Revise clinical questions and consultation rubrics

L2/README.md

@@ -0,0 +1,5 @@
+# L2: guideline application
+
+185 questions, scored by the macro-average of four capability groups. Three fields define each input: `vignette`, `question` and `options`. The reference `answer` and `explanation` must not be included in the model input.
+
+One question is marked `question_type: multiple_choice`: select every applicable option and return the complete set, for example `\boxed{C,D}`. The answer set must match exactly, independently of order. Other questions require one option.

L4/README.md

@@ -0,0 +1,7 @@
+# L4: case-level planning
+
+165 questions from 84 patients, scored by patient-macro accuracy. Combine `vignette`, `question` and `options` for the model input; do not include the answer or explanation.
+
+Two questions are marked `question_type: multiple_choice`. Return the complete applicable option set, for example `\boxed{B,D}`; extra or missing choices receive no credit. Other questions require one option.
+
+Clinical consultation and treatment dates are retained when they determine the applicable guideline version. Dates of birth and direct identifiers are removed; clinically relevant age is preserved.

L5/README.md

@@ -0,0 +1,7 @@
+# L5: longitudinal decisions
+
+305 turns from 100 patients. Use the complete `prompt` as the model input. The answer, clinical explanation and decisive-condition fields are reference annotations, not input. Each prompt already contains its options.
+
+`全程D0` denotes the patient's first admission; `全程Dk` denotes the day relative to that admission. Dates are transformed within each patient without changing resolved intervals. `某年某月` retains the imprecise timing present in the source.
+
+Score turn accuracy and the fraction of patients with every evaluated turn correct. Auxiliary history conditions have their own inputs and should be evaluated as separate paired conditions.

L5/questions.jsonl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9e39e0e7e6e9ceeeed95868c4d019d4e725dc7fb6155020ff2aa9bec9f4b0595
-size 11219547
+oid sha256:15122f6aa970f5f143ff3545475e24d41c841eff89496a7826f15405f3d192d5
+size 11201989

L6/carepath.tar.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6bafd14db5de6a2224f7e4f0b3399050d5efd967f9a0a06fe85e436edefb5195
-size 1881517860
+oid sha256:6d0cbd1076638519bfbda771a28d617a84f82e9a9fcda3f9502935ef81c72b1e
+size 1880876056

O1/README.md

@@ -0,0 +1,7 @@
+# O1: specialist consultation
+
+103 questions from 42 case descriptions, with 1,726 rubric criteria. Generate an answer from `case_input` and `question`; do not show the reference rubric to the answering model.
+
+`rubrics.jsonl` is the flattened form of each question's embedded rubric. A judge evaluates one clinical criterion against the case, question and model answer. For a positive requirement, `criteria_met=true` earns its points. For an error criterion, true means that the specified error is present and the penalty applies. Positive weights sum to 10 per question; major and minor errors deduct 2 and 1 points respectively, with a floor of zero.
+
+The 3 November 2026 revision updates clinical wording, selected case details and rubric criteria. Criterion indices identify entries within this release; they are not clinical matching keys across revisions.

README.md

@@ -66,13 +66,13 @@ nasopharyngeal carcinoma.
 | `L4/` | Case-level planning | 165 questions from 84 patients |
 | `L5/` | Longitudinal revision | 305 turns from 100 patients |
 | `L6/` | CarePath | 100 patients, 217 scored care decisions |
-| `O1/` | Open-ended consultation | 103 questions, 1,692 rubric criteria |
+| `O1/` | Open-ended consultation | 103 questions, 1,726 rubric criteria |
 
 `supplementary/` contains the separate option-format, history and physician
 calibration datasets used in additional analyses.
 
-Question files are JSONL. They retain the original question IDs, prompts, answer
-options, reference answers and task-specific grouping fields. O1 provides its
+Question files are JSONL. Question IDs and task memberships are retained. Each
+file provides the current prompts, options, references and grouping fields. O1 provides its
 criteria separately. For L3, `Path`, `Seq`, `Struct` and `T-Stg` identify the four
 image tasks; `secondary_labels/` provides additional MRI evidence labels.
 
@@ -90,8 +90,9 @@ for installation, generation and scoring commands.
 
 ## Evaluation scope
 
-These are the inputs for the current paper's evaluated snapshot. Proposed L2/L4
-question revisions remain separate until clinical review and evaluation finish.
+Data revision: **3 November 2026**. The revised L2, L4, L5, O1 and Full100 files
+retain the task memberships and include the supplied clinical corrections.
+Report the data revision when presenting evaluation results.
 Six Full100 treatment questions are unscored pending review; their reasons are
 included in the care-path data. They count as neither correct nor incorrect.
 
@@ -104,9 +105,8 @@ context and do not recursively feed earlier model care answers back. Historical
 ## Additional MRI evaluation
 
 A separate [Full100 MRI prompt condition](supplementary/full100_schema/README.md)
-provides the original Full25 JSON output schema for all 100 patients. The MRI
-images, reference labels, pathology inputs and 217 scored care decisions are
-unchanged. This condition is evaluated separately from the paper snapshot.
+provides the original Full25 JSON output schema for all 100 patients. The MRI images and pathology inputs are unchanged. Reference labels follow
+the current Full100 question revision; care questions remain in the main archive. This condition is evaluated separately from the paper snapshot.
 
 The [T300/N400 extension description](L3/extensions/README.md) summarizes the
 new cohorts and input format. These additional image assets are not included in
@@ -115,7 +115,7 @@ are being completed. They are not included in the released counts above.
 
 ## Ethics and use
 
-L4 retains clinical consultation, treatment and decision dates where they establish the applicable guideline edition. Birth-date expressions are removed; existing ages are retained, or an approximate age is stated with its reference year when only a birth date was available. Other longitudinal tasks use patient-relative days for converted complete dates. Answers, grouping and rubric weights are unchanged.
+L4 retains clinical consultation, treatment and decision dates where they establish the applicable guideline edition. Birth-date expressions are removed; existing ages are retained, or an approximate age is stated with its reference year when only a birth date was available. Other longitudinal tasks use patient-relative days for converted complete dates. The current clinical answers and rubric weights are provided in each task file.
 
 The dataset contains authorized, de-identified benchmark derivatives under the
 hospital-defined release scope and IRB B2026-465-01. It does not contain original

supplementary/full100_schema/README.md

@@ -1,15 +1,5 @@
-# Full100 MRI prompt condition
+# Full100 MRI output-schema comparison
 
-`mri_samples.jsonl` contains 100 MRI requests using AJCC/UICC edition 9.
-The 75 expanded-patient prompts now include the same JSON output example used
-for the original 25 patients. The other 25 prompts are unchanged. All patient
-IDs, images, labels and modality-source fields match the existing L6 download.
+This separate condition adds the original Full25 JSON schema to the 75 expanded-patient MRI prompts. The 100 MRI images and AJCC ninth-edition task are unchanged. Reference labels follow the 3 November 2026 Full100 question revision.
 
-Image paths are relative to `L6/full100_inputs_scored217/`, obtained by extracting
-`L6/carepath.tar.gz`. No duplicate image archive is needed. Keep the supplied
-pathology proxy-image notices and the 217-care-node scoring policy.
-
-Use `npcbench-infer --mri-prompts` with this file as described in the
-[code guide](https://github.com/pidneuralode/NPCBench/blob/main/docs/mri_extension.md).
-Care inputs must be rendered from the MRI and pathology predictions belonging
-to this condition. This file supplies inputs, not completed model results.
+Use these prompts only for the named schema comparison. They do not replace the MRI prompts inside the main CarePath archive.