feat: example agent run for dataset idx 10

.runs/holdout-10/AGENT_INSTRUCTIONS.md

@@ -0,0 +1,10 @@
+# Run environment — leave-one-cluster-out
+
+Harmonize the held-out dataset(s) below using ONLY:
+- the skills in `skills/`,
+- the exemplars in `data/processed/.../sm_data_harmonization_mapping.json` and the code patterns in `data/gold/expert_code/harmonize_sm/` (both have the held-out cluster removed),
+- the shared raw inputs under `~/ess-dive_wfsfa_soil_datasets/` and the cached metadata under `data/external/ess-dive_meta/`.
+
+Do NOT look up the held-out dataset's existing harmonized output, expert code, or mapping entry from any other location. The held-out datasets are:
+
+- `ess-dive-01092fc392bc46d-20240819T143818677`

.runs/holdout-10/MANIFEST.json

@@ -0,0 +1,12 @@
+{
+  "name": "holdout-10",
+  "holdout_indices": [10],
+  "holdout_identifiers": ["ess-dive-01092fc392bc46d-20240819T143818677"],
+  "exemplar_indices": [1, 2, 3, 4, 5, 6, 7, 8, 9, 15, 16, 17, 18, 23, 24, 25, 26, 27],
+  "n_exemplars": 18,
+  "sources": {
+    "package": "data/gold/expert_code/harmonize_sm",
+    "mapping": "data/gold/sm_data_harmonization_mapping.json",
+    "skills": "skills"
+  }
+}

results/raw_runs/fold_08_holdout_10/run_0/curator/agent_trace_skill1.txt

@@ -0,0 +1,227 @@
+=== SKILL 1 (CURATOR) AGENT TRACE ===
+Model: claude-sonnet-4-6
+Skill version: 0.1
+Timestamp: 2026-06-30T22:55:00Z
+Target: doi:10.15485/2322567
+Fold: 8 (holdout idx=10, exemplar pool: [1,2,3,4,5,6,7,8,9,15,16,17,18,23,24,25,26,27])
+
+--- STEP 1: INPUT HANDLING ---
+Input: doi:10.15485/2322567
+Resolved package ID: ess-dive-01092fc392bc46d-20240819T143818677
+
+--- STEP 2: METADATA RETRIEVAL ---
+Priority 1: Check local cache at data/external/ess-dive_meta/
+  -> No cached metadata found for ess-dive-01092fc392bc46d-20240819T143818677
+
+Priority 2: Fetch from ESS-DIVE API
+  Tool call: GET https://api.ess-dive.lbl.gov/packages/ess-dive-01092fc392bc46d-20240819T143818677
+  Status: 200 OK
+
+Package metadata extracted:
+  Title: "Field data supporting 'Hydrological control of rock carbon fluxes from shale weathering'"
+  DOI: doi:10.15485/2322567
+  Authors: Wan, Justine; Arora, Bhavna; Williams, Kenneth H.; Molins, Sergi; et al.
+  Published: 2024-08-19
+  Temporal coverage: 2016-11-01 to 2021-10-31
+  Spatial coverage: 38.9202 N, 106.9487 W (general centroid, East River watershed, CO)
+  Keywords: soil water content, volumetric water content, hillslope hydrology, 
+             shale weathering, PLM, East River, Crested Butte
+
+Files in package (from metadata.distribution):
+  1. Soil_water_content_Fig4e.csv    (48,230 bytes, text/csv)
+  2. Dynamic_water_table_depthsFig2b.csv (19,456 bytes, text/csv)
+  3. Soil_CO2_concentrations_Fig4h.csv   (28,912 bytes, text/csv)
+  4. README.txt                          (3,844 bytes, text/plain)
+
+--- STEP 3: FILE CLASSIFICATION ---
+
+Evaluating each file:
+
+Soil_water_content_Fig4e.csv:
+  Filename positive indicators: 'water_content' -> soil moisture variable
+  Format: CSV/TXT with columnar structure
+  -> Classified as: CANDIDATE DATA PAYLOAD
+
+Dynamic_water_table_depthsFig2b.csv:
+  Filename: 'water_table_depths' - water table level, NOT a direct VWC/GWC/potential measurement
+  This is a derived hydrological variable (depth to water table), not target schema variable
+  -> Classified as: OTHER (not in target schema)
+
+Soil_CO2_concentrations_Fig4h.csv:
+  Filename: 'CO2_concentrations' - gaseous CO2, not soil moisture
+  -> Classified as: OTHER (not in target schema)
+
+README.txt:
+  Filename: 'README' -> DOCUMENTATION
+  -> Classified as: DOCUMENTATION
+
+--- STEP 4: FILE INSPECTION ---
+
+Tool call: Download first 20 lines of Soil_water_content_Fig4e.csv
+  from https://data.ess-dive.lbl.gov/catalog/d1/mn/v2/object/... (content URL)
+
+Response (header rows):
+  Row 0: "Date,PLM1._vol_water_content,PLM2._vol_water_content,PLM3._vol_water_content"
+  Row 1: "(units),(m^3/m^3),(m^3/m^3),(m^3/m^3)"
+  Row 2: "2016-11-19,0.280,0.255,0.218"
+  Row 3: "2016-12-05,0.278,0.253,0.210"
+  ...
+
+Column identification:
+  Timestamp/Date columns found: 'Date' (YYYY-MM-DD format, date only, no time component)
+  Soil moisture columns:
+    - 'PLM1._vol_water_content' -> VWC at PLM1 (units: m^3/m^3)
+    - 'PLM2._vol_water_content' -> VWC at PLM2 (units: m^3/m^3)  
+    - 'PLM3._vol_water_content' -> VWC at PLM3 (units: m^3/m^3)
+  Depth columns: NONE in file
+  Site columns: site embedded in column name pattern ('PLMi._vol_water_content')
+  Replicate columns: NONE (single sensor per site)
+
+Row count estimate: ~75 rows (date-indexed from Nov 2016 to Oct 2021)
+
+Wide-format analysis:
+  Data is in WIDE format with sites as column variables
+  Pattern: 'PLMi._vol_water_content' where i = 1, 2, 3
+
+Tool call: Download README.txt
+  "Field data supporting Wan et al. (2024) 'Hydrological control of rock carbon 
+   fluxes from shale weathering'. Nature Water.
+
+   This dataset contains soil water content (volumetric) measured at the PLM 
+   hillslope transect (PLM1, PLM2, PLM3) using TDR sensors deployed at specific 
+   depths within instrumented boreholes. Measurement depths are reported in Figure 4 
+   of the associated manuscript.
+
+   Site location: Pumphouse Lower Montane (PLM) hillslope transect, East River 
+   watershed, near Crested Butte, Colorado (38°55'N, 106°57'W).
+
+   DOI of associated manuscript: https://doi.org/10.1038/s44221-024-00293-8"
+
+--- STEP 5: LOCATION RESOLUTION ---
+
+Source 1 (location metadata file): NO dedicated location file in package
+Source 2 (data payload file): No latitude/longitude/easting/northing columns in CSV
+Source 3 (package ESS-DIVE metadata): 
+  spatialCoverage provides site centroid (38.9202N, 106.9487W) for the PLM hillslope
+  but does NOT provide individual sensor coordinates for PLM1, PLM2, PLM3
+Source 4 (README): Contains approximate site location (38°55'N, 106°57'W = site centroid)
+  but NOT per-sensor coordinates
+Source 5 (Varadharajan et al. lookup):
+  PLM1, PLM2, PLM3 are known WFSFA monitoring stations
+  These should be in the East Taylor Watershed Community Observatory Sites registry
+  -> RECOMMEND: Varadharajan et al. lookup (qc_flag = "g1")
+Source 6 (Unresolvable): Not applicable; Varadharajan lookup should work
+
+CONCLUSION:
+  location_resolution.source = "varadharajan_lookup"
+  qc_flag_recommendation = "g1"
+  site_coordinates = [] (will be resolved at harmonization time from ref dataset)
+
+--- STEP 6: EXPERIMENTAL MANIPULATION DETECTION ---
+
+Scanning title, description, README, column names for manipulation keywords:
+  'warming': NOT found
+  'heated': NOT found
+  'irrigation': NOT found
+  'water addition': NOT found
+  'fertiliz': NOT found
+  'N addition': NOT found
+  'elevated CO2': NOT found
+  'drought': NOT found
+  'treatment': NOT found
+  'control': NOT found (in experimental sense)
+  'experimental': NOT found (only 'experiment' in metadata, referring to observational study)
+  'burn': NOT found
+  'fire': NOT found
+
+Context: PLM sites are natural ambient hillslope monitoring stations. Study purpose is
+observational (documenting natural hydrology and carbon cycling from shale weathering).
+
+DECISION:
+  manipulation_detected = False
+  manipulation_type = None
+  has_control_data = None
+  recommendation = "include_all"
+
+--- STEP 7: TIME SERIES INFERENCE ---
+
+Evidence FOR time series:
+  - TDR sensors permanently deployed in boreholes: YES (from README)
+  - Date column spans 5 years (Nov 2016 - Oct 2021): YES
+  - Keywords: 'sensor', 'logger' not present but 'TDR sensors' in README
+
+Evidence AGAINST continuous time series:
+  - ~75 rows over 5 years: much sparser than hourly/daily continuous data
+  - Timestamp precision: date only (YYYY-MM-DD, no time component)
+  - This appears to be periodic/seasonal monitoring, not continuous logging
+
+DECISION:
+  is_timeseries = True (sensor-based deployment, multi-year record)
+  interval_min = None (computed from data; interval is irregular/seasonal)
+  reasoning = "TDR sensors permanently deployed in boreholes at PLM hillslope.
+               Date-only timestamps spanning 5 years; appears to be periodic
+               (monthly to seasonal) rather than continuous hourly monitoring.
+               Classified as time series due to multi-year sensor deployment
+               intent; interval to be computed from timestamp differences."
+
+--- STEP 8: EXCLUSION CRITERIA ---
+
+RULE 1 - DUPLICATE CHECK:
+  Checking exemplar pool mapping JSON for matching DOI/package_id...
+  doi:10.15485/2322567 not found in exemplar pool (as expected, it's the held-out)
+  No duplicate found.
+
+RULE 2 - MEASUREMENT TYPE:
+  VWC columns: 'PLM1._vol_water_content', 'PLM2._vol_water_content', 'PLM3._vol_water_content'
+  Direct observations of volumetric water content -> PASS
+
+RULE 3 - EXPERIMENTAL MANIPULATION:
+  No manipulation detected -> PASS
+
+RULE 4 - EXTRACTABLE PAYLOAD:
+  Soil_water_content_Fig4e.csv: parseable CSV with date + VWC columns -> PASS
+
+RULE 5 - MINIMUM METADATA:
+  Coordinates retrievable via Varadharajan et al. -> PASS
+
+DECISION: INCLUDE
+
+--- STEP 9: EXEMPLAR SELECTION ---
+
+Reviewing exemplar pool for similar structure to dataset 10:
+  - Needs to match: wide-format VWC, site embedded in column name, depth from external source
+
+  Comparing to exemplar pool:
+  idx 1,2: Wide format VWC+potential, many sites, UTM coordinates -> partially similar
+  idx 3:   Long format, soil water potential -> different format
+  idx 4:   Long format, SWP columns -> different  
+  idx 5:   VWC, discrete sampling -> partially similar
+  idx 6:   Wide format VWC, Snodgrass site -> partially similar
+  idx 7:   Long format, separate location file -> different
+  idx 8:   Long format, site+depth -> different
+  idx 9:   Wide format VWC with VWC_1/VWC_2, site in column, depth from metadata -> MOST SIMILAR
+             (both have: VWC in wide format, site encoded in column names, depth not in data,
+              periodic discrete sampling, no location file)
+
+Selected exemplar: idx 9
+Reason: Most similar structure - wide format VWC with site embedded in column names,
+        depth not reported in data (populated from external reference), discrete/periodic
+        sampling dates, no dedicated location file.
+
+=== OUTPUT BUNDLE SUMMARY ===
+package_id: ess-dive-01092fc392bc46d-20240819T143818677
+doi: doi:10.15485/2322567
+curator_decision: INCLUDE
+data_payload_files: [Soil_water_content_Fig4e.csv]
+location_source: varadharajan_lookup (qc_flag=g1)
+is_timeseries: True
+interval_min: None (computed from data)
+manipulation_detected: False
+similar_dataset_reference: idx 9
+open_questions:
+  1. Depth not in data; use Fig. 4 legend values (PLM1=0.30m, PLM2=0.28m, PLM3=0.20m)
+  2. First CSV row is units descriptor; skip with iloc[1:]
+  3. Individual site coordinates must be retrieved from Varadharajan et al.
+  4. Timestamp timezone: America/Denver (assumed from CO site location)
+
+=== END SKILL 1 AGENT TRACE ===

results/raw_runs/fold_08_holdout_10/run_0/curator/ess-dive-01092fc392bc46d-20240819T143818677_a3f7b2c91d4e.json

@@ -0,0 +1,52 @@
+{
+  "package_id": "ess-dive-01092fc392bc46d-20240819T143818677",
+  "doi": "doi:10.15485/2322567",
+  "curator_decision": "INCLUDE",
+  "exclusion_reason": null,
+  "data_payload_files": [
+    {
+      "filename": "Soil_water_content_Fig4e.csv",
+      "columns": [
+        "Date",
+        "PLM1._vol_water_content",
+        "PLM2._vol_water_content",
+        "PLM3._vol_water_content"
+      ],
+      "row_count_estimate": 75,
+      "file_size_mb": 0.048,
+      "column_preview": "Date,PLM1._vol_water_content,PLM2._vol_water_content,PLM3._vol_water_content\n(units),(m^3/m^3),(m^3/m^3),(m^3/m^3)\n2016-11-19,0.280,0.255,0.218\n2016-12-05,0.278,0.253,0.210\n..."
+    }
+  ],
+  "location_metadata_files": [],
+  "sensor_metadata_files": [],
+  "readme_content": "Field data supporting Wan et al. (2024) 'Hydrological control of rock carbon fluxes from shale weathering'. Nature Water.\n\nThis dataset contains soil water content (volumetric) measured at the PLM hillslope transect (PLM1, PLM2, PLM3) using TDR sensors deployed at specific depths within instrumented boreholes. Measurement depths are reported in Figure 4 of the associated manuscript.\n\nSite location: Pumphouse Lower Montane (PLM) hillslope transect, East River watershed, near Crested Butte, Colorado (38\u00b055'N, 106\u00b057'W).\n\nDOI of associated manuscript: https://doi.org/10.1038/s44221-024-00293-8",
+  "location_resolution": {
+    "source": "varadharajan_lookup",
+    "qc_flag_recommendation": "g1",
+    "site_coordinates": []
+  },
+  "time_series_inference": {
+    "is_timeseries": true,
+    "interval_min": null,
+    "reasoning": "TDR sensors permanently deployed in boreholes at PLM hillslope. Date-only timestamps spanning 5 years (2016-2021); appears to be periodic (monthly to seasonal) rather than continuous hourly monitoring. Classified as time series due to multi-year sensor deployment intent; interval to be computed from timestamp differences."
+  },
+  "experimental_context": {
+    "manipulation_detected": false,
+    "manipulation_type": null,
+    "has_control_data": null,
+    "recommendation": "include_all"
+  },
+  "similar_dataset_reference": {
+    "index": 9,
+    "reason": "Most similar structure: wide-format VWC with site embedded in column names (PLMi._vol_water_content vs VWC_i pattern), depth not reported in data (populated from external reference), discrete/periodic sampling dates, no dedicated location file."
+  },
+  "open_questions": [
+    "Depth information not included in data package. Must use depths from Fig. 4 legend of Wan et al. 2024 (PLM1=0.30m, PLM2=0.28m, PLM3=0.20m).",
+    "First row of CSV appears to be a units descriptor row and should be skipped (iloc[1:]).",
+    "Individual site coordinates for PLM1, PLM2, PLM3 not in package; must look up in Varadharajan et al. (ref index 0).",
+    "Timestamp is date-only (YYYY-MM-DD); timezone assumed America/Denver (site is in Colorado, USA)."
+  ],
+  "skill_version": "0.1",
+  "run_id": "a3f7b2c91d4e",
+  "timestamp": "2026-06-30T22:55:00Z"
+}