diff --git a/apps/autocal/config/locales/en_US.json b/apps/autocal/config/locales/en_US.json
index ce2e1ec..641de98 100644
--- a/apps/autocal/config/locales/en_US.json
+++ b/apps/autocal/config/locales/en_US.json
@@ -263,6 +263,7 @@
     "frame": "Frame {frame} · source {source_frame} · {basis} {time} · dropped {dropped}"
   },
   "calibration": {
+    "sensor_model": "Product model",
     "sensor_sn": "Sensor SN",
     "output": "Output directory",
     "site_view": "Select site",
@@ -272,13 +273,15 @@
     "levels": "Load levels (N)",
     "hold": "Suggested hold",
     "reference": "Actual reference (N)",
-    "start": "Start recording",
-    "confirm": "Confirm level",
-    "finish": "Finish trial",
+    "start": "Start trial recording",
+    "confirm": "Confirm current level",
+    "finish": "Finish and save",
     "abort": "Abort and keep data",
     "load": "Load",
     "unload": "Unload",
     "next_step": "Next: {branch} to {load:g} N",
+    "before_start": "Start trial recording first",
+    "ready_to_finish": "All levels confirmed; finish and save",
     "no_step": "No step pending",
     "progress": "Progress {done}/{total} · {status}",
     "held": " · Current level held {seconds:.1f} s",
diff --git a/apps/autocal/config/locales/zh_CN.json b/apps/autocal/config/locales/zh_CN.json
index ffa26d7..aa5e01f 100644
--- a/apps/autocal/config/locales/zh_CN.json
+++ b/apps/autocal/config/locales/zh_CN.json
@@ -263,6 +263,7 @@
     "frame": "帧 {frame} · 源帧 {source_frame} · {basis} {time} · 丢帧 {dropped}"
   },
   "calibration": {
+    "sensor_model": "产品型号",
     "sensor_sn": "传感器 SN",
     "output": "保存目录",
     "site_view": "选择测点",
@@ -272,13 +273,15 @@
     "levels": "加载档位 (N)",
     "hold": "建议保持",
     "reference": "实际参考力 (N)",
-    "start": "开始记录",
-    "confirm": "确认到档",
-    "finish": "结束试验",
+    "start": "开始试验并录制",
+    "confirm": "确认当前档位",
+    "finish": "完成并保存",
     "abort": "中止并保留",
     "load": "加载",
     "unload": "卸载",
     "next_step": "下一档：{branch}至 {load:g} N",
+    "before_start": "请先开始试验并录制",
+    "ready_to_finish": "全部档位已确认，请完成并保存",
     "no_step": "无待确认档位",
     "progress": "进度 {done}/{total} · {status}",
     "held": " · 当前档已保持 {seconds:.1f} s",
diff --git a/apps/autocal/docs/calibration-data-format.zh.md b/apps/autocal/docs/calibration-data-format.zh.md
index e438c7d..79ac561 100644
--- a/apps/autocal/docs/calibration-data-format.zh.md
+++ b/apps/autocal/docs/calibration-data-format.zh.md
@@ -1,16 +1,16 @@
 # Autocal 标定操作与数据格式草稿
 
-状态：待评审，2026-10-11。用途：方案及操作指南。阶段一只针对 finger；单测点、单 trial 手动记录已接通，跨点 session、FT 参考及灵敏度可视化尚未实现。下列目标文件清单不等同于当前写盘格式；当前每 trial 是一个独立目录，含 `raw_0001.mp4`、`session.json`、`frames.csv`、`forces.csv`、`events.csv`。
+状态：待评审，2026-10-11。阶段一仅针对 finger；单点手动记录已接通，跨点 session、FT 参考及灵敏度图待实现。每 trial 保存视频、`session.json`、`frames.csv`、`forces.csv`、`reference.csv`、`events.csv`。目录名为 `<UTC时间>_<产品型号>_<SN>_session<采集会话号>_<唯一标识>`；完整型号和 SN 存于 `session.json`。
 
 ## 范围与软件现状
 
 | 项目 | 当前约定 |
 |---|---|
-| 标定范围 | 多测点法向力，暂定 0–30 N；剪切及组合载荷后续再做 |
+| 标定范围 | 多测点法向力，0–30 N；剪切及组合载荷后续再做 |
 | 传感器 | 首版固定为 finger；其余类型须另行验证几何、采集和处理契约 |
 | 压头 | 直径 6 mm、球面半径 3 mm 的半球压头 |
 | 参考 | 砝码或 50 N／2 N·m FT sensor，首版先接通一种 |
-| 图像 | 一个完整测试事件一个连续视频，档位间不切段 |
+| 图像 | 每测试事件一个连续视频，档位间不切段 |
 | 调试输入 | 已有视频与实时相机编号共用 finger worker 和 `SdkPacket`；相机实机尚未验收 |
 | 交付 | 单个 session 能重新打开、追溯参考载荷，计算并显示不同测点的灵敏度 |
 
@@ -24,7 +24,7 @@
 
 | 步骤 | 操作员动作 | 软件记录／通过条件 |
 |---|---|---|
-| 1. 建立会话 | 填写传感器 SN、gel ID、实验台、压头和参考设备；选保存位置 | 创建唯一 `session_id`；保存身份和配置快照 |
+| 1. 建立会话 | 填写产品型号、传感器 SN、gel ID、实验台、压头和参考设备；选保存位置 | 创建唯一 `session_id`；保存身份和配置快照 |
 | 2. 检查设备 | 完成设备设置、回零及既定检查，核对相机和参考输入 | 状态可读；运动边界按既定流程确认 |
 | 3. 确认测点 | 在有效 UV 区域选点，核对三维位置和姿态 | 分配 `site_id`，UV→3D 映射有效 |
 | 4. 设置循环 | 定义加载范围、档位、保持时间和重复数，选择手动模式 | 保存计划；自动加载留到下一阶段 |
@@ -56,16 +56,16 @@
 | `session.json` | `format_version`、`session_id`、`sensor_type`、`input_source`、身份、配置快照、坐标定义、文件清单、sites、trials、状态 | 会话建立、事件完成及结束时 |
 | `sequence.oryseq` | 实际使用的测点计划；遵循现有格式 | 开始采集前保存快照 |
 | `video/<trial_id>.<ext>` | 算法入口图像、完整循环及恢复 | 每 trial 开始至结束 |
-| `frames.csv` | `session_id`、`trial_id`、`frame_id`、`t_host_mono_ns`、`t_host_utc_ns`、`time_basis`、视频路径、帧号、丢帧标记 | 每帧，时间与帧身份同时绑定 |
+| `frames.csv` | 当前按帧身份、采集时间、视频位置、状态排序；包含 `frame_id`、`source_frame_id`、`t_rel_s`、`t_host_mono_ns`、`t_host_utc_ns`、`acquired_at`、`time_basis`、`media_time_ms`、`segment`、`segment_frame`、`in_video`、`dropped`、`capture_dropped` | 每帧，时间与帧身份同时绑定；当前与 Studio 共用录制格式 |
 | `baseline/` | 基准图像、初始状态或预录引用、`baseline_id` | 建立／重置基准时 |
-| `reference.csv` | `trial_id`、`sample_id`、`t_host_mono_ns`、`clock_id`、力、有效性、来源及延迟估计 | FT／ECAT 每个样本；砝码用事件和有效时段 |
+| `reference.csv` | 当前为 `step_index`、`branch`、`target_load_n`、区间起止电脑单调时间、`F_ref_normal_N`、预留的 `F_ref_x_N`／`F_ref_y_N`／`F_ref_z_N`、`force_frame`、`source`、`stability_verified` | 当前每个手工确认档位一行；FT／ECAT 逐样本参考尚未接通 |
 | `motion.csv` | 时间、目标／反馈位姿、有效性 | 有实际反馈时保存；缺少反馈须注明 |
 | `events.csv` | `t_host_mono_ns`、trial／step、事件类型、参考有效时段、原因 | 到档确认、阶段切换、归零、异常及中断 |
 | `derived/<processing_id>/` | 逐点灵敏度表、加载／卸载曲线、误差、分析窗口和处理配置 | 分析时另存；不覆盖原始文件 |
 
 | 参数归属 | 主要参数 | 单位／填写来源 |
 |---|---|---|
-| session 身份 | `sensor_sn`、`gel_id`、`bench_id`、`indenter_id`、`reference_source` | 操作员填写或设备读取 |
+| session 身份 | `sensor_model`、`sensor_sn`、`gel_id`、`bench_id`、`indenter_id`、`reference_source` | 操作员填写或设备读取；当前型号和 SN 必填 |
 | session 采集 | 分辨率、像素格式、编码、版本、滤波与归零配置 | 软件快照；编码待实测确认 |
 | 输入来源 | `input_source.kind` 为 `video` 或 `camera`；`value` 分别为视频路径或相机整数编号；视频另存文件哈希 | 传给同一 `FingerAcquisitionWorker.start`；来源文件与 trial 输出视频不可混淆 |
 | session 时间 | 同进程 `clock_id`、单调时钟原点、UTC 对照、时间来源、延迟及同步误差 | 电脑采集／接收时记录；不能按行号配对 |
@@ -77,11 +77,13 @@
 | 质量 | `valid`、`invalid_reason`、丢帧／中断状态 | 未知不补零；算法失败不等于图像不可用 |
 | 逐点结果 | `site_id`、`gain_n_per_raw`、`sensitivity_raw_per_n`、有效档位数、误差及滞后 | 仅对有效独立参考与原始法向信号拟合；缺数据标无效 |
 
-实时采集以电脑收到**该帧**时的单调时间为主时间。采集线程需把 `t_host_mono_ns` 与帧序号、图像原子配对；记录器再保存 UTC 对照，处理结束时间不能代替采集时间。FT／ECAT 样本也在电脑收到时打同源时间戳，按时间关联，并记录可测得的延迟与失同步；低延迟不等于零延迟。砝码的人工到档时刻只定义有效区间，稳定窗口仍需检查。
+实时采集以电脑收到**该帧**时的单调时间为主时间。采集线程需把 `t_host_mono_ns` 与帧序号、图像原子配对；记录器再保存 UTC 对照，处理结束时间不能代替采集时间。FT／ECAT 样本也在电脑收到时打同源时间戳，按时间关联，并记录可测得的延迟与失同步；低延迟不等于零延迟。砝码的人工到档时刻只定义候选区间，稳定窗口仍需检查。
 
-旧视频没有原始逐帧电脑时间时，重放取得的是本次解码时间，`time_basis=host_replay`；它可验证处理链，不能还原原实验与 FT 的同步。若视频附有原始帧时间则保留 `time_basis=original_recorded` 和原时钟身份。finger worker 已停止按配置帧率推算 `mono_t`，并在包中保留主机采集时间、源帧编号及可用的媒体时间；正式记录器仍须将这些字段写入 `frames.csv`。manifest 保留来源路径、哈希、配置、媒体帧率及各时间基准。
+当前 `reference.csv` 的起点是操作员确认时刻，终点是下一次确认或结束时刻。`F_ref_normal_N` 是人工声明的正向压缩载荷；未测量的 `F_ref_x_N`、`F_ref_y_N`、`F_ref_z_N` 及对应 `force_frame` 留空，不写零，也不将法向值直接当作平台 Z 分量。`stability_verified=0` 表示整个声明区间尚未验证稳定，后续分析必须另选稳定窗口。`forces.csv` 只保存 finger 图像处理输出，不含独立参考力。
 
-事件边界如下。`events.csv` 每行至少保存 `event_id`、`trial_id`、可为空的 `step_id`、`event_type`、`t_host_mono_ns`、`clock_id` 和 `reason`。操作员的确认时间和设备反馈时间分别标明来源；不得把按钮时刻冒充设备实际到档。所有流用 `trial_id` 与同源单调时间关联。
+旧视频没有原始逐帧电脑时间时，重放取得的是本次解码时间，`time_basis=host_replay`；它可验证处理链，不能还原原实验与 FT 的同步。finger worker 和录制器已保存主机采集时间、源帧编号及可用的媒体时间；来源路径、哈希、配置和媒体帧率仍需在 manifest 中核对。
+
+目标事件需标明 trial、step、事件时刻及来源；按钮时刻不能冒充设备实际到档。当前 `events.csv` 的 `detail` 保存 `t_host_mono_ns` 和档位信息。
 
 | 事件 | 触发与状态 |
 |---|---|
diff --git a/apps/autocal/docs/implementation-plan.zh.md b/apps/autocal/docs/implementation-plan.zh.md
index 75a6d72..efa5124 100644
--- a/apps/autocal/docs/implementation-plan.zh.md
+++ b/apps/autocal/docs/implementation-plan.zh.md
@@ -33,9 +33,9 @@
 
 M1 的视频与相机输入现共用带时间的数据包；M2 记录器与 M3 手动 UI 可在接口固定后并行。M4 可先用合成 session 开发，但阶段一只有在真实记录上重新打开并计算逐测点结果才算完成。Studio 的 `SessionController`、widget 和 `RecordingCoordinator` 保留在 Studio；Autocal 直接复用 shared `FingerAcquisitionWorker`、`SdkPacket` 和 sink，不跨应用导入私有模块。只有出现两端实际共用的逻辑时才移到 shared；SDK 不依赖 UI。
 
-界面复核：原先测点、计划和输入分散在“手指调试”“测试序列”“传感器”页。现在“标定采集”页串起单次 trial；全局“运行”按钮尚未负责标定，页面内有独立的“开始记录”。
+界面复核：原先测点、计划和输入分散在“手指调试”“测试序列”“传感器”页。现在“标定采集”页串起单次 trial；全局“运行”按钮尚未负责标定，页面内有独立的“开始试验并录制”。
 
-已实现的最小闭环在“标定采集”页：点击 UV 图粗选、以 U/V 整数像素精确修改、切换到实时二维 UV 图像，逐档填写实际参考力并确认。每个 trial 单独使用 shared `ExperimentRecorder` 生成目录、视频、帧索引、力数据和事件；`frames.csv` 附加主机采集时间及来源帧编号。当前人工参考只记录到档事件，尚无可自动判定的稳定窗口；未实现 FT 同步、跨点 session 汇总、灵敏度拟合和实体相机验收。手动 Z 控制仍使用既有设备界面，专用弹窗需先确定操作边界。
+已实现的最小闭环在“标定采集”页：点击 UV 图粗选、以 U/V 整数像素精确修改、切换到实时二维 UV 图像，逐档填写实际参考力并确认。每个 trial 单独使用 shared `ExperimentRecorder` 生成目录、视频、帧索引、力数据和事件；人工参考另存 `reference.csv`，仅声明档位区间，尚无可自动判定的稳定窗口。未实现 FT 同步、跨点 session 汇总、灵敏度拟合和实体相机验收。手动 Z 控制仍使用既有设备界面，专用弹窗需先确定操作边界。
 
 主代理负责接口、M3 和集成；一名子代理负责 M1/M2，另一名负责 M4 与合成会话验证。建议总计主代理加两名子代理，约 4–6 个工程工作日并留设备联调时间；单人约 6–9 日。估算不包括设备、FT 协议或硬件等待。当前计划不启动子代理。
 
diff --git a/apps/autocal/src/orisys_autocal/trial.py b/apps/autocal/src/orisys_autocal/trial.py
index 853e588..a6b4716 100644
--- a/apps/autocal/src/orisys_autocal/trial.py
+++ b/apps/autocal/src/orisys_autocal/trial.py
@@ -6,8 +6,10 @@ module owns only the site, planned steps and operator-confirmed event order.
 
 from __future__ import annotations
 
+import csv
 import hashlib
 import json
+import re
 import time
 from dataclasses import dataclass
 from pathlib import Path
@@ -16,6 +18,26 @@ import numpy as np
 from orisys_app_shared.record import ExperimentRecorder, RecordConfig
 
 
+REFERENCE_HEADER = (
+    "step_index",
+    "branch",
+    "target_load_n",
+    "t_start_host_mono_ns",
+    "t_end_host_mono_ns",
+    "F_ref_normal_N",
+    "F_ref_x_N",
+    "F_ref_y_N",
+    "F_ref_z_N",
+    "force_frame",
+    "source",
+    "stability_verified",
+)
+
+
+def _folder_part(value: str) -> str:
+    return re.sub(r"[^\w-]+", "-", value.strip()).strip("-_ ")[:48]
+
+
 @dataclass(frozen=True)
 class Site:
     u_px: int
@@ -84,6 +106,10 @@ class ManualTrial:
         self.aborted = False
         self.reference_alignment_possible = False
         self._last_confirm_ns: int | None = None
+        self._reference_file = None
+        self._reference_writer = None
+        self._active_reference: tuple[int, str, float, float, int] | None = None
+        self._reference_error = ""
 
     def start(
         self,
@@ -92,6 +118,7 @@ class ManualTrial:
         site: Site,
         levels: list[float],
         hold_s: float,
+        sensor_model: str,
         sensor_sn: str,
         bundle_path: str,
         source: dict,
@@ -104,6 +131,15 @@ class ManualTrial:
             raise RuntimeError("start the finger input and wait for its first frame")
         if not sensor_sn.strip():
             raise ValueError("sensor SN is required")
+        if not sensor_model.strip():
+            raise ValueError("sensor model is required")
+        model_part = _folder_part(sensor_model)
+        sn_part = _folder_part(sensor_sn)
+        if not model_part or not sn_part:
+            raise ValueError(
+                "sensor model and SN need letters or numbers for the folder"
+            )
+        label = f"{model_part}_{sn_part}"
         if not float(fps) > 0.0:
             raise ValueError("input frame rate must be positive")
         self.steps = load_path(levels)
@@ -122,6 +158,7 @@ class ManualTrial:
         metadata = {
             "purpose": "autocal_finger_manual_trial",
             "sensor_sn": sensor_sn.strip(),
+            "sensor_model": sensor_model.strip(),
             "site": {
                 "uv_xy_px": [site.u_px, site.v_px],
                 "position_sensor_mm": list(site.position_sensor_mm),
@@ -138,15 +175,36 @@ class ManualTrial:
             "time_basis": packet.time_basis,
             "reference_alignment_possible": self.reference_alignment_possible,
             "reference_source": "manual_operator_confirmation",
+            "reference_file": "reference.csv",
         }
         recorder = ExperimentRecorder(
             RecordConfig(fps=float(fps), segment_seconds=24 * 3600.0)
         )
         directory = recorder.start(
-            root=Path(root), session_id=packet.session_id, metadata=metadata
+            root=Path(root),
+            session_id=packet.session_id,
+            metadata=metadata,
+            directory_label=label,
         )
+        reference_file = None
+        try:
+            reference_file = (directory / "reference.csv").open(
+                "w", newline="", encoding="utf-8"
+            )
+            reference_writer = csv.writer(reference_file)
+            reference_writer.writerow(REFERENCE_HEADER)
+            reference_file.flush()
+        except OSError:
+            if reference_file is not None:
+                reference_file.close()
+            recorder.stop(wait=True)
+            raise
         self.recorder = recorder
         self.directory = directory
+        self._reference_file = reference_file
+        self._reference_writer = reference_writer
+        self._active_reference = None
+        self._reference_error = ""
         self.step_index = 0
         self.aborted = False
         self._last_confirm_ns = None
@@ -161,11 +219,11 @@ class ManualTrial:
             self.steps[self.step_index] if self.step_index < len(self.steps) else None
         )
 
-    def _event(self, kind: str, detail: dict) -> None:
+    def _event(self, kind: str, detail: dict, *, now_ns: int | None = None) -> None:
         recorder = self.recorder
         if recorder is None:
             return
-        now_mono = time.perf_counter_ns()
+        now_mono = now_ns if now_ns is not None else time.perf_counter_ns()
         packet = self.worker.latest()
         if packet is not None and packet.capture_mono_ns is not None:
             event_mono_t = (
@@ -185,6 +243,33 @@ class ManualTrial:
             frame_id=frame_id,
         )
 
+    def _end_reference(self, end_ns: int) -> None:
+        if self._active_reference is None:
+            return
+        step_index, branch, target, reference, start_ns = self._active_reference
+        try:
+            self._reference_writer.writerow(
+                (
+                    step_index,
+                    branch,
+                    target,
+                    start_ns,
+                    end_ns,
+                    reference,
+                    "",
+                    "",
+                    "",
+                    "",
+                    "manual_operator_confirmation",
+                    0,
+                )
+            )
+            self._reference_file.flush()
+        except OSError as exc:
+            self._reference_error = str(exc)
+        finally:
+            self._active_reference = None
+
     def confirm(self, reference_n: float) -> tuple[str, float]:
         if self.status != "recording" or self.next_step is None:
             raise RuntimeError("no load step is waiting for confirmation")
@@ -194,12 +279,20 @@ class ManualTrial:
             raise ValueError("reference load must be between 0 and 30 N")
         now_ns = time.perf_counter_ns()
         if self._last_confirm_ns is not None:
+            self._end_reference(now_ns)
+            if self._reference_error:
+                self.finish(
+                    aborted=True,
+                    reason=f"reference write failed: {self._reference_error}",
+                )
+                raise OSError(self._reference_error)
             self._event(
                 "step_ended",
                 {
                     "step_index": self.step_index - 1,
                     "held_s": (now_ns - self._last_confirm_ns) / 1e9,
                 },
+                now_ns=now_ns,
             )
         self._event(
             "step_reached",
@@ -210,7 +303,9 @@ class ManualTrial:
                 "reference_load_n": reference,
                 "reference_kind": "manual",
             },
+            now_ns=now_ns,
         )
+        self._active_reference = (self.step_index, branch, target, reference, now_ns)
         self.step_index += 1
         self._last_confirm_ns = now_ns
         return branch, target
@@ -226,13 +321,28 @@ class ManualTrial:
             raise RuntimeError("no active trial")
         if not aborted and self.next_step is not None:
             raise RuntimeError("complete the load/unload path before finishing")
+        now_ns = time.perf_counter_ns()
         if self._last_confirm_ns is not None:
+            self._end_reference(now_ns)
             self._event(
                 "step_ended",
-                {"step_index": self.step_index - 1, "held_s": self.held_s},
+                {
+                    "step_index": self.step_index - 1,
+                    "held_s": (now_ns - self._last_confirm_ns) / 1e9,
+                },
+                now_ns=now_ns,
             )
+        try:
+            self._reference_file.close()
+        except OSError as exc:
+            self._reference_error = str(exc)
+        if self._reference_error:
+            aborted = True
+            reason = f"reference write failed: {self._reference_error}"
         self._event(
-            "trial_aborted" if aborted else "trial_finished", {"reason": reason}
+            "trial_aborted" if aborted else "trial_finished",
+            {"reason": reason},
+            now_ns=now_ns,
         )
         self.aborted = aborted
         self.worker.remove_sink(self.recorder.on_packet)
diff --git a/apps/autocal/src/orisys_autocal/ui/calibration_panel.py b/apps/autocal/src/orisys_autocal/ui/calibration_panel.py
index 65f2b8c..989d0c9 100644
--- a/apps/autocal/src/orisys_autocal/ui/calibration_panel.py
+++ b/apps/autocal/src/orisys_autocal/ui/calibration_panel.py
@@ -43,6 +43,9 @@ class CalibrationPanel(QWidget):
 
         outer = QVBoxLayout(self)
         header = QHBoxLayout()
+        self.model_label = QLabel(self)
+        self.sensor_model = QLineEdit(self)
+        self.sensor_model.setObjectName("calibrationSensorModel")
         self.sn_label = QLabel(self)
         self.sensor_sn = QLineEdit(self)
         self.sensor_sn.setObjectName("calibrationSensorSn")
@@ -50,6 +53,8 @@ class CalibrationPanel(QWidget):
         self.output_path = QLineEdit(str(Path(outputs_dir()) / "calibrations"), self)
         self.output_path.setObjectName("calibrationOutputPath")
         self.output_browse = QPushButton(self)
+        header.addWidget(self.model_label)
+        header.addWidget(self.sensor_model)
         header.addWidget(self.sn_label)
         header.addWidget(self.sensor_sn)
         header.addWidget(self.output_label)
@@ -214,6 +219,7 @@ class CalibrationPanel(QWidget):
                 site=self._site,
                 levels=levels,
                 hold_s=self.hold.value(),
+                sensor_model=self.sensor_model.text(),
                 sensor_sn=self.sensor_sn.text(),
                 bundle_path=self._bundle_path,
                 source=source,
@@ -239,21 +245,21 @@ class CalibrationPanel(QWidget):
                 raise ValueError(t("calibration.reference_required"))
             self.trial.confirm(float(self.reference.text()))
             self.reference.clear()
-        except (RuntimeError, ValueError) as exc:
+        except (OSError, RuntimeError, ValueError) as exc:
             self.result_label.setText(t("calibration.error").format(detail=exc))
         self._sync()
 
     def _finish(self) -> None:
         try:
             self.trial.finish()
-        except RuntimeError as exc:
+        except (OSError, RuntimeError) as exc:
             self.result_label.setText(t("calibration.error").format(detail=exc))
         self._sync()
 
     def _abort(self) -> None:
         try:
             self.trial.finish(aborted=True, reason="operator aborted")
-        except RuntimeError as exc:
+        except (OSError, RuntimeError) as exc:
             self.result_label.setText(t("calibration.error").format(detail=exc))
         self._sync()
 
@@ -289,6 +295,7 @@ class CalibrationPanel(QWidget):
             self.levels,
             self.hold,
             self.sensor_sn,
+            self.sensor_model,
             self.output_path,
             self.output_browse,
             self.bundle_button,
@@ -299,13 +306,17 @@ class CalibrationPanel(QWidget):
         self.finish_button.setEnabled(recording and self.trial.next_step is None)
         self.abort_button.setEnabled(recording)
         step = self.trial.next_step
-        self.target_label.setText(
-            t("calibration.next_step").format(
+        if recording and step is not None:
+            target_text = t("calibration.next_step").format(
                 branch=t(f"calibration.{step[0]}"), load=step[1]
             )
-            if recording and step is not None
-            else t("calibration.no_step")
-        )
+        elif recording:
+            target_text = t("calibration.ready_to_finish")
+        elif self.trial.status == "idle":
+            target_text = t("calibration.before_start")
+        else:
+            target_text = t("calibration.no_step")
+        self.target_label.setText(target_text)
         self.progress_label.setText(
             t("calibration.progress").format(
                 done=self.trial.step_index,
@@ -333,6 +344,7 @@ class CalibrationPanel(QWidget):
         )
 
     def retranslate(self) -> None:
+        self.model_label.setText(t("calibration.sensor_model"))
         self.sn_label.setText(t("calibration.sensor_sn"))
         self.output_label.setText(t("calibration.output"))
         self.output_browse.setText(t("source.browse"))
diff --git a/apps/autocal/tests/test_manual_trial.py b/apps/autocal/tests/test_manual_trial.py
index 85eaa6a..0978918 100644
--- a/apps/autocal/tests/test_manual_trial.py
+++ b/apps/autocal/tests/test_manual_trial.py
@@ -10,7 +10,13 @@ from types import SimpleNamespace
 import numpy as np
 import pytest
 from orisys_app_shared.acquisition.models import SdkPacket
-from orisys_autocal.trial import ManualTrial, load_path, site_at, uv_origin
+from orisys_autocal.trial import (
+    REFERENCE_HEADER,
+    ManualTrial,
+    load_path,
+    site_at,
+    uv_origin,
+)
 
 
 class Worker:
@@ -95,6 +101,7 @@ def test_manual_trial_records_video_stamps_and_load_unload_events(tmp_dir):
         site=site,
         levels=[5, 10],
         hold_s=20,
+        sensor_model="Finger D40",
         sensor_sn="F-001",
         bundle_path=str(bundle),
         source={"kind": "camera", "value": 0},
@@ -108,6 +115,7 @@ def test_manual_trial_records_video_stamps_and_load_unload_events(tmp_dir):
     report = trial.recorder.stop(wait=True)
     assert report.ok
     assert trial.poll() == "complete"
+    assert "Finger-D40_F-001" in directory.name
     assert len(list(directory.glob("raw_*.mp4"))) == 1
     with (directory / "frames.csv").open(encoding="utf-8", newline="") as handle:
         row = next(csv.DictReader(handle))
@@ -115,6 +123,30 @@ def test_manual_trial_records_video_stamps_and_load_unload_events(tmp_dir):
     assert row["source_frame_id"] == "7"
     assert row["time_basis"] == "host_camera"
     assert row["capture_dropped"] == "0"
+    with (directory / "reference.csv").open(encoding="utf-8", newline="") as handle:
+        reader = csv.DictReader(handle)
+        assert tuple(reader.fieldnames) == REFERENCE_HEADER
+        references = list(reader)
+    assert [row["branch"] for row in references] == ["load", "load", "unload", "unload"]
+    assert [row["F_ref_normal_N"] for row in references] == [
+        "5.0",
+        "10.0",
+        "5.0",
+        "0.0",
+    ]
+    assert all(
+        row["F_ref_x_N"] == row["F_ref_y_N"] == row["F_ref_z_N"] == ""
+        for row in references
+    )
+    assert all(row["stability_verified"] == "0" for row in references)
+    assert all(
+        int(row["t_start_host_mono_ns"]) < int(row["t_end_host_mono_ns"])
+        for row in references
+    )
+    assert all(
+        previous["t_end_host_mono_ns"] == following["t_start_host_mono_ns"]
+        for previous, following in zip(references, references[1:])
+    )
     with (directory / "events.csv").open(encoding="utf-8", newline="") as handle:
         events = list(csv.DictReader(handle))
     reached = [event for event in events if event["kind"] == "step_reached"]
@@ -128,12 +160,15 @@ def test_manual_trial_records_video_stamps_and_load_unload_events(tmp_dir):
     with (directory / "session.json").open(encoding="utf-8") as handle:
         session = json.load(handle)
     assert session["site"]["uv_xy_px"] == [0, 1]
+    assert session["sensor_model"] == "Finger D40"
+    assert session["reference_file"] == "reference.csv"
     assert session["site"]["mesh_sha256"]
 
 
 def test_calibration_page_keeps_precise_uv_site_when_switching_to_live_view(
     make_window,
 ):
+    from orisys_app_shared.i18n import t
     from orisys_autocal.sensor.uv_flat import FlatSurface
 
     window = make_window()
@@ -151,6 +186,8 @@ def test_calibration_page_keeps_precise_uv_site_when_switching_to_live_view(
         uv_max=(1, 1),
     )
     panel = window.calibration_panel
+    assert panel.start_button.text() == t("calibration.start")
+    assert panel.target_label.text() == t("calibration.before_start")
     panel.set_flat(flat, "bundle")
     panel._picked(2, 1)
     assert (panel.u_spin.value(), panel.v_spin.value()) == (-1, 2)
diff --git a/shared/src/orisys_app_shared/record/experiment.py b/shared/src/orisys_app_shared/record/experiment.py
index 8f3b3d3..8729b5e 100644
--- a/shared/src/orisys_app_shared/record/experiment.py
+++ b/shared/src/orisys_app_shared/record/experiment.py
@@ -46,7 +46,7 @@ from orisys_app_shared.record.video_writer import VideoSegments
 
 logger = logging.getLogger(__name__)
 
-RECORD_FORMAT_VERSION = "1.3"
+RECORD_FORMAT_VERSION = "1.4"
 
 # Lifecycle states. ``failed`` and ``stopped`` are terminal; a recorder that has
 # not reached a terminal state must not be reused.
@@ -80,17 +80,17 @@ FORCES_HEADER = (
 )
 FRAMES_HEADER = (
     "frame_id",
+    "source_frame_id",
     "t_rel_s",
+    "t_host_mono_ns",
+    "t_host_utc_ns",
     "acquired_at",
+    "time_basis",
+    "media_time_ms",
     "segment",
     "segment_frame",
     "in_video",
     "dropped",
-    "t_host_mono_ns",
-    "t_host_utc_ns",
-    "source_frame_id",
-    "time_basis",
-    "media_time_ms",
     "capture_dropped",
 )
 #: ``frame_id`` was added so an operator action can be tied to a frame position.
@@ -345,8 +345,14 @@ class ExperimentRecorder:
         calibration_id: str = "",
         calibration_version: str = "",
         preroll=(),
+        directory_label: str = "",
     ) -> Path:
         """Create the experiment directory and start the writer thread."""
+        if directory_label and (
+            directory_label in {".", ".."}
+            or not all(char.isalnum() or char in "-_" for char in directory_label)
+        ):
+            raise ValueError("directory label must contain only letters, numbers, - or _")
         # A previous writer that has not reached a terminal state still owns the
         # files and the statistics; reusing it would silently mix two sessions.
         if self._thread is not None and self._thread.is_alive():
@@ -363,7 +369,8 @@ class ExperimentRecorder:
             # lifecycle checks, so caller mutations cannot change this run.
             self.config = replace(config if config is not None else self.config)
         stamp = datetime.now(timezone.utc).strftime("%Y%m%dT%H%M%SZ")
-        directory = Path(root) / f"{stamp}_session{session_id:04d}_{uuid4().hex}"
+        label = f"_{directory_label}" if directory_label else ""
+        directory = Path(root) / f"{stamp}{label}_session{session_id:04d}_{uuid4().hex}"
         try:
             directory.mkdir(parents=True, exist_ok=False)
         except Exception as exc:  # noqa: BLE001 - reported to the caller
@@ -1141,17 +1148,17 @@ class _SessionWriter:
         self._frames[1].writerow(
             (
                 sample.frame_id,
+                sample.source_frame_id if sample.source_frame_id is not None else "",
                 repr(t_rel),
+                sample.capture_mono_ns or "",
+                sample.capture_utc_ns or "",
                 acquired,
+                sample.time_basis,
+                format_value(sample.media_time_ms),
                 segment,
                 segment_frame,
                 1 if segment else 0,
                 0,
-                sample.capture_mono_ns or "",
-                sample.capture_utc_ns or "",
-                sample.source_frame_id if sample.source_frame_id is not None else "",
-                sample.time_basis,
-                format_value(sample.media_time_ms),
                 sample.capture_dropped,
             )
         )
@@ -1170,7 +1177,7 @@ class _SessionWriter:
             wall_t or time.time(), timezone.utc
         ).isoformat(timespec="milliseconds")
         self._frames[1].writerow(
-            (frame_id, repr(t_rel), acquired, 0, 0, 0, 1, "", "", "", "", "", "")
+            (frame_id, "", repr(t_rel), "", "", acquired, "", "", 0, 0, 0, 1, "")
         )
         self.gap_rows += 1
 
diff --git a/shared/tests/test_experiment_recorder.py b/shared/tests/test_experiment_recorder.py
index d4db1c4..ac78d85 100644
--- a/shared/tests/test_experiment_recorder.py
+++ b/shared/tests/test_experiment_recorder.py
@@ -92,6 +92,19 @@ def test_directory_layout_and_headers(root):
             assert tuple(next(csv.reader(handle))) == header, name
 
 
+def test_frame_columns_are_grouped_by_identity_time_video_and_status(root):
+    directory, report = _record(root, [3], config=RecordConfig(video=False))
+    assert report.ok
+    with (directory / "frames.csv").open(encoding="utf-8", newline="") as handle:
+        reader = csv.reader(handle)
+        assert tuple(next(reader)) == (
+            "frame_id", "source_frame_id", "t_rel_s", "t_host_mono_ns",
+            "t_host_utc_ns", "acquired_at", "time_basis", "media_time_ms",
+            "segment", "segment_frame", "in_video", "dropped", "capture_dropped",
+        )
+        assert len(next(reader)) == len(FRAMES_HEADER)
+
+
 def test_unavailable_values_are_empty_not_zero(root):
     recorder = ExperimentRecorder(RecordConfig(video=False))
     directory = recorder.start(root=root, session_id=1, metadata={}, calibration_id="cal-1")
@@ -180,7 +193,7 @@ def test_queue_overflow_is_counted_and_alarmed(root):
     # A queue that holds one item, with a writer we deliberately stall.
     recorder = ExperimentRecorder(RecordConfig(video=True, max_queue=16))
     recorder.set_alarm_handler(alarms.append)
-    directory = recorder.start(root=root, session_id=1, metadata={})
+    recorder.start(root=root, session_id=1, metadata={})
 
     # Push far more than the queue can hold without letting the writer catch up.
     for index in range(4000):