Harden feedback-driven continuity capture

ITERATION_LOG.md

@@ -16,6 +16,22 @@
 
 ---
 
+## 2026-08-16：中文 owner 范围收拢与外部 Windows 案例收口
+
+### FLG-ITER-20260816-01：中文 owner 范围收拢在短句与长口述中重复漏提取 [confirmed — P0]
+
+**场景**：2026-08-11，FlightModeAI owner 用长段中文口述把产品收拢为飞行学习包，说明飞行前选主题、飞行中完成、落地后分享的闭环，也指出端侧模型体验与互动方式仍待验证。`closeout` 只触发 lessons，正式候选为 0。2026-08-16，owner 又用一句“我们主要做 Canvas Prompt FlowGrid 其他我觉得意义不大”收拢资源范围，`closeout` 再次返回空候选。
+
+**发现**：两次样本长度不同，暴露的是同一缺陷。中文 owner 通过“收拢、定义为、主要做、其他意义不大”等自然口语表达范围取舍时，现有入口和提取器不能稳定保存来源明确的候选。[confirmed]
+
+**影响**：安全 review gate 阻止了错误自动写入，但 owner 已经明确的范围可能静默丢失。宿主随后只能人工使用 `decision add` 或 confirmed capture 补账，正式决策还可能夹带无证据的模板字段。
+
+**处理方向**：一句带明确 owner 来源的短判断走 source-backed capture 草稿；会议、长讨论和包含范围、闭环、体验证据、待验证问题的口述走 `closeout`。长口述只有在保留 `User:` 归因、范围收拢信号和流程信号时生成 pending candidate；普通项目名列表 abstain。正式决策只写真实存在的字段，继续保留 review 与来源门。
+
+**本地候选验证**：新增 8 月 11 日真实长口述 fixture、8 月 16 日短 capture 回归和无来源实体列表反例。外部 Windows + WorkBuddy 案例另建隐私安全的冷启动输入预检，普通项目文件提供 1/6 个连续性不变量，FlowGrid Context Pack 提供 6/6；该结果只度量输入信息，不证明 Agent 输出更好、用户满意或跨会话增益。
+
+---
+
 ## 2026-08-08：主线优化审计与合并收口
 
 **场景**：项目主编排同时核对映射 runtime、Vault 治理层、遗留开发 worktree、`flg status`、`flg capture list`、`doctor --strict` 和全量测试。

README.md

@@ -11,36 +11,13 @@
 [![CI](https://github.com/dlxeva/FlowGrid/actions/workflows/ci.yml/badge.svg?branch=master)](https://github.com/dlxeva/FlowGrid/actions/workflows/ci.yml)
 ![Python](https://img.shields.io/badge/python-%3E%3D3.10-3776AB?logo=python&logoColor=white)
 [![License](https://img.shields.io/github/license/dlxeva/FlowGrid)](./LICENSE)
-[![AML Academic Textual #8](https://img.shields.io/badge/AML%20Academic%20Textual-%238-7C3AED)](https://agentmemories.ai/leaderboard/academic/textual)
 
 FlowGrid helps business-project knowledge workers turn messy AI work sessions into state-aware, traceable, and resumable project context.
 
 > **Current status:** the codebase reports `v0.3.0` and is in v0.4 core validation. The current focus is entry reliability, rebuildable ledger state, and real-project continuation; v0.4 is not presented as a released version yet.
 
 It is built for long-running work where the deliverable is not just a document, but a defensible judgment chain: why this proposal makes sense, why this direction was chosen, what alternatives were rejected, and when a past judgment should be revised.
 
-## Public Benchmark Result
-
-The independent [FlowGrid AML Retriever](https://github.com/dlxeva/flowgrid-aml-retriever)
-ranked **#8** in the first public Agent Memory Leaderboard Academic Textual
-track, with an overall score of **43.98**, **1.08 points behind the top-ranked
-entry**. [View the public leaderboard →](https://agentmemories.ai/leaderboard/academic/textual)
-
-The competition entry and FlowGrid Core share ideas about evidence provenance,
-temporal state, conflict preservation, and traceable retrieval, but they are
-different systems:
-
-- **AML Retriever** is a benchmark-specific, deterministic Add/Search memory
-  retriever. It stores messages, builds retrieval views, and returns ranked
-  evidence under the competition contract.
-- **FlowGrid Core** is a local project-continuity layer. It preserves reviewed
-  judgments, constraints, rejected paths, current actions, and handoff state for
-  long-running human-AI projects.
-
-The ranking validates the competition retriever under AML's unified evaluation.
-It does not establish FlowGrid Core's overall product quality, user adoption, or
-universal superiority over other memory systems.
-
 ## How It Carries a Project Forward
 
 ![FlowGrid project continuation lifecycle](./docs/assets/flowgrid-lifecycle.svg)
@@ -136,6 +113,11 @@ flg context --mode manifest
 flg handoff
 ```
 
+Hosts should route one short, explicitly attributed owner judgment to a
+source-backed capture draft, for example `flg capture add`. A meeting, long
+discussion, or multi-signal product explanation goes through `flg closeout`.
+An entity list with no attributed trade-off produces no decision candidate.
+
 After that, you will have:
 
 - a local project ledger: `PROJECT.md`, `FRAMING.md`, `DECISIONS.md`, `SNAPSHOT.md`, `PROGRESS.md`
@@ -152,6 +134,22 @@ remains pending.
 
 You can run that flow from Codex, Hermes, OpenClaw, Claude, or any AI agent work product that can read files and run commands.
 
+## Public Benchmark Result
+
+[![AML Academic Textual #8](https://img.shields.io/badge/AML%20Academic%20Textual-%238-7C3AED)](https://agentmemories.ai/leaderboard/academic/textual)
+
+The independent [FlowGrid AML Retriever](https://github.com/dlxeva/flowgrid-aml-retriever)
+ranked **#8** in the first public Agent Memory Leaderboard Academic Textual
+track, with an overall score of **43.98**, **1.08 points behind the top-ranked
+entry**. [View the public leaderboard →](https://agentmemories.ai/leaderboard/academic/textual)
+
+The competition entry and FlowGrid Core share ideas about evidence provenance,
+temporal state, conflict preservation, and traceable retrieval, but they are
+different systems. AML Retriever is a benchmark-specific deterministic
+Add/Search system. FlowGrid Core is the local continuity product described
+above. The ranking does not establish Core product quality, user adoption, or
+universal superiority.
+
 ## Independent Runtime Experiments
 
 FlowGrid's local-first CLI and protocol are the main product. Three independent

README.zh-CN.md

@@ -11,30 +11,13 @@
 [![CI](https://github.com/dlxeva/FlowGrid/actions/workflows/ci.yml/badge.svg?branch=master)](https://github.com/dlxeva/FlowGrid/actions/workflows/ci.yml)
 ![Python](https://img.shields.io/badge/python-%3E%3D3.10-3776AB?logo=python&logoColor=white)
 [![License](https://img.shields.io/github/license/dlxeva/FlowGrid)](./LICENSE)
-[![AML 学术文本榜 #8](https://img.shields.io/badge/AML%20Academic%20Textual-%238-7C3AED)](https://agentmemories.ai/leaderboard/academic/textual)
 
 FlowGrid 帮助业务项目型知识工作者，把混乱的 AI 协作过程转成有状态边界、可追溯、可恢复的项目上下文。
 
 > **当前状态：** 代码包版本仍为 `v0.3.0`，项目正在进行 v0.4 核心验证。当前重点是原始会话稳定入账、账本状态可重建，以及真实项目续接；目前不把 v0.4 宣称为正式发布版本。
 
 它适合长期推进的模糊项目。此类项目的交付物通常不只是一份文档，还包括一条能被解释和质询的判断链：为什么这个方案成立，为什么选择这个方向，哪些备选被放弃，什么情况下旧判断需要修正。
 
-## 公开评测成绩
-
-独立参赛项目 [FlowGrid AML Retriever](https://github.com/dlxeva/flowgrid-aml-retriever)
-在 Agent Memory Leaderboard 首期公开评测的学术方法文本记忆榜中位列
-**第 8 名**，综合分 **43.98**，与榜首相差 **1.08 分**。
-[查看公开榜单 →](https://agentmemories.ai/leaderboard/academic/textual)
-
-参赛 Retriever 与 FlowGrid Core 共享证据来源、时间状态、冲突保留和可追溯检索等思想，
-但两者不是同一个系统：
-
-- **AML Retriever** 是针对统一评测契约构建的确定性 Add/Search 记忆检索器，负责保存消息、构建检索视图并返回排序后的原始证据。
-- **FlowGrid Core** 是本地项目连续性层，负责保存经过审核的判断、约束、被否决路径、当前行动和长期人机项目的交接状态。
-
-这项成绩验证的是参赛 Retriever 在 AML 统一环境中的表现，不等同于 FlowGrid Core
-整体产品能力、用户采用情况，也不能证明它普遍优于其他记忆系统。
-
 ## 一张图看懂如何续接项目
 
 ![FlowGrid 项目续接生命周期](./docs/assets/flowgrid-lifecycle.svg)
@@ -112,6 +95,10 @@ flg context --mode resume
 flg handoff
 ```
 
+宿主应把一句带明确 owner 来源的短判断送入有来源的 capture 草稿，例如
+`flg capture add`。会议、长讨论或包含多个产品信号的口述走 `flg closeout`。
+只有项目名、没有来源归因和取舍关系的列表不生成决策候选。
+
 跑完以后，你会得到：
 
 - 一套本地项目账本：`PROJECT.md`、`FRAMING.md`、`DECISIONS.md`、`SNAPSHOT.md`、`PROGRESS.md`
@@ -124,6 +111,19 @@ flg handoff
 
 这条流程可以运行在 Codex、Claude、OpenClaw、Hermes 或其他 AI agent work 产品里，不要求用户改用新的工作台。
 
+## 公开评测成绩
+
+[![AML 学术文本榜 #8](https://img.shields.io/badge/AML%20Academic%20Textual-%238-7C3AED)](https://agentmemories.ai/leaderboard/academic/textual)
+
+独立参赛项目 [FlowGrid AML Retriever](https://github.com/dlxeva/flowgrid-aml-retriever)
+在 Agent Memory Leaderboard 首期公开评测的学术方法文本记忆榜中位列
+**第 8 名**，综合分 **43.98**，与榜首相差 **1.08 分**。
+[查看公开榜单 →](https://agentmemories.ai/leaderboard/academic/textual)
+
+参赛 Retriever 与 FlowGrid Core 共享证据来源、时间状态、冲突保留和可追溯检索等思想。
+AML Retriever 是面向榜单契约的确定性 Add/Search 系统，FlowGrid Core 是上文描述的
+本地项目连续性产品。这项成绩不构成 Core 产品质量、用户采用或普遍优越性的证明。
+
 ## 独立运行时实验
 
 FlowGrid 的主产品仍是本地优先 CLI 与项目协议。三个独立仓库在受控运行环境或评测契约下验证部分思路：

evals/README.md

@@ -7,6 +7,20 @@ scenario; Context Pack is not assumed to outperform a clean raw history.
 
 This directory contains scenario fixtures for evaluating that claim.
 
+## External cold-start fixture preflight
+
+The privacy-safe Windows + WorkBuddy fixture compares which known continuity
+invariants are available in ordinary project files and a FlowGrid Context Pack:
+
+```bash
+python evals/cold_start_ab.py --output-dir evals/results
+```
+
+The deterministic result is recorded in
+[`results/external-windows-workbuddy-cold-start.md`](results/external-windows-workbuddy-cold-start.md).
+It measures input information only. A fresh-agent output A/B remains required
+before claiming a continuation gain.
+
 Real-project, dogfood, customer-field, and host-acceptance evidence is indexed
 separately in [`case-registry.json`](case-registry.json). The registry records
 both the supported claim and limitation of every case; it must not label

evals/cold-start/external-windows-workbuddy/case.md

@@ -0,0 +1,19 @@
+# External Windows + WorkBuddy cold-start fixture
+
+This fixture is a privacy-safe reconstruction of behavior visible in screenshots
+from one external Windows + WorkBuddy use of FlowGrid. It is not a transcript,
+testimonial, satisfaction claim, or cross-session result.
+
+Observed behavior boundary:
+
+- an external user installed and used FlowGrid on a real project;
+- the agent used project framing, constraints, and decisions to resist a
+  physical-gardening/e-commerce interpretation of a mobile-game CSR project;
+- D-001 through D-003 retained decision rationale and rejected paths while a
+  deck changed across three revisions;
+- DAU, budget, average order value, and contracting entity remained unknown;
+- Windows path handling caused friction;
+- the screenshots came from one agent with its existing chat context.
+
+The fixture tests only whether a cold-start input exposes the known continuity
+invariants. It does not test model quality by itself.

evals/cold-start/external-windows-workbuddy/flg-context-pack.md

@@ -0,0 +1,31 @@
+# FlowGrid Context Pack excerpt
+
+## Current scope
+
+The project is a CSR marketing campaign for a mobile game. It is not a
+physical gardening, e-commerce, or physical-fulfillment project.
+
+## Reviewed decisions
+
+- D-001: Keep the campaign inside the mobile-game CSR scope. Reject physical
+  gardening and physical fulfillment because they change the product and
+  delivery object.
+- D-002: Keep the channel brief aligned with the approved CSR mechanism while
+  the presentation deck changes visual form. A visual revision does not reopen
+  the project mechanism.
+- D-003: Do not invent DAU, budget, average order value, or the contracting
+  entity. Keep each item as an owner-confirmation question until evidence is
+  supplied.
+
+## Rejected paths
+
+- physical gardening
+- e-commerce
+- physical fulfillment
+
+## Unknowns
+
+- DAU
+- budget
+- average order value
+- contracting entity

evals/cold-start/external-windows-workbuddy/invariants.json

@@ -0,0 +1,15 @@
+{
+  "case_id": "external-windows-workbuddy-cold-start-v1",
+  "modes": {
+    "no_flg": "no-flg-project-files.md",
+    "flg": "flg-context-pack.md"
+  },
+  "invariants": [
+    {"id": "scope-mobile-game-csr", "needles": ["mobile game", "CSR"]},
+    {"id": "reject-physical-gardening", "needles": ["reject", "physical gardening"]},
+    {"id": "reject-ecommerce", "needles": ["e-commerce", "rejected paths"]},
+    {"id": "preserve-d001-d003", "needles": ["D-001", "D-002", "D-003"]},
+    {"id": "do-not-invent-owner-facts", "needles": ["Do not invent", "DAU", "budget", "average order value", "contracting entity"]},
+    {"id": "visual-revision-does-not-reopen", "needles": ["visual revision", "does not reopen"]}
+  ]
+}

evals/cold-start/external-windows-workbuddy/no-flg-project-files.md

@@ -0,0 +1,6 @@
+# Ordinary project files
+
+The project is a CSR marketing campaign for a mobile game. A presentation deck
+has gone through three visual revisions. The next deliverable is a channel
+brief. DAU, budget, average order value, and the contracting entity have not
+been supplied in these files.

evals/cold-start/external-windows-workbuddy/task.md

@@ -0,0 +1,6 @@
+# Cold-start task
+
+Review the incoming request to turn the project into a physical gardening and
+e-commerce campaign. Produce a channel brief that preserves confirmed project
+scope, does not revive rejected paths, and lists facts that still require owner
+confirmation.

evals/cold_start_ab.py

@@ -0,0 +1,97 @@
+"""Deterministic preflight for the external-case cold-start A/B fixture."""
+
+from __future__ import annotations
+
+import argparse
+import json
+from datetime import date
+from pathlib import Path
+
+
+ROOT = Path(__file__).resolve().parent
+DEFAULT_CASE = ROOT / "cold-start" / "external-windows-workbuddy"
+
+
+def evaluate(case_dir: Path) -> dict:
+    config = json.loads((case_dir / "invariants.json").read_text(encoding="utf-8"))
+    results = {}
+    for mode, filename in config["modes"].items():
+        content = (case_dir / filename).read_text(encoding="utf-8").lower()
+        checks = []
+        for invariant in config["invariants"]:
+            passed = all(needle.lower() in content for needle in invariant["needles"])
+            checks.append({"id": invariant["id"], "passed": passed})
+        results[mode] = {
+            "passed": sum(item["passed"] for item in checks),
+            "total": len(checks),
+            "checks": checks,
+        }
+    return {
+        "case_id": config["case_id"],
+        "method": "deterministic input-invariant preflight",
+        "results": results,
+        "boundary": (
+            "Measures information available to a cold-start agent. Does not measure "
+            "agent output quality, user satisfaction, or cross-session improvement."
+        ),
+    }
+
+
+def render_markdown(report: dict) -> str:
+    no_flg = report["results"]["no_flg"]
+    flg = report["results"]["flg"]
+    lines = [
+        "# External Windows + WorkBuddy cold-start A/B preflight",
+        "",
+        f"Date: {date.today().isoformat()}",
+        "",
+        "## Result",
+        "",
+        f"- Ordinary project files: {no_flg['passed']}/{no_flg['total']} continuity invariants available",
+        f"- FlowGrid Context Pack: {flg['passed']}/{flg['total']} continuity invariants available",
+        "",
+        "| Invariant | Ordinary files | FlowGrid pack |",
+        "| --- | ---: | ---: |",
+    ]
+    no_flg_checks = {item["id"]: item["passed"] for item in no_flg["checks"]}
+    flg_checks = {item["id"]: item["passed"] for item in flg["checks"]}
+    for invariant_id in no_flg_checks:
+        lines.append(
+            f"| `{invariant_id}` | {'pass' if no_flg_checks[invariant_id] else 'missing'} "
+            f"| {'pass' if flg_checks[invariant_id] else 'missing'} |"
+        )
+    lines.extend(
+        [
+            "",
+            "## Interpretation boundary",
+            "",
+            report["boundary"],
+            "",
+            "The observed screenshots establish real external use and behavior inside one existing chat context. "
+            "A fresh-agent output comparison is still required before claiming a continuation gain.",
+            "",
+        ]
+    )
+    return "\n".join(lines)
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--case-dir", type=Path, default=DEFAULT_CASE)
+    parser.add_argument("--output-dir", type=Path)
+    args = parser.parse_args()
+    report = evaluate(args.case_dir)
+    print(json.dumps(report, ensure_ascii=False, indent=2))
+    if args.output_dir:
+        args.output_dir.mkdir(parents=True, exist_ok=True)
+        (args.output_dir / "external-windows-workbuddy-cold-start.json").write_text(
+            json.dumps(report, ensure_ascii=False, indent=2) + "\n", encoding="utf-8"
+        )
+        (args.output_dir / "external-windows-workbuddy-cold-start.md").write_text(
+            render_markdown(report), encoding="utf-8"
+        )
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

evals/results/external-windows-workbuddy-cold-start.json

@@ -0,0 +1,31 @@
+{
+  "case_id": "external-windows-workbuddy-cold-start-v1",
+  "method": "deterministic input-invariant preflight",
+  "results": {
+    "no_flg": {
+      "passed": 1,
+      "total": 6,
+      "checks": [
+        {"id": "scope-mobile-game-csr", "passed": true},
+        {"id": "reject-physical-gardening", "passed": false},
+        {"id": "reject-ecommerce", "passed": false},
+        {"id": "preserve-d001-d003", "passed": false},
+        {"id": "do-not-invent-owner-facts", "passed": false},
+        {"id": "visual-revision-does-not-reopen", "passed": false}
+      ]
+    },
+    "flg": {
+      "passed": 6,
+      "total": 6,
+      "checks": [
+        {"id": "scope-mobile-game-csr", "passed": true},
+        {"id": "reject-physical-gardening", "passed": true},
+        {"id": "reject-ecommerce", "passed": true},
+        {"id": "preserve-d001-d003", "passed": true},
+        {"id": "do-not-invent-owner-facts", "passed": true},
+        {"id": "visual-revision-does-not-reopen", "passed": true}
+      ]
+    }
+  },
+  "boundary": "Measures information available to a cold-start agent. Does not measure agent output quality, user satisfaction, or cross-session improvement."
+}