earthbench-boot-al-global-static-v21-noqid-generalize-repair2

name	earthbench-boot-al-global-static-v21-noqid-generalize-repair2
description	Global static Earth-Bench executor skill for routing remote-sensing questions by filename and task family, then applying the correct thermal, index, aggregation, scene, change, counting, or geometry tools. Use for benchmark tasks over spectrum, raster products, and RGB imagery when the executor must stay within the current question directory, pick tools from file cues, reduce outputs into scalar evidence, and emit a calibrated final answer.
allowed-tools	["ATI","ChangeOS","InstructSAM","MSCN","RemoteSAM","SAM2","SM3Det","analyze_hotspot_direction","apply_cloud_mask","average_ratio_exceeding_threshold","band_ratio","bboxes2centroids","calc_batch_fire_pixels","calc_batch_image_hotspot_percentage","calc_batch_image_hotspot_tif","calc_batch_image_max","calc_batch_image_mean","calc_batch_image_mean_mean","calc_batch_image_mean_threshold","calc_batch_image_sum","calc_threshold_value_mean","calculate_area","calculate_band_mean_by_condition","calculate_batch_nbr","calculate_batch_ndsi","calculate_batch_ndti","calculate_batch_ndvi","calculate_batch_ndwi","calculate_bbox_area","calculate_max_lst_by_ndvi","calculate_mean_lst_by_ndvi","calculate_multi_band_threshold_ratio","[Truncated]"]
metadata	{"version":"1.0","benchmark":"Earth-Bench","release":"boot-al-global-static-v21-noqid-generalize-repair2","parent":"train124-r1-source","graph_policy":"internal_graph_global_static"}

Earth-Bench global static skill.

Skill metadata:

• benchmark: Earth-Bench
• version: boot-al-global-static-v21-noqid-generalize-repair2
• parent: train124-r1-source
• graph_policy: internal_graph_global_static

Induction source:

• This route policy is induced from train124 folded G2 supervision using gold_tool_names, gold_tool_chain_truth, and gold_trajectory.
• Treat that induction as a reusable routing prior only. Do not memorize instances, hidden answers, prompt snippets, task ids, sample ids, concrete filenames, or question-specific constants.
• Use placeholders such as <current-question>, current file list, current filenames, benchmark/data/<current-question>, and benchmark/out/<current-question>.

Global route scaffold:

• INIT -> CLASSIFY -> one of SPECTRUM / PRODUCTS / RGB -> CALIBRATE -> CONCLUDE.
• This graph is internal planning only. Never print phase names, next tags, transition markers, or state.
• Any node may emit a final <ANSWER>...</ANSWER> immediately when evidence is sufficient.
• Use only relative paths inside the current task directory tree. Read from benchmark/data/<current-question> and write or consume derived outputs under benchmark/out/<current-question>.

Node: INIT

Goal

• Build a minimal evidence inventory for the current task without leaking across tasks.

Instructions

1. Always start with get_filelist on benchmark/data/<current-question>.
2. Inspect the returned filenames before choosing any algorithm. File extension and recurring tokens are the primary router.
3. Mentally record: image type, number of dates, paired bands, whether there are support masks such as QA_PIXEL, whether outputs must be generated first, and whether the question asks for a scalar, ranking, date, trend, count, ratio, area, or class tally.
4. Keep all later tool arguments anchored to the current filenames only. Never borrow an output path from another task.
5. Use read_file only for small support files in the same question directory when the prompt clearly depends on side metadata.
6. If evidence is already direct, for example a single raster and a single mean request, you may solve from here by moving directly to calibration and answer emission.

Tool patterns

• get_filelist(dir_path='benchmark/data/<current-question>')
• read_file(file_path='benchmark/data/<current-question>/<support-file>')

Exit evidence

• You have the current file list.
• You know whether the task is spectrum, products, or RGB.
• You know the likely family and the expected final output form.

Node: CLASSIFY

Goal

• Convert filename cues plus prompt wording into one modality route and one family route.

Instructions

1. Route to RGB when the current file list is mostly .jpg, .jpeg, or .png and the task asks about scene category, counting visible objects, change masks between two images, or object geometry.
2. Route to SPECTRUM when filenames imply sensor-band retrieval or thermal physics, for example tokens such as LST, NDVI together with thermal data, BT10, BT_31, BT_32, Band 31, Band 32, Bands 10-12, Bands 10-14, Day, Night, emissivity, ATI inputs, or reflectance triplets used to derive temperature or drought variables.
3. Route to PRODUCTS when filenames are already geophysical or simple raster products: precipitation, nighttime light, built volume, MaxFRP, NDVI/NDWI/NDSI/NDTI/NBR rasters, cloud-masked surface reflectance, QA masks, turbidity, or annual / monthly comparisons over rasters.
4. Within SPECTRUM choose the family by cue:

• NDVI plus LST or drought language -> TVDI.
• BT10 plus red / NIR style cues -> single-channel LST.
• BT_31 plus BT_32 -> split-window or multi-channel LST.
• Bands 10-14 or emissivity separation -> TES.
• Bands 10-12 -> TTM.
• Day plus Night MODIS thermal -> MODIS day-night LST.
• day/night temperature plus albedo -> ATI.
• reflectance bands 2/5/17/18/19 -> band-ratio atmospheric absorption or water vapor.

5. Within PRODUCTS choose the family by cue:

• QA_PIXEL plus SR_B3/SR_B5 -> cloud-mask then NDWI.
• surface reflectance with index names or bands for NDVI / NDWI / NDSI / NDTI / NBR / turbidity -> derived-index.
• precipitation, avg_rad, B10 temperature summaries, simple time averages -> timeseries-aggregation.
• built_volume_total or built_volume_nres -> arithmetic.
• MaxFRP or change in fire intensity -> hotspot-fire.

6. Within RGB choose the family by cue:

• scene class counting over many chips -> scene.
• 'how many', 'sort by count', facilities, ships, courts -> counting.
• before/after buildings or disaster damage -> area-change.
• centroid, closest, farthest, westmost, northmost, roundabout, storage tank, harbor, field, plane -> grounding-geometry.

7. Prefer filename-first routing even if the natural-language prompt is vague.

Tool patterns

• get_filelist(dir_path='benchmark/data/<current-question>')
• Family selection examples: BT10 + b4 + b5 -> lst_single_channel; QA_PIXEL + SR_B3 + SR_B5 -> apply_cloud_mask + calculate_batch_ndwi; A.png/B.png pre/post -> ChangeOS; many .jpg chips -> MSCN.

Exit evidence

• One modality route is chosen.
• One tool family is chosen.
• You know the first computational tool to call.

Node: SPECTRUM

Goal

• Derive thermal, drought, emissivity, ATI, or band-ratio evidence and reduce it to the requested scalar, ranking, trend, or date.

Instructions

1. Use this route for physics-style raster derivations. Start from the raw band filenames and create derived outputs in benchmark/out/<current-question> only.
2. TVDI route:

• If the task asks for drought index, dryness, spikes in TVDI, threshold exceedance, or drought severity from NDVI and LST, call compute_tvdi using the current paired date files.
• Then use reducers according to the question: calc_batch_image_mean, calc_batch_image_mean_mean, calculate_tif_average, count_spikes_from_values, calculate_threshold_ratio, count_images_exceeding_threshold_ratio, calc_threshold_value_mean, or compute_linear_trend.

3. Single-channel Landsat LST route:

• If the current files imply BT10 plus red and NIR inputs, call lst_single_channel.
• For vegetation-conditioned summaries use calculate_mean_lst_by_ndvi or calculate_max_lst_by_ndvi.
• For overall means or threshold area use calc_batch_image_mean, calc_batch_image_mean_mean, calculate_threshold_ratio, calc_batch_image_mean_threshold, or kelvin_to_celsius when choices are in Celsius.

4. Split-window / multi-channel thermal route:

• If the task uses BT_31 and BT_32, choose lst_multi_channel or split_window according to the wording. Both belong to the same family of two-band temperature retrieval.
• Then reduce by count, mean, mean-threshold, ratio, or trend with calc_batch_image_mean, calc_batch_image_mean_mean, calculate_threshold_ratio, count_images_exceeding_threshold_ratio, count_images_exceeding_mean_multiplier, calc_batch_image_mean_threshold, max_value_and_index, min_value_and_index, or compute_linear_trend.

5. TES route:

• If the task uses ASTER thermal bands 10-14 or mentions emissivity separation, call temperature_emissivity_separation.
• Then use calculate_multi_band_threshold_ratio, calculate_band_mean_by_condition, image_division_mean, or calculate_threshold_ratio depending on whether the question is about combined LST-emissivity conditions, mean emissivity within hot regions, LST/epsilon, or emissivity variation above threshold.

6. TTM route:

• If the task uses ASTER bands 10-12, call ttm_lst.
• For urban-rural or region comparisons use calculate_band_mean_by_condition, calculate_threshold_ratio, calc_batch_image_max, calc_batch_image_mean_mean, difference, or centroid-derived geometry if the prompt also asks about region location.

7. MODIS day-night route:

• If the current files contain day and night brightness temperature plus emissivity, call modis_day_night_lst.
• Then use calculate_threshold_ratio, count_images_exceeding_threshold_ratio, average_ratio_exceeding_threshold, or calc_batch_image_mean_threshold for extreme-heat style tasks.

8. ATI route:

• If filenames imply day temperature, night temperature, and albedo or the prompt names Apparent Thermal Inertia, call ATI.
• Then use calc_batch_image_mean, calc_batch_image_mean_mean, calc_batch_image_mean_threshold, calculate_threshold_ratio, difference, or multiply when the threshold is a scaled mean such as 115 percent of average.

9. Band-ratio atmospheric absorption route:

• If reflectance bands 2, 5, 17, 18, and 19 are present, call band_ratio.
• Then use calc_batch_image_mean, calc_batch_image_mean_mean, mean, count_spikes_from_values, compute_linear_trend, count_images_exceeding_mean_multiplier, or calculate_threshold_ratio after scaling the mean with multiply.

10. Common reducer rules:

• calculate_tif_average creates an averaged raster. Follow it with mean, sum, hotspot percentage, or another reducer only if the question asks for a statistic of the averaged raster.
• calc_batch_image_mean returns per-image means as a list. Use mean only when a second-stage aggregation across dates or cities is required.
• calc_batch_image_mean_mean already behaves like a higher-level aggregate. Do not add extra mean unless the prompt explicitly asks for a second aggregation.
• Use max_value_and_index or min_value_and_index when the answer is a date or choice tied to the extreme value.
• Use compute_linear_trend for linear slope questions and sens_slope for Sen's Slope questions.

11. Threshold repair rules:

• If the tool gives proportion above threshold but the question asks for below threshold, compute the complement with difference from 100.
• If the threshold is a scaled mean like 110 percent or 115 percent of average, first compute the mean, then multiply, then apply the threshold ratio or count tool.
• Distinguish percent outputs from fraction outputs. Many ratio tools already return percent-like values; do not multiply by 100 unless the evidence clearly remains in [0,1].

12. Temporal grouping rules:

• Group by date string in the current filenames only.
• For month, season, or year comparisons, reduce within each period before comparing periods.
• For adjacent-date change logic, keep the current chronological order implied by filenames.

Tool patterns

• compute_tvdi(lst_path='benchmark/data/<current-question>/<lst-file>')
• lst_single_channel(bt_path='benchmark/data/<current-question>/<bt10-file>')
• lst_multi_channel(band31_path='benchmark/data/<current-question>/<bt31-file>')
• temperature_emissivity_separation(tir_band_paths=['benchmark/data/<current-question>/<band10-file>', '...'])
• ttm_lst(tir_band_paths=['benchmark/data/<current-question>/<band10-file>', '...'])
• modis_day_night_lst(BT_day_path='benchmark/data/<current-question>/<day-file>')
• ATI(day_temp_path='benchmark/data/<current-question>/<day-file>')
• band_ratio(sur_refl_b02_path='benchmark/data/<current-question>/<b02-file>')
• calculate_threshold_ratio(image_paths=['benchmark/out/<current-question>/<derived-file>'])
• count_images_exceeding_threshold_ratio(image_paths=['benchmark/out/<current-question>/<derived-file>', '...'])
• compute_linear_trend(y=[<period-means>])

Exit evidence

• You have derived the requested physical quantity.
• You have reduced it to a comparable scalar, ordered list, trend, date, count, or ratio.
• You can now map that evidence to answer choices or emit a direct numeric answer.

Node: PRODUCTS

Goal

• Solve raster-product, index, aggregation, arithmetic, turbidity, and fire-summary tasks that do not require RGB object grounding.

Instructions

1. Use this route when the task already lives in raster-product space or only needs simple index derivation plus aggregation.
2. Cloud-mask plus NDWI route:

• If the current directory contains QA_PIXEL, apply apply_cloud_mask to each required reflectance band before NDWI.
• Then call calculate_batch_ndwi on the cloud-masked current files.
• Reduce with calc_batch_image_mean, mean, compute_linear_trend, difference, or coefficient_of_variation depending on whether the question asks for mean, trend, change, or volatility.

3. Derived-index route:

• NDVI: calculate_batch_ndvi, then calc_batch_image_mean, mean, calc_batch_image_hotspot_percentage, max_value_and_index, or difference.
• NDSI: calculate_batch_ndsi, then calc_batch_image_mean, calculate_tif_average, calc_batch_image_hotspot_percentage, percentage_change, or difference.
• NDTI: calculate_batch_ndti, then calc_batch_image_mean, mean, max_value_and_index, calculate_tif_average, multiply, and calculate_threshold_ratio for high-turbidity area tasks.
• NBR: calculate_batch_nbr, then calc_batch_image_mean, mean, sens_slope, calc_batch_image_hotspot_tif, or analyze_hotspot_direction.
• Turbidity from red reflectance: calculate_water_turbidity_ntu, then calc_batch_image_mean, mean, max_value_and_index, skewness, or multi-day rolling comparison.

4. Timeseries-aggregation route:

• For precipitation, nighttime light, plain NDVI rasters, B10 temperatures, or simple daily products, call calc_batch_image_mean first.
• Use mean for period averages, percentage_change for consecutive or period-to-period change, max_value_and_index / min_value_and_index for extremes, compute_linear_trend for slope, kurtosis or skewness for distribution shape, and kelvin_to_celsius when the underlying values are Kelvin.
• When the question asks for the average of an already aggregated list, do the second reduction explicitly with mean.

5. Arithmetic route:

• For built volume or similar paired metrics, gather the necessary means or sums first.
• Use calc_batch_image_sum when the question is about total intensity or total volume after an averaged raster.
• Use division for ratios such as non-residential over total or light intensity per unit building volume.
• Use difference and percentage_change after the base quantities are aligned.

6. Fire / hotspot route:

• For MaxFRP counts over time, use calc_batch_fire_pixels, then compute_linear_trend, kurtosis, or max_value_and_index.
• For fire-intensity increases across periods, create comparable average rasters with calculate_tif_average, use subtract or create_fire_increase_map, and then count_above_threshold.
• For vegetation change hotspot maps across adjacent dates, use repeated calculate_tif_difference, then calc_batch_image_mean, calc_batch_image_hotspot_percentage, and max_value_and_index.
• For directional hotspot concentration after NBR, use calculate_batch_nbr -> calc_batch_image_hotspot_tif -> repeated analyze_hotspot_direction, then identify the dominant compass direction.

7. Common reducer rules:

• mean is for a list of scalar summaries, not for raw rasters.
• calculate_tif_average is for creating a multi-date mean map before later area or threshold analysis.
• When a question asks for the date of peak pollution or max vegetation threshold exceedance, pair the ordered statistics with filename order and use max_value_and_index.
• For average daily value over a period, first get per-image means, then reduce those means.

8. Statistical repair rules:

• coefficient_of_variation, skewness, and kurtosis summarize the distribution of already reduced time-series values.
• If a single bad value is returned as NaN, avoid inventing a fix; rely on the statistic tool if it still returns a usable benchmark value.
• For descriptive choices, match both sign and interpretation: increasing vs decreasing, right-skewed vs left-skewed, leptokurtic vs platykurtic.

Tool patterns

• apply_cloud_mask(sr_band_path='benchmark/data/<current-question>/<sr-file>')
• calculate_batch_ndwi(input_nir_paths='benchmark/out/<current-question>/<masked-nir-file>')
• calculate_batch_ndsi(green_file_list='benchmark/data/<current-question>/<green-file>')
• calculate_batch_ndti(input_red_paths='benchmark/data/<current-question>/<red-file>')
• calculate_batch_nbr(input_nir_paths='benchmark/data/<current-question>/<nir-file>')
• calculate_water_turbidity_ntu(input_red_path='benchmark/data/<current-question>/<red-file>')
• calc_batch_image_mean(file_list=['benchmark/data/<current-question>/<product-file>'])
• calculate_tif_average(file_list=['benchmark/out/<current-question>/<derived-file>', '...'])
• calc_batch_fire_pixels(file_list=['benchmark/data/<current-question>/<maxfrp-file>', '...'])
• percentage_change(a=<old-value>, b=<new-value>)

Exit evidence

• You have a reduced statistic or ordered comparison that matches the product-style question.
• You know whether the answer is a raw number, a trend description, a date, or a multiple-choice mapping.

Node: RGB

Goal

• Solve scene classification, visible-object counting, change detection, and geometry queries on RGB imagery.

Instructions

1. Scene route:

• If the prompt asks how many image chips belong to a land-use class such as industrial, airport, mountain, bridge, or farmland, run MSCN on each current image and count the chips whose predicted class matches the target class.
• Use the predicted top class first. Only rely on lower-ranked classes when the prompt explicitly asks for uncertainty handling.

2. Counting route:

• For questions about courts, diamonds, ships, or other repeated visible objects, prefer InstructSAM when the task is pure counting from a short text prompt.
• Prefer SM3Det when you also need bounding boxes, area ranking, or geometric follow-up on detected objects.
• For ranking tasks, compare counts or computed areas across A/B/C images and produce the full order, not only the largest.

3. Area-change route:

• For before/after building damage or change, run ChangeOS on the paired pre/post images in the current question directory.
• Then use calculate_area for changed area in pixels or real units if GSD is supplied.
• Use count_connected_components when the answer asks for number of fully destroyed buildings or number of changed building blobs.
• For two-region comparisons, run the full route independently for each region before comparing.

4. Geometry / grounding route:

• When the prompt names a semantic target such as roundabout, harbor, storage tank, tennis court, ground track field, or plane, use RemoteSAM if the query is language-rich and refers to attributes like largest, westernmost, northernmost, closest, or farthest.
• Use SM3Det when object class detection with boxes is sufficient.
• Convert boxes to centroids with bboxes2centroids.
• For closest or farthest pairs, call centroid_distance_extremes on the centroid list, then recover the corresponding objects with get_list_object_via_indexes if the task asks for boxes, or multiply the pixel distance by GSD with multiply if the task asks for meters.

5. Object-area route:

• If the prompt asks for area of detected facilities from boxes, use SM3Det followed by calculate_bbox_area.
• Then optionally use division and ceil_number to convert into equivalent numbers of courts or other reference units.

6. Mask-area route:

• If the prompt asks for actual object area rather than box proxy, use SAM2 or ChangeOS to obtain a mask, then calculate_area.
• When comparing two images with different GSD, compute each area in real units before difference.

7. Grounding repair rules:

• 'westernmost' means smallest x among relevant candidates.
• 'northernmost' means smallest y.
• 'largest' should be resolved before or together with the positional constraint when the prompt combines both.
• If the tool already supports the composite text prompt, ask for the composite target directly instead of resolving the pieces separately.

8. Counting repair rules:

• For ranking A/B/C images, keep each image's evidence separate and only then compare.
• Use exact counts when available. Do not infer ties without counting.
• When the answer choices are qualitative rankings, numeric counts or areas are only intermediate evidence.

Tool patterns

• MSCN(input_image_path='benchmark/data/<current-question>/<chip>.jpg')
• InstructSAM(text_prompt='basketball court')
• SM3Det(text_prompt='tennis court')
• ChangeOS(pre_image_path='benchmark/data/<current-question>/<pre-image>')
• calculate_area(input_image_path='benchmark/out/<current-question>/<mask-file>')
• count_connected_components(image_path='benchmark/out/<current-question>/<change-mask>')
• RemoteSAM(text_prompt='the largest roundabout on the westernmost side')
• bboxes2centroids(bboxes=[[x1, y1, x2, y2], ...])
• centroid_distance_extremes(centroids=[[cx, cy], ...])
• calculate_bbox_area(bboxes=[[x1, y1, x2, y2], ...])

Exit evidence

• You have class counts, object counts, mask area, centroid coordinates, pairwise distances, or bounding boxes.
• You can now map those quantities to the provided options or emit the direct numeric/string answer.

Node: CALIBRATE

Goal

• Convert raw tool outputs into the exact answerable form expected by the benchmark.

Instructions

1. Normalize the statistic before comparing choices: know whether the evidence is a pixel count, real area, mean value, proportion, percentage, slope, intercept, date, order, or direction label.
2. Preserve sign semantics. Increasing vs decreasing, warmer vs cooler, north vs south, June lower vs August lower, and urban vs rural all matter before nearest-number matching.
3. Use complement logic when needed:

• If the question asks for below-threshold share and the tool returned above-threshold share, use difference from 100.
• If the question asks for unaffected share after obtaining changed share, use the same complement rule.

4. Use scale logic when needed:

• For '115 percent of mean' or similar phrases, compute the mean first, then multiply, then apply the threshold tool.
• Do not multiply a value by 100 unless it is clearly still a fraction and the choices are percentages.

5. Use ordering logic when needed:

• max_value_and_index / min_value_and_index give the winning statistic; map the index back to the current filenames or ordered choice set.
• For closest or farthest object tasks, use the centroid-extremes output to recover the correct pair before answering.

6. Use comparison logic when needed:

• difference is for absolute gap or signed gap between two already aligned values.
• percentage_change is for relative change from an old baseline to a new value.
• division is for ratios such as per-unit-area intensity or nres over total.

7. For descriptive choices, match all constraints, not only the number:

• trend direction plus slope magnitude
• date plus value
• category order plus tie structure
• location word plus count
• skewness or kurtosis plus narrative description

8. For open numeric questions with no options, obey precision requests exactly, usually with three decimals or a single concise numeric/string output inside <ANSWER>...</ANSWER>.
9. If several options are numerically close, reject any option that conflicts with units, direction, time period, or object identity.
10. Once one choice fully fits all evidence, stop and answer. Do not keep calling tools after calibration is decisive.

Tool patterns

• difference(a=<value-a>, b=<value-b>)
• percentage_change(a=<baseline>, b=<new-value>)
• multiply(a=<mean>, b=<multiplier>)
• max_value_and_index(x=[<values>])
• min_value_and_index(x=[<values>])
• kelvin_to_celsius(kelvin=<kelvin-value>)

Exit evidence

• Exactly one answer candidate remains consistent with the evidence.
• The answer can be emitted in the required capture format.

Node: CONCLUDE

Goal

• Emit the final benchmark answer in the exact capture format.

Instructions

1. Output exactly two lines and nothing after them:

• <THOUGHT>brief evidence-based summary</THOUGHT>
• <ANSWER>final</ANSWER>

2. When choices are present, emit exactly one multiple-choice letter in the answer tag. Use the interface letter only. In this benchmark, the default expected form is one of <ANSWER>A</ANSWER>, <ANSWER>B</ANSWER>, <ANSWER>C</ANSWER>, or <ANSWER>D</ANSWER>.
3. When no choices are present, emit the final numeric or string answer only inside the answer tag and honor the requested precision.
4. Keep the thought line short, factual, and grounded in tool evidence.
5. Do not output JSON, bullet lists, phase names, extra explanations, or multiple answer tags.
6. If evidence becomes sufficient in INIT, CLASSIFY, SPECTRUM, PRODUCTS, RGB, or CALIBRATE, you may conclude immediately without waiting for any later node.

Tool patterns

• No more tool calls when you are already certain.

Exit evidence

• The final answer has been emitted in the exact capture format.