 和 `BLUEPRINT_REASON`,按原证据核受影响SEC_ID。普通调参、重复测试、无实质Debug不触发。不新增用户确认;已批准重大结构变动仍须真实作者确认;`WRITING_LOCKED`/绘图门禁不受影响。细节见 [DUAL_DOCUMENT_WORKFLOW.md](DUAL_DOCUMENT_WORKFLOW.md),`core.doc_integrity` 的静态通过不代表科学正确。 **最小防漏检查**:核当前活动、适用 H 节点、原证据和实际操作许可;按 [research_activity_protocol.v1.json](contracts/research_activity_protocol.v1.json) 将进入条件、H 决定、交接声明分别审查,禁止用 `declared_activity_ready_not_authenticated` 自授行动。H7/H9 可据证据说明不适用;`REVIEW:H1` ≠ `HEADING:H1`。 From de5d06762095485aa08651926547a6f60011ad5a Mon Sep 17 00:00:00 2001 From: 91haoran Date: Mon, 16 Nov 2026 13:23:49 +0900 Subject: [PATCH 14/25] test: replace obsolete v1.2 agent-owned H6/H8 expectations with v1.3 author strategy and H7 autonomy --- tests/test_dev_autonomy.py | 27 +++++++++++++++------------ 1 file changed, 15 insertions(+), 12 deletions(-) diff --git a/tests/test_dev_autonomy.py b/tests/test_dev_autonomy.py index d9c1a99..4e261ee 100644 --- a/tests/test_dev_autonomy.py +++ b/tests/test_dev_autonomy.py @@ -1,4 +1,4 @@ -"""Deterministic H1-H5 user / H6-H8 agent decision regression. +"""Current v1.3 H6/H8 user strategy and H7 agent DEV autonomy regression. Declarations are not authentication, real tool permission or Agent behavior. """ @@ -33,7 +33,7 @@ def packet(target, included): "protocol_version": CONTRACT["protocol_version"], "target": target, "decisions": { - step: choice(step, agent=step in {"H6", "H7", "H8"}) + step: choice(step, agent=step == "H7") for step in included }, } @@ -41,21 +41,24 @@ def packet(target, included): class DevAutonomyContractTests(unittest.TestCase): def test_owner_partition_and_protocol_version(self): - self.assertEqual(CONTRACT["protocol_version"], "pd-rw-human-review-v1.2") - for step in ("H1", "H2", "H3", "H4", "H5", "H9", "H10"): + self.assertEqual(CONTRACT["protocol_version"], "pd-rw-human-review-v1.3") + for step in ("H1", "H2", "H3", "H4", "H5", "H6", "H8", "H9", "H10"): self.assertEqual(CONTRACT["steps"][step]["decision_owner"], "USER") - for step in ("H6", "H7", "H8"): - self.assertEqual(CONTRACT["steps"][step]["decision_owner"], "AGENT") + self.assertEqual(CONTRACT["steps"]["H7"]["decision_owner"], "AGENT") - def test_h6_h7_h8_are_agent_decisions_without_fake_user_grants(self): + def test_h6_h8_are_author_decisions_h7_agent_autonomy(self): done = ("H1", "H2", "H3", "H4", "H5", "H6", "H7", "H8") for step in ("H6", "H7", "H8"): with self.subTest(step=step): r = review(packet(step, done)) self.assertTrue(r["can_advance_declarations_only"]) - self.assertEqual(r["target_decision_owner"], "AGENT") - self.assertFalse(r["target_requires_user_scientific_choice"]) - self.assertIn(step, r["agent_decision_declarations"]) + expected = "AGENT" if step == "H7" else "USER" + self.assertEqual(r["target_decision_owner"], expected) + self.assertEqual(r["target_requires_user_scientific_choice"], step != "H7") + if step == "H7": + self.assertIn(step, r["agent_decision_declarations"]) + else: + self.assertNotIn(step, r["agent_decision_declarations"]) self.assertFalse(r["allows_download_training_sealed_or_writing"]) # H7 remains optional for scientific discovery/no new algorithm. r = review(packet("H8", ("H1", "H2", "H3", "H4", "H5", "H6", "H8"))) @@ -67,12 +70,12 @@ def test_no_h5_no_agent_forging_or_missing_provenance(self): self.assertFalse(r["can_advance_declarations_only"]) self.assertIn("H5", r["missing_prerequisites"]) p = packet("H6", ("H1", "H2", "H3", "H4", "H5", "H6")) - del p["decisions"]["H6"]["h5_scope_ref"] + del p["decisions"]["H6"]["original_user_message_ref"] r = review(p) self.assertIn("H6", r["malformed_declarations"]) self.assertFalse(r["can_advance_declarations_only"]) p = packet("H6", ("H1", "H2", "H3", "H4", "H5", "H6")) - del p["decisions"]["H6"]["evidence_refs"]["falsifier_and_comparator"] + del p["decisions"]["H6"]["evidence_refs"][CONTRACT["steps"]["H6"]["evidence_minimum"][0]] r = review(p) self.assertIn("H6", r["missing_evidence_refs"]) self.assertFalse(r["can_advance_declarations_only"]) From 22472295d34f8c4fa6953119c2b0ddce844529ab Mon Sep 17 00:00:00 2001 From: 91haoran Date: Mon, 16 Nov 2026 13:24:02 +0900 Subject: [PATCH 15/25] fix: malformed unhashable claim status must return BLOCK_HANDOFF, not raise TypeError --- core/execution_preflight.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/core/execution_preflight.py b/core/execution_preflight.py index dcd7597..b7fe0ab 100644 --- a/core/execution_preflight.py +++ b/core/execution_preflight.py @@ -218,7 +218,7 @@ def science_checkpoint(packet: dict) -> dict: if not isinstance(packet.get("claim_status"), str) or packet.get("claim_status") not in {"HYPOTHESIZED", "EVIDENCE_CANDIDATE", "SUPPORTED", "CONFIRMED"}: missing.append("claim_status") - if packet.get("claim_status") in {"SUPPORTED", "CONFIRMED"}: + if isinstance(packet.get("claim_status"), str) and packet["claim_status"] in {"SUPPORTED", "CONFIRMED"}: actual_ref = packet.get("actual_evidence_ref_or_not_run") if (not _present(actual_ref) or actual_ref.strip().upper() in { "NOT_RUN", "UNKNOWN", "NONE", "N/A", "UNAVAILABLE", "INVALID"}): @@ -235,7 +235,7 @@ def science_checkpoint(packet: dict) -> dict: or _present(packet.get("legacy_unchanged_claim_ref")))): missing.append("author_selected_main_claim_or_unchanged_legacy") independent_ref = packet.get("independent_validation_ref_or_unavailable") - if (packet.get("claim_status") == "CONFIRMED" + if (isinstance(packet.get("claim_status"), str) and packet["claim_status"] == "CONFIRMED" and isinstance(independent_ref, str) and independent_ref.strip().upper() in { "UNAVAILABLE", "UNKNOWN", "NOT_RUN", "NONE", "N/A", "INVALID"}): From 96259588743d4c907f345a3e4639a4f938def029 Mon Sep 17 00:00:00 2001 From: 91haoran Date: Mon, 16 Nov 2026 13:25:16 +0900 Subject: [PATCH 16/25] fix(policy): align NEXT_ACTION_POLICY.md with current H6/H8 ownership and decision output --- NEXT_ACTION_POLICY.md | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/NEXT_ACTION_POLICY.md b/NEXT_ACTION_POLICY.md index 8d40289..3beace3 100644 --- a/NEXT_ACTION_POLICY.md +++ b/NEXT_ACTION_POLICY.md @@ -25,6 +25,8 @@ **可执行的最小路径核查(已有本地Python且确有文件交付时)**:在现有交付根目录中运行 `python -m core.delivery_check --root OUTPUT_DIR --primary MAIN_RESULT --file OTHER_RESULT`(`--file`按需重复);工具只核路径存在、是否仍处于交付根目录,**不生成清单、不读取文件内容,也不证明内容正确、权限合规或SEALED隔离**。工具缺失/不可运行时由宿主直接检查文件并报告未验证范围;`PASS_PATHS_ONLY`不得写成科研/隐私/权限验证PASS。普通对话/无文件输出不调用。详情代码见 [core/delivery_check.py](core/delivery_check.py)。 +**输出语义不混用**:`NEXT_AGENT_INSTRUCTION` 仅用于已授权且可执行的后续科研任务;`USER_DECISION_REQUIRED` 表示确实需要作者选择的 H6/H8 或独立操作授权;`NO_NEXT_ACTION` 表示没有后续任务。**两种结果不可混淆**:不能把已完成工作表述成待执行命令,也不能在用户尚未决定时提前发送执行命令(禁止预先输出Agent执行指令)。真实原消息与 CPU/GPU、SEALED、写作/绘图许可分别核验,只有被真实授权覆盖的 DEV 技术工作可自主连续执行。 + **已执行结果与后续行动分开**:本节规定结果反馈;下文 A/B 继续唯一决定是否给下一步Agent指令。待用户 H 节点选择、改变主Task、增加权限时,只给推荐与必要依据,**不提供依赖待决选择的执行命令**。已有合法任务可继续,则提供精简可复制的下一步Agent指令;全部完结且无需续作时报告 `NO_NEXT_ACTION`。不另建 `COMPLETED/PARTIAL/BLOCKED` 强制状态枚举,文本状态永不等于授权。 **Token 优化是全任务成本目标,而不是单次回复最短**: @@ -44,11 +46,11 @@ - `AGENT_DECIDABLE`:已在用户真实授权、研究范围及预算内的**技术实现选择**,例如定位并修复普通Bug、复用已有Run、选最有信息量的只读诊断、重跑已授权且协议不变的受影响测试、补写非论文的科学证据摘要。Agent可自行选出**一项最佳后续动作**。 - `STRONG_RECOMMENDATION`:关键证据充分、下一技术方案明显优于备选且**无需新的用户授权或价值判断**;直接给最优执行方案,不为常规技术分歧反复要求确认。强烈建议**不覆盖权限**。 -- `USER_DECISION_REQUIRED`:H1–H5、**H6初始核心Task与研究策略、H8实质核心Claim/重大升级选择**、适用H9/H10,需要真实作者选择或限定委托;此外超H5科学目标、项目整体STOP/PIVOT、新数据下载、付费/云资源、危险工具、SEALED、外传、实际绘图、正式写作和投稿仍属现有用户门禁。普通H7算法、同H6/H8已选范围内的DEV代码/训练/修复及增量证据归AI自主;不得因每个Run/小模型修改重复要人类确认。AI推荐不是用户决定。 +- `USER_DECISION_REQUIRED`:新项目 H1/H2 复用启动配置、H3–H5 由 Agent 自行准备,**只在 H6 请求一次具体核心Task/科研方案选择;H8仅重大核心Claim/升级选择**、适用H9/H10,需要真实作者选择或限定委托;此外超H5科学目标、项目整体STOP/PIVOT、新数据下载、付费/云资源、危险工具、SEALED、外传、实际绘图、正式写作和投稿仍属现有用户门禁。普通H7算法、同H6/H8已选范围内的DEV代码/训练/修复及增量证据归AI自主;不得因每个Run/小模型修改重复要人类确认。AI推荐不是用户决定。 ### 深度参与模式中的用户科学决策(复用已有 USER_DECISION_REQUIRED) -新 PD-RW 课题默认 `USER_DEEP_PARTICIPATION`:H1–H5完成后,AI在H6提供**推荐的Task+科研组合(C/R/P)及可比较备选**让作者选择或明确委托;关键成果需要冻结主Claim/重大升级时,H8提供**保留、升级、收窄/重构**等可行路线和首选供作者决定;H7保持AI自主。卡片标准与重新触发规则见 [AUTHOR_STRATEGY_REVIEW.md](AUTHOR_STRATEGY_REVIEW.md)。真实原作者选择、选项范围与证据引用缺一不可;旧项目已有真实选择不重问;新证据只重开确有影响的战略节点。技术Bug和普通实验不进入确认循环。 +新 PD-RW 课题默认 `USER_DEEP_PARTICIPATION`:H1/H2从启动配置取得,Agent 先完成 H3–H5 候选,不逐项要求作者回复;只在 H6 提供**推荐的Task+科研组合(C/R/P)及可比较备选**让作者选择或明确委托;关键成果需要冻结主Claim/重大升级时,H8提供**保留、升级、收窄/重构**等可行路线和首选供作者决定;H7保持AI自主。卡片标准与重新触发规则见 [AUTHOR_STRATEGY_REVIEW.md](AUTHOR_STRATEGY_REVIEW.md)。真实原作者选择、选项范围与证据引用缺一不可;旧项目已有真实选择不重问;新证据只重开确有影响的战略节点。技术Bug和普通实验不进入确认循环。 **待用户决策时**输出 `DECISION_ID | RECOMMENDATION+SOURCE | OPTIONS(通常2–3项)+TRADEOFFS | REQUIRED_USER_CHOICE | DEPENDENT_ACTION_HELD`。未确认节点只暂停**依赖该决定**的定稿/资源动作,仍可在既有授权范围内进行独立合法只读核查。明确委托的选题决策不授予`DOWNLOAD/TRAIN/SEALED/WRITING`权限;项目与主张事实不能由用户偏好升级成 VERIFIED。用户决定已在本轮给出时直接引用,不机械增加一次对话。 From 312ff61f41afebd0094a5635a947a6740fc98066 Mon Sep 17 00:00:00 2001 From: 91haoran Date: Mon, 16 Nov 2026 13:25:19 +0900 Subject: [PATCH 17/25] fix(policy): align PAPER_FIRST.md with current H6/H8 ownership and decision output --- PAPER_FIRST.md | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/PAPER_FIRST.md b/PAPER_FIRST.md index 0c5d585..e892031 100644 --- a/PAPER_FIRST.md +++ b/PAPER_FIRST.md @@ -2,12 +2,14 @@ **Claim 证据类型界线**:实证性能/泛化主张必须以真实有效 Run、选择史与独立确认资格为依据;纯理论、限定仿真或审查机制结论可以分别记录 `NON_DATA` 类型的有界假设、证明和独立验证,不应产生虚假数据集、指标或确认 Run。混合主张拆分后各自承担证据义务;结构合同 PASS 不认证论文质量/科学真实性。 -**活动05与04并行的科学价值交接([研究活动合同](contracts/research_activity_protocol.v1.json))**:首个真实可信 Run、重要负证据或独立理论结果即可只读审查 Claim/具名最近邻/反解释/负结果/重要证据缺口,不强制等待全部模型开发。成熟H8属于AI在已批准H5项目范围内的科学贡献/DEV资源决定;真正项目STOP、超H5主目标变化仍属于用户决定。早期贡献筛查仅候选,不能因结构齐全、模型高分或发表愿望宣称 `SUPPORTED/CONFIRMED` 或 Q2 成功。未经用户审查总结与单独授权不写正式论文。 +**活动05与04并行的科学价值交接([研究活动合同](contracts/research_activity_protocol.v1.json))**:首个真实可信 Run、重要负证据或独立理论结果即可只读审查 Claim/具名最近邻/反解释/负结果/重要证据缺口,不强制等待全部模型开发。H8 的核心科学Claim及重大战略升级由作者选择或明确限域委托;AI 可在已选 H5/H6 范围内自主回收证据、进行普通 DEV 技术研究,不能因 Run/低成本补证重复要求作者确认。真正项目 STOP、超 H5 主目标变化仍属于用户决定。早期贡献筛查仅候选,不能因结构齐全、模型高分或发表愿望宣称 `SUPPORTED/CONFIRMED` 或 Q2 成功。未经用户审查总结与单独授权不写正式论文。 **目标**:在可控数据、时间和算力下优先形成**有重要独立知识差额、能够支持JCR Q2+投稿潜力的可信科研证据链**;额外机制/理论/模型深度仅在能弥补关键证据、提高投稿适配或经用户明确选择时追加。**不能保证接收论文**,更不能把开发集高分、工程PASS或普通基线复现包装成原创研究。本文件仅在活动03–05选题、结果诊断或构建论文路线时按需读取;不增加独立审核阶段。 **科学价值先于低档小稿完成性**:见[HIGH_VALUE_RESEARCH_DESIGN.md](HIGH_VALUE_RESEARCH_DESIGN.md)。若现有Run可以支持小贡献,但数据/对照无法排除关键竞争解释或具名近邻没有实质知识差额,则先如实标为探索资产;优先补一项决定性科学证据或比较具有更高Q2潜力的合法任务,而不是自动将低档小稿视为成功。不得借此提前撰写论文。 +**Q2 状态边界**:`Q2_POTENTIAL_SUPPORTED` 仅在具名期刊近邻、独立科学增量与真实证据充分时成立;`Q2_UNCERTAIN` 或仅探索证据不得自动晋升为成功,Q3/Q4 稿件的可写性不等同于既定 Q2 成果目标。 + **Q2科研目标**:本章所有论文路线都须按[Q2_PUBLICATION_GOAL.md](Q2_PUBLICATION_GOAL.md)评估**有证据支持的JCR Q2+投稿潜力**,优先Q1;不要求实际录用。普通专业期刊、B/C档小稿及`DRAFTABLE`均不能自动判科研项目成功。 **论文类型与高价值探索**:实证算法稿需支持方法/结果/公平对照;Critical Review/Perspective 的学术价值可体现为原始证据的批判综合、新的概念分类、重要争议澄清和可检验未来议程,**不强制要求本论文产生算法 F1 增益**。具名近邻、资料真实性、证据边界和目标期刊同类型文章仍须核实。科学问题重要但暂时无法定位 Q2 期刊时,可在科研记录文字描述 `HIGH_VALUE_UNCERTAIN_VENUE`,只作为保留探索机会的注记,**不是新状态/成功认证**;明确一项现实可证伪预测与成本/授权边界,不因期刊暂未知自动停题。 From 24ea74c60a7dab4109fea16da48a2a2760d39f4b Mon Sep 17 00:00:00 2001 From: 91haoran Date: Mon, 16 Nov 2026 13:25:23 +0900 Subject: [PATCH 18/25] fix(policy): align WORKFLOW.md with current H6/H8 ownership and decision output --- WORKFLOW.md | 2 ++ 1 file changed, 2 insertions(+) diff --git a/WORKFLOW.md b/WORKFLOW.md index 840bd34..beb56da 100644 --- a/WORKFLOW.md +++ b/WORKFLOW.md @@ -11,6 +11,8 @@ ## 1. 三个决策循环(六项活动) +**状态兼容保证**:本次 H6 交互修复不新增 Claim 状态或阶段,仍使用原 H1–H10、现有科研记录与权限合同。 + **新项目执行入口约束**:可信项目路由器应调用 `core.human_decision_review.review_new_project_h6()`(CLI 使用 `--new-project-h6`),活动03交接应调用 `core.contracts.review_activity_handoff(packet, new_project=True)`,DEV 预检需从宿主提供 `new_project=True`。新的 H3–H6 选择必须绑定相同原始作者回复、具体方案选项与 `plan_sha256`;仅持有此摘要/声明仍不是用户真实身份认证。旧项目使用原恢复路径,未核项目身份不得自动按新项目或旧项目裁定。诊断 CLI 的退出码 3 表示未取得宿主认证,不能当作要求用户重复回答。 **新项目 H6 单次方案选择(前期交互优先规则)**:启动 Prompt 已明确的 H1 目标/资源与 H2 偏好直接作为用户原始配置,绝不再要求重复填写。Agent 在 H3–H5 自主检索、构建和反证候选科学问题、D1/PROOF_PATH 与方案边界;这些内容在用户选择前仅是 `PROPOSED`,不能伪记为 `USER_CONFIRMED`。随后向用户一次展示具体 H6 科研方案卡(首选、真正不同的备选、近邻、真值/独立分组、混杂风险、决定性最小验证、成本与不确定性),用户的一条**针对当前方案的明确回复**同时确定所选方案对应的 H3、H4、H5 和 H6。四节点仍分别记录原语义及**同一真实用户消息**,不拆成四次确认。泛泛的“继续”“按工作流执行”、启动 Prompt 中的宽泛探索委托、沉默或 Agent 自荐不能代替当前 H6 选择。未完成时只暂停依赖新方案的 Task 冻结与主线 DEV 执行;合法独立只读检查可继续。旧项目已有效的选择不重审,普通 H7 迭代不重问;H8 与下载、SEALED、费用、写稿等独立权限保持原样。 From 5132c7e145d0d8a1fffb27cee95a4230932ed8e8 Mon Sep 17 00:00:00 2001 From: 91haoran Date: Mon, 16 Nov 2026 13:25:26 +0900 Subject: [PATCH 19/25] fix(policy): align START.md with current H6/H8 ownership and decision output --- START.md | 2 ++ 1 file changed, 2 insertions(+) diff --git a/START.md b/START.md index 3a26807..403c66f 100644 --- a/START.md +++ b/START.md @@ -11,6 +11,8 @@ **选择不等于证据认证**:H4若是理论/形式证明或无需静态数据库的仿真,确认可审计的 `PROOF_PATH`,不虚构D1;必须记录 H3/H4/H5 对同一科研问题的科学适配,不允许因数据方便替换用户选题。记录人工节点用 `REVIEW:H#`,论文蓝图Markdown标题的 H1/H2/H3 用 `HEADING:H#` 区分;上游决定或真值被新证据推翻时,依赖下游须重新核有效性。 +**六活动机器合同**:科研阶段交接以 [research_activity_protocol.v1.json](contracts/research_activity_protocol.v1.json) 为准;新项目进入活动03时以 H6 一次具体作者选择为前置,合法只读探索不依赖该选择。 + **新项目统一交互路线(执行镜像 [`contracts/human_review_protocol.v1.json`](contracts/human_review_protocol.v1.json))**:使用 `protocol_version=pd-rw-human-review-v1.3`;H1–H5仍为原有科研决策维度,但新项目无需逐项询问,**H6初始核心Task/科研组合、H8核心Claim/重大升级由AI先推荐选项后由作者选择或有限委托,H7及普通DEV工作由AI自主**。仅旧项目按原记录兼容读取v1.2,不得把旧AGENT_DECIDED冒充新版作者选择;`H1→H2→H3→H4→H5` 的**节点名称、必审维度和顺序固定**,而论文/实验内容与搜索深度可自适应。开始先从用户原指令逐项映射已明确的决定,不能因为只知道研究领域就假定用户选择了六类课题来源或五类贡献路线;对待定 H3–H5 先给候选证据,最终只在 H6 具体方案卡中让用户一次选择;未收到该选择前不得登记为已获作者确认。已有项目旧记录不得仅按同名H编号无证据地自动转新版。每轮展示简洁的 H1–H5 进度和**当前节点**,不是让AI自行选取当前要问的科学问题。 **新项目默认 `USER_DEEP_PARTICIPATION`**:按 [WORKFLOW.md](WORKFLOW.md) 复用用户启动配置的 H1/H2,在 Agent 自主准备 H3–H5 研究问题、D1 数据资源与下一步计划后,于 H6 一次完成方案的**真实用户选择或明确委托**。AI 在每轮先做足合法只读核查与备选比较;用户可一条消息决定多个已披露事项,已有项目复用旧选择而不强制重审。`PREFERRED`、下载计划或 AI 自选库均不能冒充用户批准;选库、下载、拟合、SEALED 和写稿权限严格分开。待决期间可继续不依赖该节点的合法只读研究,不中断普通技术任务。 From 43733ce34eaaedbe0ba56eb7e2cf17e0d93d55f1 Mon Sep 17 00:00:00 2001 From: 91haoran Date: Mon, 16 Nov 2026 13:25:28 +0900 Subject: [PATCH 20/25] fix(policy): align AGENTS.md with current H6/H8 ownership and decision output --- AGENTS.md | 2 ++ 1 file changed, 2 insertions(+) diff --git a/AGENTS.md b/AGENTS.md index b0a7728..17f7a3d 100644 --- a/AGENTS.md +++ b/AGENTS.md @@ -13,6 +13,8 @@ **H6 可信入口提示**:新项目 H3–H6 一次用户选择的声明检查调用 `review_new_project_h6`,活动03交接使用 `review_activity_handoff(..., new_project=True)`;这两个接口本身并不认证原始聊天消息,未连接宿主可信身份时不能声称不可绕过。Agent 看到真实当前对话的具体 H6 回复后不重复询问,继续已获合法 DEV 范围内的工作。 +**新项目决策模式**:`USER_DEEP_PARTICIPATION` 的实际人工交互是启动配置(已提供的不重复问)与 H6 一次方案选择,不等于 H3/H4/H5 分别触发三次审批。 + **人工科学决策冷启动**:新项目 H1/H2 直接复用已有启动配置、H3–H5 由 Agent 先作未批准的候选分析,只在 H6 提交一张具体科研方案卡供作者一次选择(同时确定 H3–H6);**H6核心Task/研究组合与H8核心Claim/重大升级需先由AI给带依据的首选和可行备选,再由作者选择或明确范围委托**。H7技术方案、既定策略内DEV算法/训练/修复仍由AI自主,不逐次询问。旧项目H6/H8原记录保留但旧AGENT_DECIDED不能伪装为新作者确认,只有实质战略变动时按新版审核。详见 [AUTHOR_STRATEGY_REVIEW.md](AUTHOR_STRATEGY_REVIEW.md)、[WORKFLOW.md](WORKFLOW.md);原H9/H10、SEALED与写稿门禁不变。 **论文蓝图检查闭环**:科研方向或Task明确即自主创建 `PAPER_BLUEPRINT.md`(默认六章、H1/H2、逐节拟写内容和Claim–Evidence);重要Run、关键负结果、核心Claim/路线变化、阶段交接时,Agent在已有 `RESEARCH_RECORD.md` 第5节事件填写 `EVENT_KIND`、`BLUEPRINT_CHECK: UPDATED/NO_CHANGE_NEEDED`、`BLUEPRINT_SECTIONS`、`BLUEPRINT_EVIDENCE` 和 `BLUEPRINT_REASON`,按原证据核受影响SEC_ID。普通调参、重复测试、无实质Debug不触发。不新增用户确认;已批准重大结构变动仍须真实作者确认;`WRITING_LOCKED`/绘图门禁不受影响。细节见 [DUAL_DOCUMENT_WORKFLOW.md](DUAL_DOCUMENT_WORKFLOW.md),`core.doc_integrity` 的静态通过不代表科学正确。 From 52aecd53a54714e5c566fc295a2fbd3dd2adc962 Mon Sep 17 00:00:00 2001 From: 91haoran Date: Mon, 16 Nov 2026 13:25:48 +0900 Subject: [PATCH 21/25] test: align tests/test_case_registry.py with current scientific evidence and H6 semantics --- tests/test_case_registry.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tests/test_case_registry.py b/tests/test_case_registry.py index c13bd5a..a78df12 100644 --- a/tests/test_case_registry.py +++ b/tests/test_case_registry.py @@ -35,7 +35,7 @@ def test_status_axes_do_not_falsely_promote_evidence(self): self.assertIn("来源简称", self.case) else: self.assertIn(required, self.case) - self.assertIn("37 条 SEALED", self.case) + self.assertIn("37 test Episode", self.case) # official test Episodes are not automatically independent SEALED self.assertIn("5 人同数据源确认", self.case) self.assertIn("待", self.case) From f6b8b348f135d0c9ebd115b26514c92aaf95d15e Mon Sep 17 00:00:00 2001 From: 91haoran Date: Mon, 16 Nov 2026 13:25:51 +0900 Subject: [PATCH 22/25] test: align tests/test_human_decision_flow.py with current scientific evidence and H6 semantics --- tests/test_human_decision_flow.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tests/test_human_decision_flow.py b/tests/test_human_decision_flow.py index 7abd10b..6f8142a 100644 --- a/tests/test_human_decision_flow.py +++ b/tests/test_human_decision_flow.py @@ -148,7 +148,7 @@ def test_prominent_entrypoint_and_heading_namespace(self): def test_documents_have_human_choice_at_each_owner(self): rules = { - "WORKFLOW.md": ("USER_DEEP_PARTICIPATION", "H1", "H10", "HUMAN_REVIEW_PENDING"), + "WORKFLOW.md": ("USER_DEEP_PARTICIPATION", "H1", "H10", "AUTHOR_CHOICE_REQUIRED"), "START.md": ("USER_DEEP_PARTICIPATION", "H1–H5", "PREFERRED"), "SEARCH_DISCOVERY.md": ("H3", "H5", "BREADTH_SCAN"), "DATA_DISCOVERY.md": ("H4", "H5", "PRE_DOWNLOAD_SCIENCE_FIT"), From b98622ae4b73b607430c9a5efc1011b8e32d5e22 Mon Sep 17 00:00:00 2001 From: 91haoran Date: Mon, 16 Nov 2026 13:26:00 +0900 Subject: [PATCH 23/25] test: H7 machine declaration must be agent-owned; keep H6/H8 author-owned --- tests/test_research_activity_handoff.py | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) diff --git a/tests/test_research_activity_handoff.py b/tests/test_research_activity_handoff.py index bfecdfa..cd91d27 100644 --- a/tests/test_research_activity_handoff.py +++ b/tests/test_research_activity_handoff.py @@ -192,9 +192,11 @@ def _human_packet(*nodes): "target": nodes[-1], "decisions": { step: { - "kind": "USER_CONFIRMED", - "original_user_message_ref": f"synthetic/original/{step}", - "chosen_option": f"synthetic choice {step}", + "kind": "AGENT_DECIDED" if step == "H7" else "USER_CONFIRMED", + **({"agent_decision_ref": "synthetic/agent/H7", + "h5_scope_ref": "synthetic/original/H5"} if step == "H7" else { + "original_user_message_ref": f"synthetic/original/{step}", + "chosen_option": f"synthetic choice {step}"}), "dimension_decisions": { key: "synthetic scoped decision" for key in H_CONTRACT["steps"][step]["required_dimensions"] From 7a19c56fbeaf5500caac9b91281f296ffa91d1a1 Mon Sep 17 00:00:00 2001 From: 91haoran Date: Mon, 16 Nov 2026 13:27:13 +0900 Subject: [PATCH 24/25] =?UTF-8?q?docs:=20state=20H1=E2=80=93H10=20identiti?= =?UTF-8?q?es=20remain=20stable=20while=20H6=20interaction=20is=20bundled?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- AGENTS.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/AGENTS.md b/AGENTS.md index 17f7a3d..dc4866b 100644 --- a/AGENTS.md +++ b/AGENTS.md @@ -13,7 +13,7 @@ **H6 可信入口提示**:新项目 H3–H6 一次用户选择的声明检查调用 `review_new_project_h6`,活动03交接使用 `review_activity_handoff(..., new_project=True)`;这两个接口本身并不认证原始聊天消息,未连接宿主可信身份时不能声称不可绕过。Agent 看到真实当前对话的具体 H6 回复后不重复询问,继续已获合法 DEV 范围内的工作。 -**新项目决策模式**:`USER_DEEP_PARTICIPATION` 的实际人工交互是启动配置(已提供的不重复问)与 H6 一次方案选择,不等于 H3/H4/H5 分别触发三次审批。 +**新项目决策模式**:H1–H10 保留科研语义,`USER_DEEP_PARTICIPATION` 的实际人工交互是启动配置(已提供的不重复问)与 H6 一次方案选择,不等于 H3/H4/H5 分别触发三次审批。 **人工科学决策冷启动**:新项目 H1/H2 直接复用已有启动配置、H3–H5 由 Agent 先作未批准的候选分析,只在 H6 提交一张具体科研方案卡供作者一次选择(同时确定 H3–H6);**H6核心Task/研究组合与H8核心Claim/重大升级需先由AI给带依据的首选和可行备选,再由作者选择或明确范围委托**。H7技术方案、既定策略内DEV算法/训练/修复仍由AI自主,不逐次询问。旧项目H6/H8原记录保留但旧AGENT_DECIDED不能伪装为新作者确认,只有实质战略变动时按新版审核。详见 [AUTHOR_STRATEGY_REVIEW.md](AUTHOR_STRATEGY_REVIEW.md)、[WORKFLOW.md](WORKFLOW.md);原H9/H10、SEALED与写稿门禁不变。 From 67934a46fd1f7075dea1c31ea3cea00b0b144fae Mon Sep 17 00:00:00 2001 From: 91haoran Date: Mon, 16 Nov 2026 13:27:27 +0900 Subject: [PATCH 25/25] test: preserve v1.2 legacy H6 agent evidence without promoting to v1.3 user consent --- tests/test_dev_autonomy.py | 30 +++++++++++++++++++++++++++--- 1 file changed, 27 insertions(+), 3 deletions(-) diff --git a/tests/test_dev_autonomy.py b/tests/test_dev_autonomy.py index 4e261ee..a237007 100644 --- a/tests/test_dev_autonomy.py +++ b/tests/test_dev_autonomy.py @@ -99,10 +99,34 @@ def test_invalidated_h5_stales_agent_decisions_and_h9(self): def test_legacy_science_decisions_can_be_read_but_not_reclassified(self): p = packet("H6", ("H1", "H2", "H3", "H4", "H5", "H6")) - p["protocol_version"] = "pd-rw-human-review-v1.1" - self.assertFalse(review(p)["can_advance_declarations_only"]) + p["protocol_version"] = "pd-rw-human-review-v1.2" + p["decisions"]["H6"] = { + "kind": "AGENT_DECIDED", + "agent_decision_ref": "synthetic/legacy-agent", + "h5_scope_ref": "synthetic/user-H5", + "dimension_decisions": { + x: "legacy scoped" for x in + ("task_hypothesis", "legal_inputs", "independent_evaluation") + }, + "evidence_refs": { + x: "synthetic/legacy-source" for x in + ("valid_dev_or_theory_basis", "falsifier_and_comparator") + }, + } + legacy = review(p) + self.assertTrue(legacy["can_advance_declarations_only"]) + self.assertEqual(legacy["target_decision_owner_under_source_protocol"], "AGENT_LEGACY") + # Identical old agent decision is NOT v1.3 author approval. + p["protocol_version"] = CONTRACT["protocol_version"] + current = review(p) + self.assertFalse(current["can_advance_declarations_only"]) + self.assertIn("H6", current["malformed_declarations"]) + # A new explicit author declaration meets the structural requirement, + # but still does not authenticate the actual user or authorize execution. p["decisions"]["H6"] = choice("H6", agent=False) - self.assertTrue(review(p)["can_advance_declarations_only"]) + updated = review(p) + self.assertTrue(updated["can_advance_declarations_only"]) + self.assertFalse(updated["allows_download_training_sealed_or_writing"]) def test_activity04_accepts_standing_dev_scope_not_human_h7_signoff(self): from tests.test_research_activity_handoff import packet as activity_packet