Datasets:

vimageiitb
/

GeoDisaster

This is the authoritative public specification for the benchmark. It explains the dataset, runtime contract, evaluation policy, release procedure, and all 43 question types as they are defined in the published version.

1. Benchmark purpose

GeoDisaster tests whether an agent can select and execute the correct tools, pass artifacts causally between them, combine heterogeneous geospatial sources, and produce a schema-valid answer. Upstream data acquisition is done during dataset construction. The cached rasters, vectors, tables, metadata, and imagery are intentional benchmark inputs: they make training and evaluation fast, deterministic, reproducible, and independent of live API availability, authentication, quotas, and later source changes.

The benchmark measures multi-source geospatial reasoning and tool use. For fast, repeatable execution, its inputs are downloaded and cached. The worker contracts also allow compatible new sources and scenarios to be registered and processed through the same agent workflow.

2. Authoritative dataset and evaluation contract

Canonical file: data/core/geodisaster_2921_source_with_splits.jsonl
Canonical SHA-256: 03988b960e6b09ae01fb622fe45b35d763eb705485cddc3cea68bcac7ae63fe1
Inventory: 2,921 records, 43 logical question types, 2,629 train and 292 public-reference test records.
Split isolation: no shared sample group, declared path, normalized routing-content group, or exact complete model-visible trajectory between train and test.
Agent-visible test file: data/core/geodisaster_2921_test_inputs.jsonl. It omits expected answers, GT chains, and internal build/replay provenance.
The test set is public reference gold, not a secret leaderboard. A fair agent run mounts only the isolated agent-visible inputs, not reference or teacher files.
Final answers are complete JSON values scored under geodisaster_structured_answer_v2: exact object keys, types, strings, booleans, null states, array lengths/order, and finite numeric values. Numeric leaves use absolute tolerance 1e-6; relative tolerance is disabled so accepted error does not grow with value magnitude. Finding the right number in the wrong field is not correct.
data/core/scoring_policy_43.json is the executable scoring authority. The nested correction_provenance.publication_contract.structured_scoring_policy value retained in canonical records names the historical publication pass that produced those records; it does not select or override the active v2 evaluator. Values under correction_provenance.evidence are opaque build-run labels, not runtime filesystem paths or agent-visible inputs.

3. How an agent solves a record

The evaluator creates a sanitized session and exposes the question plus named initial handles. The agent does not receive evaluator gold.
ReadData opens a declared input handle and returns a scene handle. Unsupported roles, times, options, or missing required paths fail.
PrepareScene validates compatible scene handles and creates a work handle. It is orchestration, not an undeclared data fetch.
Analytical workers consume the current handle and emit new result/artifact handles. A later analytical step consumes the preceding result when the GT chain declares a causal dependency.
After each successful tool call, the runtime attempts the declared schema-specific projection from worker evidence. It publishes ANSWER_JSON:answer_1 only when the complete projection validates.
TaskComplete closes a specialist subgoal; it does not contain the final answer. Multi-agent ORC finishes with Terminate(answer_ref=...), while the single agent finishes with FinalAnswer(answer_ref=...). The runtime then resolves and revalidates the referenced JSON for scoring.

The same worker contracts support both published agent forms. In the single-agent conversations, one agent plans, calls the workers, and answers. In the multi-agent conversations, ORC plans, delegates each analytical step to its declared specialist, receives the resulting handles/evidence, and terminates with the published answer reference. ORC does not replace or silently rerun a specialist computation or generate the resolved JSON.

Multi-agent owner	Workers	Responsibility
GA	`ReadData`, `PrepareScene`, `MeasureArea`, `ExtractObjects`, `MapFlood`, `RenderMap`	Open declared inputs, establish geospatial lineage, create or measure spatial artifacts, and render requested images.
PA	`ComputeStats`, `CompareSources`, `DetectChange`, `ClassifyChange`, `ScoreHazard`, `EstimateExposure`, `RankRegions`, `PlanRoutes`, `CompareRoutes`	Perform quantitative analysis, comparison, classification, exposure/ranking, and routing decisions.
VRA	`AssessDamage`, `DescribeFlood`, `CompareScenes`	Interpret damage/flood evidence and compare scene-level results.

Each question-type section below gives the exact subset, argument contract, handle handoff, calculation, and answer fields used for that task. A tool is not a generic answer lookup: it validates its declared arguments and consumes only the opened lineage required by that question.

Every runtime read is restricted to the exact repository-relative paths in the opened handle lineage. Hidden answers, ground truth, trajectories, and undeclared sibling files are denied. Null can be a valid schema value; a tool error, missing required input, non-finite required result, failed specialist contract, or unavailable terminal answer reference is not a valid empty answer.

4. Data, provenance, and release policy

All source paths are repository-relative. The canonical file contains 14,963 path references representing 5,681 unique paths. In the complete development snapshot, deep inspection resolved the declared paths and opened all 10,831 source files with zero empty files or read failures. The public thin package intentionally omits the Sen1Floods11 payload; run the checksum-verified hydrator described in ATTRIBUTIONS.md before executing those 500 records. For those records, an artifact card's exists: true states the hydrated execution contract, not that the separately distributed thin archive already contains the file. Thin-mode audit permits only the exact 200 upstream-derived paths named by the pinned manifest; hydration materializes and verifies them.

The package is mixed-license. ATTRIBUTIONS.md is the authoritative mapping from every packaged file family to its upstream product, citation, terms, and transformation. Project-authored questions, schemas, and chains do not relicense upstream imagery or data. The public release excludes the retained Sen1Floods11 payload and supplies a checksum-verified upstream hydrator.

5. Final validation baseline

2,921/2,921 prescribed chains executed successfully in the complete, hydrated development snapshot.
2,921/2,921 records passed field-aware substantive validation.
43/43 logical types are fully correct under their declared contracts.
The same 2,921/2,921 result reproduces in the hash-locked Python 3.10.12 replay environment with zero dependency or compatibility warnings.
A gold-free filesystem bundle containing only the 573 declared test files passes 292/292 reference-test records across all 43 types, with zero undeclared-path authorization violations.
Core projections and all multi-agent, single-agent, turn-level, and role-SFT exports have zero drift from the canonical source.
Invalid cross-family and causal-input mutations were rejected in 43/43 tests.
All 816 declared public question JSON sidecars are valid and contain no private AOI-type hints, GT paths, or final answers. An expected-answer shape describes required output fields and types; it is not answer gold.
Generated artifacts used by the published chains are nonempty.

The following sections specify each question type separately, including its question, declared files, worker arguments, causal sequence, calculation, output schema, edge cases, and verified population. Every type also states:

what the agent must do — the observable actions and final response;
what the agent must infer — the conclusion that must be derived from worker evidence rather than copied from an input; and
what the task measures — why that question belongs in an agentic geospatial benchmark.

6. Dataset construction, split, and answer registry

The canonical JSONL is the sole authority. Core train/test files, public test inputs, multi-agent and single-agent conversations, turn rows, and role-SFT rows are generated projections rather than independent datasets. Their IDs, splits, chains, answers, handles, and weights match the canonical source and the declared worker contract.

The public multi-agent role-SFT projections are stored only in data/multi_agent/sft_train/ and data/multi_agent/sft_test/. Every published conversation and SFT row carries id, sample_id, and question_type so it can be joined directly to the canonical record. Source inputs remain repository-relative. Non-distributed replay products referenced by teacher evidence use generated-artifact:// logical locators rather than build-host filesystem paths.

Released projection	Train rows	Test/eval rows
Core tasks	2,629	292
Full multi-agent episodes	2,629	292
Full single-agent episodes	2,629	292
Multi-agent ORC SFT turns	11,895	1,318
Multi-agent GA SFT turns	11,127	1,235
Multi-agent PA SFT turns	5,899	656
Multi-agent VRA SFT turns	2,763	302
Single-agent SFT turns	18,410	2,043

The role and single-agent turn projections contain 55,648 SFT rows in total. Single-agent training uses SFT plus deterministic/live evaluation. GRPO/RCEA is defined only for the four-role multi-agent system and does not require or imply a single-agent GRPO stage.

The split contract is geodisaster_source_path_trajectory_grouped_split_v4_20261015. Records connected by a source sample or declared input path remain in one split. Routing additionally groups normalized analytical-content duplicates, so path-renamed copies cannot cross the split. Exact complete model-visible single-agent episodes (messages plus tool_sequence, excluding identity metadata) are also grouped, even when their source samples differ. The final 2,629/292 split has zero shared source groups, declared paths, routing-content fingerprints, or complete-trajectory fingerprints. Repeated episodes within one split are retained for source provenance and receive inverse complete-trajectory multiplicity through training_weight. The published projection contains 41 such groups covering 114 records and zero cross-split groups; the 2,921 rows must not be described as 2,921 statistically independent scenes.

During SFT, the code repository constructs its internal validation partition from connected components of both split_group_id and trajectory_content_group_id; a component cannot contribute turns to both the internal training and validation subsets. The compact multi-agent GRPO prompt selector also keeps only one member of each canonical non-unit-weight duplicate trajectory group while preserving six prompts for each of the 43 logical types. These are derived training choices and do not modify this canonical release.

GT-chain steps are causal: a downstream worker consumes the declared current handle, not a hidden answer or a broad sample directory. A worker need not emit the final answer object verbatim. The runtime answer-contract layer deterministically projects requested status, units, fractions, and nested output fields from substantive worker evidence. The model receives only the published reference and never emits the resolved JSON. Correctness is decided only on the complete resolved JSON under the per-type schema; matching isolated numbers or an intermediate payload is insufficient.

All 2,921 multi-agent records and all 2,921 single-agent records use ANSWER_JSON:answer_1. Every one of the 5,842 terminal SFT targets uses the same exact reference-only action. The resolved worker projection is value-identical to canonical gold for 2,058 records; the other 863 records have the same complete structure and differ only by numeric display precision within the published tolerance. All 2,921 receive strict structured correctness.

The 292-row test source is public reference gold, not a secret leaderboard. For fair runtime evaluation, mount only the gold-free test-input projection and its declared files in an isolated filesystem. On 2026-10-14 the task owner also recorded human verification of all 2,921 canonical samples. Automated replay and source checks provide separate reproducibility evidence.

7. Reproducible runtime and file authorization

The publication replay environment is CPython 3.10.12 on Linux x86-64. The hash-locked environment fixes 13 packages, including Affine 2.4.0, NumPy 1.26.4, SciPy 1.11.4, Rasterio 1.4.3, PyProj 3.7.1, Shapely 2.1.2, and Pillow 11.3.0. The lock SHA-256 is 6c0419601f785ed45b60ce5c2d2927ad070d6aab5d987c424c764b8d6e4a6456. Binary wheels and hashes are required; source-build or dependency substitution is not permitted by the creation script. The reference native stack is GDAL 3.9.3, PROJ 9.5.1, and GEOS 3.13.1.

Create and verify it from the repository root:

release_tools/create_c11_replay_environment.sh
.venv-geodisaster-replay/bin/python release_tools/verify_c11_environment.py

Every tool call uses opened_handle_exact_path_v1: its read allowlist contains only exact file cards reachable through the question/input/work handles opened by that call. Ground truth, hidden/reference answers, trajectories, selected-route files, undeclared siblings, and broad directory fallback are denied at the read boundary. Deforestation quality reports are regenerated from declared rasters; fetch-time reports are provenance only. Missing optional epochs are reported rather than discovered by probing conventional filenames.

The final full replay audited 2,921 sessions, 14,611 tool calls, and 25,672 recorded reads with zero authorization violations. The isolated test replay audited 292 sessions, 1,459 calls, and 2,587 reads with zero violations. Strict isolation uses GEODISASTER_STRICT_DATA_ROOT=1, preventing fallback from the test bundle into the development repository.

All 14,803 generated worker artifacts in the final full replay are nonempty. The zero-byte answer_support_gaps.jsonl is intentionally an empty audit ledger meaning that no record lacks answer evidence; it is not an agent answer or a generated geospatial artifact.

8. Shared geospatial and source-data contracts

ATTRIBUTIONS.md is the authoritative file-family registry. It maps every packaged source and internal support category to its upstream product, citation, applicable terms, transformation, and redistribution status. The package is mixed-license: the project's CC-BY-4.0 terms for authored questions, schemas, chains, and documentation do not relicense upstream imagery, labels, or derived products. Earth Engine and OSMnx are processing facilities, not substitutes for citing the underlying products.

The retained development-snapshot audit classifies all 10,831 physical source files, every canonical reference, and every multi-hazard source label, with zero uncited or unregistered categories. The thin public release contains 9,328 files and its hydrated execution form contains 9,628; the mode-aware audit verifies the appropriate inventory and permits no unexpected missing required path. OpenAQ files record empty-query provenance only and are never answer evidence. WorldPop and OpenStreetMap are actual exposure inputs and retain their attribution requirements. The project-supplied xBD intermediate package is not treated as a separate external source; underlying xBD/xView2 imagery and labels retain their original attribution and non-commercial conditions.

Shared numerical rules are:

nodata is excluded while a valid numeric zero remains a measurement;
geographic raster areas sum WGS84 ellipsoidal cell areas, while projected grids use the affine determinant;
comparisons require the declared shape, CRS, and affine compatibility;
source geometry is transformed through the raster's real CRS and affine—no synthetic origin-zero transform is permitted;
null is used only where the schema declares a valid undefined/absent state; it is not a substitute for a missing file or failed computation; and
shared workers branch on declared domain/handle metadata, never on a sample ID or expected answer.

Official Sen1Floods11 S1Hand.tif bytes remain immutable provenance. The hydrator creates portable S1Portable.tif inputs with finite nodata -9999 while preserving finite pixels, dimensions, CRS, transform, Otsu threshold, and prediction. Cached Otsu/hand-label support products have zero canonical input references and are prohibited answer inputs.

9. Honest scope and interpretation limits

The benchmark's published claim includes the following scientific and evaluation limits:

Tasks use fixed cached inputs for reproducibility; they do not measure live upstream discovery, API reliability, or real-time hazard status.
Routing validates and renders packaged candidates. It is not a fresh road- graph search, a proof of globally nearest shelter, or evacuation advice.
Deforestation causes, drought/air classes, and multi-source agreement are explicit remote-sensing or benchmark-relative proxy constructs, not field- verified causality, health categories, or regulatory determinations.
Cached OSM way/relation centers are representative points, not complete line geometries. Questions and answers use that wording deliberately.
Sen1 A–E analyze a deterministic Otsu flood-candidate prediction, not its hand-label accuracy. The evaluator-only 100-scene check measures macro IoU/F1 0.206884/0.278917 and micro IoU/F1 0.206212/0.341916; this supports a reproducible high-recall/low-precision candidate baseline, not operational flood mapping.
The published population includes class imbalance, valid zero masks/counts, null branches, identical routing candidates, and source disagreement. It is designed for task-family coverage rather than balanced target classes.
Automated checks establish execution, dataflow, source consistency, and reproducibility. They do not waive upstream license conditions or create an unperformed independent scientific field validation.

10. Public-release procedure

The complete development workspace retains Sen1Floods11 files for local reproducibility. Because the upstream repository does not state an authoritative redistribution license for that payload, a public archive must exclude source_assets/sen1_flood/ and retain:

release_tools/fetch_sen1floods11_subset.py;
release_tools/manifests/sen1floods11_required_100.json;
release_tools/manifests/c10_excluded_legacy_rasters.json; and
release_tools/evaluate_c12_sen1_accuracy.py for evaluator-only construct validation, without mounting LabelHand files in the agent runtime.

In a clean public-release root, hydrate directly from the official source, verify checksums and deterministic derivatives, then replay all 500 Sentinel questions:

python3 release_tools/fetch_sen1floods11_subset.py
python3 release_tools/fetch_sen1floods11_subset.py --verify-only
.venv-geodisaster-replay/bin/python release_tools/replay_all_2921.py \
  --dataset-root . \
  --code-root /path/to/GeoDisaster/code \
  --input-jsonl data/core/geodisaster_2921_source_with_splits.jsonl \
  --output-dir /path/outside/repository/public_sen1_replay \
  --family sen1_flood --workers 8
python3 release_tools/audit_c01_source_registry.py --mode hydrated \
  --output /path/outside/repository/source_registry_hydrated.json

The source registry/path audit and 500/500 replay must pass before upload. The replay command projects a fresh final answer from the newly executed worker payloads and applies the active strict structured scorer. Stored historical source_replay.matches_expected flags and scalar evidence coverage are reported as provenance/diagnostics only and cannot produce an end-to-end pass. Preserve the complete development copy separately; construct the public archive from a clean release root.

Question type 1 of 43 — deforestation `forest_loss_detection`

Task and coverage

The agent decides whether real forest loss occurred in the stated event window. It must return deforestation or no_deforestation together with Hansen loss area, optical NDVI drop, Sentinel-1 VH change, and the rules that produced the decision.

Records: 48 AOIs — 43 train and 5 test.
Event years represented: 2019, 2020, 2021, 2022, and 2023.
Final labels: 27 deforestation, 21 no_deforestation.
One answer schema and one tool-name sequence are used for all 48 records.
All 48 records have training_weight=1.0; this type contains no downweighted content aliases.

What the agent must do and infer

Do. Open the optical, SAR, and Hansen handles; prepare one compatible forest-change work context; request a cross-source comparison; classify that comparison with the declared binary rule; and produce sufficient evidence for the runtime to publish the exact schema-valid answer reference.

Infer. Decide whether the combined measurements represent genuine deforestation. The agent cannot equate any visible color change with forest loss or copy a prewritten label. It must use the analytical result that distinguishes confirmed forest loss from plantation harvest, water/flood edges, burn/bare soil, agricultural change, and no strong change signal. It then maps only the generic deforestation cause to the positive binary label.

Meaning in the benchmark. This task measures multi-modal evidence integration and causal tool use: optical imagery can show vegetation change, SAR can provide cloud-independent structural evidence, and Hansen supplies an independent loss-year/baseline-forest constraint. Success demonstrates that the agent can route three complementary real-source inputs through the correct workers, preserve their provenance, interpret the resulting rule evidence, and produce a strictly structured conclusion. It does not test live satellite download or ask the model to classify RGB pixels without tools.

Declared sources and files

All files for an AOI are below source_assets/deforestation/<sample_id>/ and are exposed through exactly three analytical input handles.

Handle	Declared files	Purpose
`optical_stack_1`	Five or six `cache/s2/{pre_,post_}.tif` files and matching `visible/s2_rgb/*.png` panels	Sentinel-2 B2/B3/B4/B8/B11/B12 data for NDVI/NBR; PNGs provide visual context. Twenty-seven AOIs have six epochs and 21 have five; every AOI retains usable pre and post epochs.
`sar_stack_1`	Five `cache/s1_grd/s1_*.tif` files	Sentinel-1 VV/VH evidence. Every AOI has two declared pre-event and three post-event acquisitions. Each path carries its real acquisition date and `temporal_role`.
`forest_mask_1`	`masks/loss_year_<event_year>.tif`, `masks/lossyear.tif`, and `masks/treecover2000.tif`	Event-year Hansen loss, full loss-year codes, and baseline canopy.

The direct upstream products are Sentinel-2 L2A SR Harmonized, Sentinel-1 GRD, and Hansen/UMD Global Forest Change v1.13. Google Earth Engine was the acquisition/processing platform, not an additional data source.

Fetch-time quality_report.json files may remain beside the data for provenance, but they are not declared analytical inputs. Runtime quality is recomputed from the three Hansen rasters opened through forest_mask_1.

Required causal tool sequence

Step	Call	What it validates or computes	Output used by
1	`ReadData(question_1, optical_stack_1, role=pre_post, time=event_window)`	Declared optical files and metadata	`scene_1`
2	`ReadData(question_1, sar_stack_1, role=pre_post, time=event_window)`	Declared SAR files, dates, and roles	`scene_2`
3	`ReadData(question_1, forest_mask_1, role=hansen)`	Hansen paths and a fresh raster-quality report	`scene_3`
4	`PrepareScene(inputs={optical:scene_1, sar:scene_2, forest:scene_3}, mode=forest_change)`	Input types, roles, path status, and lineage	`work_1`
5	`CompareSources(work_1, signals=optical_sar_hansen, criterion=forest_loss_support)`	Optical, SAR, Hansen, and generic cause classification	`compare_1`
6	`ClassifyChange(compare_1, rule=forest_loss, labels={positive:deforestation, negative:no_deforestation})`	Consumes the comparison result and maps its cause to the binary task label	`class_1` and final answer evidence
7	`TaskComplete`	Closes the specialist subgoal; carries no terminal answer	ORC termination by published answer reference

ClassifyChange consumes compare_1; it does not bypass the comparison and silently recompute from work_1. Thus every analytical step in this chain is causal.

Calculation and decision contract

For each available Sentinel-2 epoch: NDVI = (B8 - B4) / (B8 + B4) and NBR = (B8 - B12) / (B8 + B12).
The optical drop is mean(pre) - mean(post). If the annual Hansen mask has positive loss pixels, the optical means are restricted to that loss region; otherwise the calculation uses the valid AOI pixels.
Sentinel-1 values are aggregated in dB using only the declared temporal roles. sar_vh_delta_db = post_mean_db - pre_mean_db; a negative value is a backscatter decrease.
Event-year loss is the sum of positive valid annual-mask cell areas. Geographic rasters use WGS84 ellipsoidal cell areas and projected rasters use the affine determinant. Hectares are rounded to six decimals.
If the annual mask contains no valid pixels, the worker—not the agent—uses the declared lossyear.tif code event_year - 2000. Valid code 0 means no loss. Nodata is 255 and is excluded.
The generic cause classifier applies these ordered rules:
- Hansen loss at least 0.5 ha and baseline canopy at least 50%: deforestation;
- Hansen loss at least 0.5 ha but canopy below 50%: plantation_harvest;
- no Hansen confirmation, but NDVI drop above 0.25 and VH delta below -1.5 dB: deforestation with a missing-Hansen flag;
- otherwise it may identify water/flood edge, burn/bare soil, agriculture, or no change from the published generic thresholds.
This binary task returns deforestation only when the generic cause is exactly deforestation; every other cause maps to no_deforestation.

No rule branches on an AOI ID or expected answer. Unsupported comparison criteria, labels, rules, roles, or times are rejected.

Final answer contract

The answer is exactly:

{
  "label": "deforestation | no_deforestation",
  "evidence": {
    "hansen_loss_area_ha": 0.0,
    "ndvi_drop": 0.0,
    "rules_fired": ["ordered rule strings"],
    "sar_vh_delta_db": 0.0
  }
}

The numbers above illustrate types, not target values. All four evidence fields are required, extra fields are rejected, and rules_fired order is significant. The agent formats this schema from class_1; it must not answer from the RGB panel alone.

Edge cases and verified result

Missing one optical epoch is allowed because 21 AOIs legitimately contain five rather than six epochs; both temporal sides must still be measurable.
Missing SAR role metadata or an empty pre/post SAR group is an error.
Zero canopy and zero loss are valid measurements; 255 is nodata.
An all-nodata annual mask triggers the declared full-loss-year fallback; it is not treated as a zero-sized or successful empty output.
Required evidence must be finite. No record in this type returns an empty answer or empty required artifact.

Validated contract. All declared files exist, the three sources are used for their stated roles, the chain is causal, the worker logic matches the question and schema, and 48/48 records passed execution plus field-aware answer validation.

Question type 2 of 43 — deforestation `area_estimation`

Task and coverage

The agent measures the total Hansen event-year forest-loss area inside the AOI and reports Sentinel-2 NDVI/NBR change over the same detected region as validation evidence. This is a quantitative area task; it does not request or return a binary deforestation label.

Records: 48 AOIs — 43 train and 5 test.
Event years represented: 2019, 2020, 2021, 2022, and 2023.
Results: 28 positive-area and 20 valid zero-area records.
Observed area range: 0 to 265.910267 ha.
One answer schema and one tool-name sequence are used for all 48 records.
All 48 records have training_weight=1.0.

What the agent must do and infer

Do. Open the Hansen and optical handles, prepare the forest-change work context, extract the event-year loss patches, measure their total area in hectares, and request NDVI/NBR validation against Hansen. The runtime projects the complete worker evidence into the exact area-and-evidence JSON schema.

Infer. Identify the correct event year from the question, interpret valid Hansen loss pixels rather than nodata, and distinguish a genuine zero-area measurement from an absent annual observation. The agent must combine the area result and the parallel optical-validation result. It must not invent a binary class or reject a source-derived area merely because an optical drop is small or negative.

Meaning in the benchmark. This task measures quantitative geospatial tool use: selecting the correct temporal raster, extracting a spatial phenomenon, measuring it on its real grid, and supporting that measurement with a second sensor-derived modality. It tests whether the agent can combine parallel analytical evidence into a strict final schema without confusing area measurement with classification.

Declared sources and files

Handle	Declared files	Purpose
`forest_mask_1`	`loss_year_<event_year>.tif`, `lossyear.tif`, and `treecover2000.tif`	Hansen event-year mask, fallback loss-year codes, and baseline canopy metadata.
`optical_stack_1`	Five or six Sentinel-2 six-band GeoTIFFs plus matching RGB panels	NDVI and NBR validation. Twenty-seven AOIs have six optical epochs and 21 have five; all have usable pre and post observations.

The upstream sources are Hansen/UMD Global Forest Change v1.13 and Sentinel-2 L2A SR Harmonized. No SAR handle is exposed because this question neither asks for nor uses SAR. Across this type, all 726 declared path references exist and remain inside the repository.

Required causal tool sequence

Step	Call	What it validates or computes	Output role
1	`ReadData(question_1, forest_mask_1, role=hansen)`	Hansen files plus a fresh runtime quality report	`scene_1`
2	`ReadData(question_1, optical_stack_1, role=pre_post, time=event_window)`	Optical paths and temporal role	`scene_2`
3	`PrepareScene(inputs={forest:scene_1, optical:scene_2}, mode=forest_change)`	Compatible inputs, path status, and lineage	`work_1`
4	`ExtractObjects(work_1, target=forest_loss_patches, filters={event_year:<question year>})`	Event-year mask selection, connected-patch summary, pixel count, and grid-aware area report	`objects_1`
5	`MeasureArea(objects_1, target=forest_loss, unit=hectares)`	Consumes the extracted-object report and returns area in hectares and square metres	`area_1`
6	`CompareSources(work_1, signals=ndvi_nbr_vs_hansen, criterion=validation)`	Computes loss-region NDVI/NBR drops and the independently reported Hansen area	`compare_1`
7	`TaskComplete`	Closes the specialist subgoal after `area_1` and `compare_1` are available; carries no terminal answer	ORC termination by published answer reference

ExtractObjects -> MeasureArea is causal. MeasureArea and CompareSources are intentionally parallel final-evidence branches: the first supplies the requested measurement and the second supplies the requested cross-source validation. Neither output is decorative or bypassed.

Calculation contract

ExtractObjects requires exactly target=forest_loss_patches and an event_year matching the question. A missing, malformed, additional, or conflicting filter is rejected.
Valid positive cells in loss_year_<event_year>.tif define the preferred loss mask. Declared nodata and non-finite cells are excluded.
If the annual mask has no valid cells, the worker selects code event_year - 2000 from the declared full lossyear.tif. Code 0 is valid no loss; nodata is 255. The fallback is a data-validity decision made by the worker, not a file choice delegated to the agent.
Area is summed cell by cell: WGS84 ellipsoidal area for geographic grids or affine-determinant area for projected grids. The result is rounded to six decimals in hectares.
ExtractObjects records the component count and largest-patch size in its intermediate report. MeasureArea validates the requested target/unit and consumes this result before exposing the final area measurement.
CompareSources computes NDVI=(B8-B4)/(B8+B4) and NBR=(B8-B12)/(B8+B12), then reports mean(pre)-mean(post). When positive Hansen loss pixels exist, optical evidence is restricted to that region; otherwise valid AOI pixels are used.
Optical values are validation evidence, not an alternative area estimate and not a hidden threshold that changes the Hansen area.

These are sample-independent rules. ExtractObjects, MeasureArea, and CompareSources reject unsupported target, year, unit, signal, or criterion arguments rather than silently applying a default.

Final answer contract

The answer is exactly:

{
  "loss_area_ha": 0.0,
  "evidence": {
    "hansen_loss_area_ha": 0.0,
    "ndvi_drop": 0.0,
    "nbr_drop": 0.0
  }
}

The numbers illustrate types only. loss_area_ha is the requested result; evidence.hansen_loss_area_ha identifies and confirms its source, so the two values must agree. NDVI and NBR drops may be positive, zero, or negative and must remain finite. The schema intentionally contains no label field.

Edge cases and verified result

Twenty records correctly return 0.0 ha; these are measurements, not empty answers.
Ten AOIs have annual masks with no valid cells. Every one correctly uses the declared full lossyear.tif fallback and resolves to valid code-0 no loss.
A five-epoch optical stack is valid when both pre and post evidence remain.
No record has a missing, null, non-finite, or contradictory required field.
The top-level and evidence Hansen areas agree in 48/48 records.

Validated contract. The canonical event-year arguments are correct in all 48 chains. The shared ExtractObjects worker rejects a filter year that conflicts with the question. The fresh focused run passed 48/48 executions, 48/48 field-aware answers, 726/726 paths, and produced 336 nonempty artifacts. A repository-wide regression passes 2,921/2,921 executions and 2,921/2,921 field-aware checks across all 43 types. Downstream MA, SA, turn, and role-SFT representations are verified exact projections of the canonical records.

Question type 3 of 43 — deforestation `change_window_estimation`

Task and coverage

The agent localizes the dominant optical clearing interval, conditions that interval on independently confirmed Hansen event-year loss, and reports whole-period optical/SAR context plus the selected-window score.

Records: 48 AOIs — 43 train and 5 test.
Results: 28 deforestation records with a window and 20 no_deforestation records with a null window.
Confirmed-loss area range: 13.326184 to 265.910267 ha; all unconfirmed records have valid 0 ha measurements.
One answer schema and one tool-name sequence are used for all 48 records.
All 48 records have training_weight=1.0.

What the agent must do and infer

Do. Open the optical time series, SAR time series, and Hansen handles; prepare one temporal forest work context; detect the dominant adjacent optical interval; classify the resulting detection under the Hansen gate; and return the label, nullable interval, and six exact evidence fields.

Infer. Determine which adjacent pair of available optical epochs has the largest NDVI_drop + 0.5*NBR_drop, then decide whether that optical fluctuation may be reported as a clearing interval. A candidate becomes a reported interval only when the independent event-year Hansen area is at least 0.5 ha. The agent must distinguish a valid null interval from an execution failure and must not treat SAR as the interval-selection signal.

Meaning in the benchmark. This task measures temporal localization, cross-source confirmation, missing-epoch reasoning, and causal artifact use. Sentinel-2 identifies when the strongest optical transition occurs, Hansen decides whether the transition is supported as event-year forest loss, and Sentinel-1 supplies independent whole-period structural context. It tests whether the agent can preserve these distinct source roles rather than collapsing every change signal into a positive answer.

Declared sources and files

Handle	Declared files	Purpose
`optical_timeseries_1`	Five or six Sentinel-2 six-band `cache/s2/{pre_,post_}.tif` files and matching RGB panels	AOI-wide per-epoch NDVI/NBR means for interval selection and optical support evidence.
`sar_timeseries_1`	Five Sentinel-1 `cache/s1_grd/s1_*.tif` files	Whole-period VH context from two declared pre-event and three post-event acquisitions with per-path dates and temporal roles.
`forest_mask_1`	`loss_year_<event_year>.tif`, `lossyear.tif`, and `treecover2000.tif`	Grid-aware event-year Hansen loss confirmation, fallback codes, and baseline canopy metadata.

The upstream products are Sentinel-2 L2A SR Harmonized, Sentinel-1 GRD, and Hansen/UMD Global Forest Change v1.13. Across this type, all 966 declared path references exist, are nonempty, and remain inside the repository.

Required causal tool sequence

Step	Call	What it validates or computes	Output role
1	`ReadData(question_1, optical_timeseries_1, role=ndvi_nbr, time=pre_post_epochs)`	Declared Sentinel-2 epochs and temporal metadata	`scene_1`
2	`ReadData(question_1, sar_timeseries_1, role=vh, time=pre_post_epochs)`	Declared Sentinel-1 files, dates, and pre/post roles	`scene_2`
3	`ReadData(question_1, forest_mask_1, role=hansen)`	Hansen files and fresh runtime raster quality	`scene_3`
4	`PrepareScene(inputs={optical:scene_1, sar:scene_2, forest:scene_3}, mode=forest_timeseries)`	Compatible types, roles, path status, and lineage	`work_1`
5	`DetectChange(work_1, change=dominant_clearing_interval, window=pre_post_epochs)`	Candidate interval, Hansen gate, and all requested evidence	`change_1`
6	`ClassifyChange(change_1, rule=change_window, labels={none:null, changed:start_epoch_to_end_epoch})`	Validates and formats the preceding detection result	`class_1` and final-answer evidence
7	`TaskComplete`	Closes the specialist subgoal; carries no terminal answer	ORC termination by published answer reference

ClassifyChange consumes change_1 directly. It performs zero raw-file reads and cannot recompute a different bundle from work_1; an incorrect input handle is rejected.

Calculation and decision contract

For every declared optical epoch, the worker computes AOI-wide valid-pixel means for NDVI=(B8-B4)/(B8+B4) and NBR=(B8-B12)/(B8+B12).
Only epochs with finite values for both indices are eligible. Eligible epochs retain canonical order pre_3, pre_2, pre_1, post_1, post_2, post_3; an undeclared epoch is skipped.
For each adjacent eligible pair (a,b), NDVI_drop=NDVI[a]-NDVI[b], NBR_drop=NBR[a]-NBR[b], and score=NDVI_drop+0.5*NBR_drop. The maximum score wins; an exact tie uses the earliest pair in canonical order.
Hansen event-year area follows the same nodata, annual-mask fallback, and grid-aware cell-area rules documented for area estimation. Area at least 0.5 ha confirms the candidate; a smaller area produces no_deforestation, a null window, and null selected-window fields.
The separately reported whole-period NDVI drop is mean(pre)-mean(post) within positive event-year loss pixels when present, otherwise over valid AOI pixels. It supports the answer but does not select the interval.
sar_vh_delta_db=post_mean_db-pre_mean_db uses the declared Sentinel-1 temporal roles. SAR is contextual evidence, not a hidden gate or optical interval selector.
Confirmed Hansen loss without at least two eligible optical epochs is an explicit processing error, not a misleading null/no-deforestation answer.

No rule branches on sample identity or target values. Unsupported change, window, rule, label, or source-handle arguments fail.

Final answer contract

The answer is exactly:

{
  "label": "deforestation | no_deforestation",
  "change_window": {"start_epoch": "pre_1", "end_epoch": "post_1"},
  "evidence": {
    "hansen_loss_area_ha": 0.0,
    "ndvi_drop": 0.0,
    "sar_vh_delta_db": 0.0,
    "selected_window_ndvi_drop": 0.0,
    "selected_window_nbr_drop": 0.0,
    "composite_drop_score": 0.0
  }
}

The values illustrate types only. For no_deforestation, change_window and the three selected_window_*/composite_drop_score fields are null; Hansen, whole-period NDVI, and SAR evidence remain finite numeric measurements. For deforestation, the interval object and all selected-window values are required, and the composite score equals the declared formula.

Edge cases and verified result

Twenty-one AOIs contain five optical epochs rather than six. The worker forms adjacency over the declared available sequence; seven selected windows therefore legitimately span the name of one missing epoch.
Twenty null windows are intentional negative outcomes, not empty files, empty masks, or missing answers.
All 28 reported windows have finite NDVI, NBR, and composite values; the formula agrees exactly in 28/28 records.
All 48 classification steps consumed change_1 with zero raw reads.
No required numeric field is non-finite and no required source is missing.

Validated contract. The prompt requests every schema field and states both the available-epoch and spatial-scope rules. Missing NBR is an error, and the classification worker directly consumes the detection artifact. A fresh focused replay passed 48/48 executions, 48/48 field-aware validations, 966/966 paths, and produced 336 nonempty artifacts.

Question type 4 of 43 — deforestation `false_positive_rejection`

Task and coverage

The agent applies a declared ordered rule set to distinguish the operational remote-sensing pattern behind an apparent change. The six outputs are benchmark proxy classes, not independently field-verified statements of real-world causation.

Records: 48 AOIs — 43 train and 5 test.
Outputs: 27 deforestation, 2 plantation_harvest, 3 burn_bare_soil, 5 agriculture, and 11 no_change.
water_flood_edge is a valid reachable schema/rule state but does not occur in the current 48-record population.
One answer schema and one tool-name sequence are used for all 48 records.
All 48 records have training_weight=1.0.

What the agent must do and infer

Do. Open the optical, SAR, and Hansen handles; prepare a false-positive analysis context; compare all three sources under cause_attribution; classify that comparison using the six-class ordered proxy rule; and return the exact class plus every structured measurement and ordered rule string required by the schema.

Infer. Determine the first matching operational pattern in the published rule order. The agent must preserve the distinction between a Hansen-confirmed loss pattern, a strong unconfirmed optical/SAR pattern, possible water-like SAR, burn/bare-soil-like optical change, agriculture-like vegetation change, and no strong signal. It must not present the proxy name as proof of actual land-use cause.

Meaning in the benchmark. This task measures confuser-aware reasoning and ordered decision execution. It asks whether the agent can combine multiple sensors, respect precedence between rules, carry a comparison artifact into a classification worker, and report enough exact evidence for the result to be audited. It measures consistent tool-mediated attribution under a declared benchmark taxonomy, not field causal discovery.

Declared sources and files

Handle	Declared files	Purpose
`optical_stack_1`	Five or six Sentinel-2 six-band GeoTIFFs and matching RGB panels	NDVI/NBR change patterns over the loss region when present, otherwise the AOI.
`sar_stack_1`	Five Sentinel-1 VV/VH GeoTIFFs	Whole-period VH change and post-event VV mean from declared pre/post acquisition roles.
`forest_mask_1`	Event-year mask, full `lossyear.tif`, and `treecover2000.tif`	Hansen loss area and AOI baseline canopy mean.

The sources are Sentinel-2 L2A SR Harmonized, Sentinel-1 GRD, and Hansen/UMD Global Forest Change v1.13. All 966 declared path references exist, are nonempty, and remain inside the repository.

Required causal tool sequence

Step	Call	What it validates or computes	Output role
1	`ReadData(question_1, optical_stack_1, role=pre_post, time=event_window)`	Optical files and epoch coverage	`scene_1`
2	`ReadData(question_1, sar_stack_1, role=pre_post, time=event_window)`	SAR files, dates, and temporal roles	`scene_2`
3	`ReadData(question_1, forest_mask_1, role=hansen)`	Hansen files and runtime raster quality	`scene_3`
4	`PrepareScene(inputs={optical:scene_1, sar:scene_2, forest:scene_3}, mode=forest_false_positive)`	Compatible types, roles, paths, and lineage	`work_1`
5	`CompareSources(work_1, signals=optical_sar_hansen, criterion=cause_attribution)`	All source measurements and ordered proxy classification	`compare_1`
6	`ClassifyChange(compare_1, rule=false_positive_type, labels=<six declared classes>)`	Validates and formats the preceding comparison	`class_1` and final-answer evidence
7	`TaskComplete`	Closes the specialist subgoal; carries no terminal answer	ORC termination by published answer reference

ClassifyChange requires a cause_attribution comparison and consumes compare_1 directly. Across all 48 records it performed zero raw-file reads; passing work_1 is rejected rather than triggering hidden recomputation.

Ordered proxy-rule contract

The worker first requires finite NDVI drop, NBR drop, VH delta, post-event VV mean, canopy mean, and Hansen loss area. Negative loss area or canopy outside 0–100% is an error. It then applies the first matching rule:

Hansen loss area at least 0.5 ha and canopy mean at least 50%: deforestation.
Hansen loss area at least 0.5 ha and canopy below 50%: plantation_harvest.
No confirmed Hansen loss, NDVI drop above 0.25, and VH delta below -1.5 dB: deforestation with a missing_hansen_confirmation intermediate flag.
No confirmed Hansen loss and post-event VV mean below -16 dB: water_flood_edge.
No confirmed Hansen loss, NBR drop above 0.20, and NBR drop greater than NDVI drop: burn_bare_soil.
No confirmed Hansen loss and NDVI drop above 0.10: agriculture.
Otherwise: no_change.

Optical drops use mean(pre)-mean(post) and are restricted to positive event-year Hansen pixels when available, otherwise valid AOI pixels. VH delta is post-pre; post-event VV is the declared post-group mean. Canopy is the valid-pixel mean of treecover2000.tif. Hansen area uses the documented grid-aware/fallback contract. These fixed thresholds are operational taxonomy rules, not universal ecological definitions.

Final answer contract

{
  "cause": "deforestation | plantation_harvest | water_flood_edge | burn_bare_soil | agriculture | no_change",
  "evidence": {
    "hansen_loss_area_ha": 0.0,
    "nbr_drop": 0.0,
    "ndvi_drop": 0.0,
    "rules_fired": ["ordered rule strings"],
    "sar_vh_delta_db": 0.0,
    "sar_vv_post_mean_db": 0.0,
    "treecover2000_mean_pct": 0.0
  }
}

The numbers illustrate types only. All six measurements are finite and all seven evidence keys are required. rules_fired is nonempty and ordered; extra answer fields are rejected.

Edge cases and verified result

The schema retains all six declared classes, including classes absent from some source populations.
A five-epoch optical stack remains valid when pre and post means exist.
Zero Hansen loss is a valid measurement and activates only the later unconfirmed-signal rules; it is not an empty mask result.
The answer explicitly exposes canopy and post-event VV as structured evidence.
No required field is null, non-finite, or missing in any of the 48 records.

Validated contract. The class labels, numeric decisions, and exact rule inputs are structured and auditable in all 48 targets. The focused replay passed 48/48 executions, 48/48 field-aware validations, 966/966 paths, and 336 nonempty artifacts. The shared forest-loss path also passed 48/48, and the repository-wide regression passed 2,921/2,921 executions and field-aware checks across all 43 types.

Question type 5 of 43 — deforestation `baseline_canopy_assessment`

Task and coverage

The agent computes mean baseline canopy density from the Hansen treecover2000 raster and assigns exactly one density class.

Records: 48 AOIs — 43 train and 5 test.
Classes: 31 dense_forest, 7 open_canopy, and 10 sparse.
Observed means range from 2.1365201473236084% to 99.99832916259766%.
All 48 records have training_weight=1.0.

What the agent must do and infer

Do. Open the one declared Hansen baseline raster, prepare the canopy work context, request its valid-pixel AOI mean, classify that computed statistic, and produce sufficient evidence for the runtime to publish the exact two-field JSON answer reference.

Infer. Determine the class from the computed mean: dense_forest when the mean is greater than 60%, open_canopy when it is from 30% through 60% inclusive, and sparse below 30%. The class is not stored in the input and cannot be inferred from a loss mask.

Meaning in the benchmark. This task measures whether an agent can select the minimal relevant raster, invoke a quantitative statistic, preserve the statistic-to-class causal dependency, and obey exact boundary and nodata semantics. It is a baseline canopy-density proxy, not a field survey or a claim that the mean alone defines ecological forest condition.

Declared source and file

The only analytical input is:

Handle	Declared file	Purpose
`forest_mask_1`	`source_assets/deforestation/<sample_id>/masks/treecover2000.tif`	Hansen percent tree cover for year 2000 on the packaged AOI grid.

The upstream product is Hansen/UMD Global Forest Change v1.13. No optical, SAR, annual-loss, or full lossyear.tif input is exposed or read for this task. Across the type, all 96 declared path references—the question JSON and tree-cover raster for each record—exist inside the repository.

Required causal tool sequence

Step	Call	What it validates or computes	Output used by
1	`ReadData(question_1, forest_mask_1, role=hansen)`	The exact declared tree-cover path; it does not build a loss-quality report	`scene_1`
2	`PrepareScene(inputs={forest:scene_1}, mode=baseline_canopy)`	Input type, role, path status, and lineage	`work_1`
3	`ComputeStats(work_1, metrics={treecover_mean_pct:true}, region=aoi)`	Valid-pixel arithmetic mean and valid-pixel count	`stats_1`
4	`ClassifyChange(stats_1, rule=canopy_density, labels=<declared thresholds>)`	Validates and classifies the preceding statistic	`class_1`
5	`TaskComplete`	Closes the specialist subgoal; carries no terminal answer	ORC termination by published answer reference

ClassifyChange accepts only the matching baseline-canopy stats_1 result, checks that its embedded class agrees with the mean, and performs no raw-file read. Passing work_1, an unsupported metric/region, or inconsistent result is an error.

Calculation and decision contract

Read band 1 of the declared treecover2000.tif only.
Exclude non-finite pixels and the declared finite nodata value 255. Tree-cover value 0 is valid, and all valid values must lie in 0–100.
Fail if no valid pixels remain; otherwise compute the unrounded arithmetic mean over all valid pixels in the packaged AOI raster.
Classify the mean with mutually exclusive, exhaustive boundaries:
- dense_forest: mean > 60;
- open_canopy: 30 <= mean <= 60;
- sparse: mean < 30.

The rules are sample-independent and contain no AOI-ID or expected-answer branch. Both statistic and classification workers reject missing, non-finite, negative, or above-100 canopy values.

Final answer contract

{
  "baseline_class": "dense_forest | open_canopy | sparse",
  "treecover2000_mean_pct": 0.0
}

The number illustrates the type, not a target value. Both keys are required, the class is enum-constrained, and extra keys are rejected.

Edge cases and verified result

Exact means 30 and 60 belong to open_canopy; values immediately below 30 are sparse, and values immediately above 60 are dense_forest.
Every packaged raster has at least one valid pixel. No record returns null, NaN, infinity, an empty artifact, or an unknown class.
Direct source recomputation reproduces all 48 stored means and classes.
The 48-record focused replay passed 48/48 executions and 48/48 field-aware validations, produced 192 nonempty artifacts, and found zero missing, absolute, or outside-root paths.

Validated contract. The final worker uses a dedicated tree-cover-only computation, the public inventory declares only that required raster, the 30/60 boundaries are unambiguous, and the classification step is causally tied to stats_1. The repository-wide regression remains 2,921/2,921 for execution and field-aware validation across all 43 types.

Question type 6 of 43 — deforestation `multi_year_loss_attribution`

Task and coverage

The agent attributes the largest mapped Hansen tree-cover-loss area in the AOI to one calendar year within 2001–2023.

Records: 48 AOIs — 43 train and 5 test.
Results: 35 records with a winning year and 13 valid no_loss records.
Positive winners span 2002–2023 and 0.07951–616.31121 ha.
No current record has a tie for the largest six-decimal area.
All 48 records have training_weight=1.0.

What the agent must do and infer

Do. Open the full Hansen loss-year raster, prepare a loss-history work context, compute grid-aware area independently for codes 1–23, detect the largest annual value, and return exactly its calendar year and six-decimal area—or the declared no_loss result.

Infer. Convert Hansen code n to year 2000+n, select the maximum from the computed annual histogram, apply the earliest-year rule if equal rounded areas occur, and distinguish valid code 0 from missing data. Neither the winning year nor its area is stored in the exposed input.

Meaning in the benchmark. This task measures temporal attribution over a categorical geospatial raster, grid-aware area calculation, deterministic selection, and causal use of a statistics artifact. “Loss” means Hansen's mapped tree-cover loss signal; it is not by itself a field-verified causal attribution or a live forest-monitoring result.

Declared source and file

Handle	Declared file	Purpose
`forest_loss_history_1`	`source_assets/deforestation/<sample_id>/masks/lossyear.tif`	Hansen loss-year code per AOI pixel.

The upstream product is Hansen/UMD Global Forest Change v1.13. No annual event mask, treecover2000, optical, or SAR input is exposed or read. All 96 declared references—the question JSON and one lossyear.tif per record—exist inside the repository.

Required causal tool sequence

Step	Call	What it validates or computes	Output used by
1	`ReadData(question_1, forest_loss_history_1, role=hansen, time=2001_2023)`	Exact loss-year path and metadata; no event-year quality report is generated	`scene_1`
2	`PrepareScene(inputs={loss_history:scene_1}, mode=forest_loss_history)`	Input type, role, path status, and lineage	`work_1`
3	`ComputeStats(work_1, metrics={loss_area_ha_by_year:true}, region=aoi)`	Validates codes and computes the 2001–2023 area histogram	`stats_1`
4	`DetectChange(stats_1, change=largest_loss_year, window=2001_2023)`	Validates the statistics contract and returns the deterministic winner	`change_1` and final answer
5	`TaskComplete`	Closes the specialist subgoal; carries no terminal answer	ORC termination by published answer reference

DetectChange consumes stats_1 and performs no raw raster read. A work_1 input, wrong window/mode, malformed year, or non-positive/non-finite histogram entry is rejected rather than recomputed or silently accepted.

Calculation and decision contract

Read band 1 of the declared lossyear.tif. Exclude non-finite pixels and nodata 255; require at least one valid integer code in 0–25.
Analyze only codes 1–23, mapping them to 2001–2023. Code 0 is valid no loss; codes 24–25 are valid source values but outside this task's time window.
For each analyzed code, sum actual grid-cell areas. Geographic rasters use WGS84 ellipsoidal cell area; projected rasters use the affine determinant. Express hectares to six decimal places.
Select the largest six-decimal annual area. If multiple years are equal, choose the earliest calendar year.
If codes 1–23 are absent, return year="no_loss" and loss_ha=0.0. An all-nodata raster is an execution error, not no_loss.

Final answer contract

{
  "year": "integer 2001..2023 | no_loss",
  "loss_ha": 0.0
}

Both keys are required, loss_ha must be non-negative, and extra keys are rejected. The JSON Schema permits only an integer in 2001–2023 or the exact string no_loss for year.

Edge cases and verified result

Every raster has valid pixels; 13 contain no analyzed loss code and correctly return the nonempty no_loss object.
Direct raster recomputation reproduces all 48 stored answers. No current maximum tie exists, but the generic earliest-year rule is published.
The final detector preserves the source-derived six-decimal area.
The focused replay passed 48/48 executions and 48/48 field-aware checks, with 192 nonempty artifacts and zero missing/outside paths.

Validated contract. The worker reads only the necessary loss-year raster, validates its domain, returns a schema-valid exact winner, and is causally tied to the preceding statistics result. The full regression remains 2,921/2,921 for execution and field-aware validation across 43 types.

Question type 7 of 43 — deforestation `cross_modal_concordance`

Task and coverage

The agent compares whole-AOI Sentinel-2 vegetation-index change with Sentinel-1 VH backscatter change and reports whether their binary indications agree.

Records: 48 AOIs — 43 train and 5 test.
Classes: 11 concordant, 14 discordant, and 23 no_change.
Every AOI has five Sentinel-1 acquisitions (two pre and three post).
Optical coverage is five or six epochs, always with usable pre and post data.
All 48 records have training_weight=1.0.

What the agent must do and infer

Do. Open the optical and SAR handles, prepare one cross-modal context, request their comparison, causally classify that comparison, and return the class plus exact NDVI, contextual NBR, and VH evidence.

Infer. Apply the two independent threshold decisions and combine them: both positive is concordant, exactly one positive is discordant, and neither positive is no_change. NBR must be reported but does not determine the class.

Meaning in the benchmark. This task measures sensor-fusion reasoning and the ability to distinguish agreement from mere change. Optical vegetation response and radar structural/backscatter response are complementary signals; the benchmark tests whether the agent preserves their temporal conventions, uses both, and follows a causal comparison-to-classification chain. The fixed thresholds are operational benchmark proxies, not universal physical laws.

Declared sources and files

Handle	Declared files	Purpose
`optical_stack_1`	Five or six Sentinel-2 six-band GeoTIFFs plus matching RGB panels	Whole-AOI NDVI/NBR pre/post evidence and agent-visible context.
`sar_stack_1`	Five Sentinel-1 VV/VH GeoTIFFs with declared acquisition dates and pre/post roles	Whole-AOI VH pre/post change.

The upstream products are Sentinel-2 L2A SR Harmonized and Sentinel-1 GRD. No Hansen mask is declared or read. All 822 declared path references exist inside the repository; the public sidecar inventory lists only optical, SAR, and SAR-window metadata.

Required causal tool sequence

Step	Call	What it validates or computes	Output used by
1	`ReadData(question_1, optical_stack_1, role=pre_post, time=event_window)`	Optical paths and temporal coverage	`scene_1`
2	`ReadData(question_1, sar_stack_1, role=pre_post, time=event_window)`	SAR paths, acquisition dates, and roles	`scene_2`
3	`PrepareScene(inputs={optical:scene_1, sar:scene_2}, mode=forest_cross_modal)`	Compatible inputs, path status, and lineage	`work_1`
4	`CompareSources(work_1, signals=ndvi_nbr_vs_vh, criterion=concordance)`	Finite whole-AOI evidence, thresholds, and preliminary class	`compare_1`
5	`ClassifyChange(compare_1, rule=concordance, labels=<three classes>)`	Verifies and formats only the preceding comparison	`class_1` and final answer
6	`TaskComplete`	Closes the specialist subgoal; carries no terminal answer	ORC termination by published answer reference

ClassifyChange accepts only the exact concordance comparison, verifies its evidence, thresholds, and embedded class, performs zero raw reads, and rejects work_1 or any other incorrect causal handle.

Calculation and decision contract

For each available optical epoch, compute AOI-valid-pixel NDVI and NBR. Each drop is the mean of available pre-epoch means minus the mean of available post-epoch means.
Aggregate Sentinel-1 VH in dB by each path's declared temporal role. vh_delta_db = post_mean_db - pre_mean_db.
Optical change is strictly ndvi_drop > 0.10; an exact 0.10 is false.
SAR change is strictly vh_delta_db < -1.0; an exact -1.0 dB is false.
Both true gives concordant; one true gives discordant; both false gives no_change. NBR is contextual evidence only.

All three evidence values must be numeric and finite. Missing temporal groups, NaN, infinity, unsupported criteria, or inconsistent comparison content are execution errors rather than no_change defaults.

Final answer contract

{
  "concordance": "concordant | discordant | no_change",
  "evidence": {
    "nbr_drop": 0.0,
    "ndvi_drop": 0.0,
    "vh_delta_db": 0.0
  }
}

The numbers illustrate types only. All fields are required, the class is enum-constrained, values must be finite, and extra keys are rejected.

Edge cases and verified result

Explicit boundary tests cover exact and immediately adjacent threshold values for both worker stages.
Direct raster recomputation exactly reproduces all 48 evidence objects and labels within the declared numerical tolerance.
No record has missing/non-finite evidence or an empty required artifact.
The focused replay passed 48/48 executions and field-aware checks, with 240 nonempty artifacts and zero missing, absolute, or outside-root paths.

Validated contract. Inputs, temporal roles, equations, threshold boundaries, schema, and causal handoff are explicit and consistent. The repository-wide regression remains 2,921/2,921 across all 43 types.

Question type 8 of 43 — deforestation `severity_proxy_numeric`

Population: 48 records — 43 train and 5 public-reference test. Each record has training weight 1.0.

What the task asks and why

Do. Open only the Sentinel-1 pre/post stack, prepare its SAR-change context, compute the whole-AOI VH change, score that numeric result, and return the exact finite delta and one severity bucket.

Infer. Use every available acquisition according to its declared temporal role. Compute post_mean_db - pre_mean_db, then apply the stated half-open threshold intervals. The bucket must come from the computed number, not from a filename, cached label, or unrelated optical/mask input.

Meaning in the benchmark. This task measures whether an agent preserves a signed radar-change convention, passes a computed statistic causally into a second worker, and handles numeric decision boundaries exactly. The four buckets are operational benchmark proxies, not universal forest-severity classes.

Declared sources and files

Handle	Declared files	Purpose
`sar_stack_1`	Five Sentinel-1 VV/VH GeoTIFFs, each with an acquisition date and `pre` or `post` role	Whole-AOI VH pre/post statistics.

The upstream product is Sentinel-1 GRD. No Sentinel-2 or Hansen file is declared or read. The 48 public sidecars expose only s1_grd and s1_window_metadata; all 288 declared path references exist inside the repository.

Required causal tool sequence

Step	Call	What it validates or computes	Output used by
1	`ReadData(question_1, sar_stack_1, role=pre_post, time=event_window)`	SAR paths, dates, roles, and availability	`scene_1`
2	`PrepareScene(inputs={sar:scene_1}, mode=sar_change)`	SAR context and lineage	`work_1`
3	`ComputeStats(work_1, metrics={vh_delta_db:true}, region=aoi)`	Declared-window whole-AOI VH delta	`stats_1`
4	`ScoreHazard(stats_1, hazard=forest_disturbance, scale=vh_delta_severity)`	Validates the exact upstream metric contract and applies the buckets	`score_1` and final answer
5	`TaskComplete`	Closes the specialist subgoal; carries no terminal answer	ORC termination by published answer reference

ScoreHazard accepts only the matching stats_1 result carrying the declared VH-delta measurement contract. It rejects work_1, a different metric result, an unsupported hazard/scale, or missing/non-finite evidence; it never reopens raw rasters or silently recomputes the preceding step.

Calculation and decision contract

Group the five opened Sentinel-1 acquisitions by their declared pre and post roles; invented dates or inferred filename chronology are forbidden.
Compute valid-pixel AOI mean VH backscatter for each acquisition, then the mean for each temporal group.
Set vh_delta_db = post_mean_db - pre_mean_db.
Apply: severe if < -3; moderate if -3 <= delta < -1; mild if -1 <= delta < 0; none if >= 0.
Therefore exact -3 is moderate, exact -1 is mild, and exact 0 is none.

Both temporal groups must exist and the result must be numeric and finite.

Final answer contract

{
  "vh_delta_db": 0.0,
  "severity": "severe | moderate | mild | none"
}

The number illustrates the type only. Both fields are required, severity is enum-constrained, and extra keys are rejected.

Edge cases and verified result

Direct source recomputation exactly reproduced all 48 numeric targets and buckets: 2 severe, 15 moderate, 18 mild, and 13 none.
Explicit probes on both sides of each threshold and at -3, -1, and 0 produced the documented classes.
No record has a missing temporal group, non-finite value, empty required artifact, missing path, or outside-root path.
The focused replay passed 48/48 executions and field-aware checks with 192 nonempty generated artifacts. The full regression remains 2,921/2,921.

Validated contract. The sole declared modality is sufficient, the sign convention and boundaries are explicit, the chain is causal, and all answers exactly match fresh computation.

Question type 9 of 43 — multi-hazard wildfire `A`

All multi-hazard GeoTIFFs use the corresponding sample AOI as their spatial contract. Source rasters are written on an EPSG:4326 grid derived from that AOI; derived rasters preserve the reference grid. A release audit checks every one of the 1,063 packaged raster/tabular GeoTIFFs across 108 samples for CRS, AOI extent, and within-sample co-registration. A nominal (0, 0) origin is invalid unless genuinely implied by the sample AOI. The current package has zero failures.

Population: 15 records — 13 train and 2 public-reference test. Each record has training weight 1.0.

What the task asks and why

Do. Open the wildfire measurement handle, prepare the scene, compute the burned footprint area from the declared event mask, and return a complete JSON answer containing the value, hectare units, source evidence, and completion status.

Infer. Treat finite mask value 1 as burned and 0 as valid unburned. Sum the actual grid-cell areas of the burned pixels and round once to three decimals; do not estimate area from a nominal 500 m scale or from an image panel.

Meaning in the benchmark. This task tests direct quantitative extraction from a source-specific hazard product, including nodata handling, geographic cell area, source attribution, units, and final-answer assembly. It does not ask the agent to classify burn severity.

Declared sources and files

Declared file	Purpose
`available_handles.json`	Confirms that the named cached source handle exists.
`rasters/mcd64a1__event.tif`	Binary event-window burned-area mask.
`rasters/mcd64a1__event.json`	Product provenance, source dates, units, shape, and raster checksum.

The upstream source is NASA MODIS MCD64A1.061 Burned Area, whose BurnDate product was transformed during dataset construction into the declared event-window binary mask. No dNBR, Sentinel-2, VIIRS, NIFC, panel, population, OSM, GT, or trajectory file is exposed to this task. All 60 path references across the 15 records are repository-relative, unique, and present.

Required causal tool sequence

Step	Call	What it validates or computes	Output used by
1	`ReadData(question_1, hazard_stack_1, role=wildfire, time=event)`	Exact handle inventory, MCD64 raster, and metadata paths	`scene_1`
2	`PrepareScene(inputs={hazard:scene_1}, mode=wildfire_burn)`	Wildfire context and declared-file lineage	`work_1`
3	`ComputeStats(work_1, metrics={burned_area:true}, region=aoi)`	Binary-mask validation, burned pixel count, grid-aware area, units, and source handle	`stats_1` and final answer
4	`TaskComplete(status=TASK_COMPLETE)`	Closes the specialist subgoal; carries no terminal answer	ORC termination by published answer reference

ComputeStats rejects a valid-but-wrong cross-domain metric such as forest_loss, as well as invented metrics or the wrong region. It reads only the three declared runtime dependencies and emits the area method and mask contract; no sample ID or expected answer participates in the calculation.

Calculation contract

Read the single-band MCD64 event mask and convert its declared finite nodata value to invalid pixels.
Require at least one valid pixel and require every valid value to be exactly 0 or 1. Empty burned support is permitted in principle and would yield 0.000 ha; an entirely invalid raster is an error.
Select valid pixels with value 1.
For EPSG:4326 grids, compute every selected cell's WGS84 ellipsoidal area; for a projected grid, use the affine determinant with CRS unit conversion.
Sum in hectares and round the final area to three decimal places.

Final answer contract

{
  "status": "TASK_COMPLETE",
  "answer": {
    "value": 0.0,
    "units": "ha"
  },
  "evidence": {
    "primary_area_ha": 0.0,
    "primary_handle": "mcd64a1__event"
  }
}

The numbers illustrate types only. answer.value and primary_area_ha must agree, both must be non-negative, the status/units/handle strings are fixed by schema, and extra keys are rejected. The runtime projects this object from stats_1 plus the literal terminal status; it never copies a stored answer.

Edge cases and verified result

All 15 rasters have valid binary 0/1 support, matching declared SHA-256 hashes, and nonempty burn masks. Fresh areas exactly match all 15 targets.
Areas span 548.919–23,052.567 ha; no NaN, infinity, empty analytical result, missing dependency, or fallback estimate occurs.
All 108 multi-hazard A records pass their domain-specific metric contracts; wildfire A passes 15/15 field-aware checks and produces 45 nonempty replay artifacts. The independent 20-type multi-hazard logic audit reports zero undeclared or missing runtime dependencies.
The full regression remains 2,921/2,921 and all 43 field-aware types pass.

Validated contract. The prompt, declared files, binary-mask semantics, grid-area calculation, worker output, schema, and targets are explicitly aligned.

Question type 10 of 43 — multi-hazard wildfire `B`

Population: 15 records — 13 train and 2 public-reference test. Each record has training weight 1.0.

What the task asks and why

Do. Open the dNBR and MCD64 inputs, prepare the wildfire scene, classify valid dNBR pixels only inside the MCD64 burned footprint, and return the dominant class/fraction plus complete per-class evidence.

Infer. Apply the four declared half-open dNBR intervals, count the valid masked population, select the class with the largest count, and divide that count by the complete classified population. Respect exact boundaries and the declared tie order.

Meaning in the benchmark. This task tests masked categorical reasoning: the agent must distinguish a source-defined analysis footprint from a continuous severity index, preserve class-boundary semantics, and synthesize a structured answer from worker evidence. The bins are an operational burn severity convention, not a field-validated damage assessment.

Declared sources and files

Declared file	Purpose
`available_handles.json`	Confirms both cached handles.
`rasters/dnbr.tif` and `.json`	Sentinel-2 pre-minus-post NBR change and its provenance/grid metadata.
`rasters/mcd64a1__event.tif` and `.json`	Binary MCD64 event burn footprint and metadata.

The upstream inputs are Sentinel-2 L2A SR Harmonized and NASA MODIS MCD64A1.061 Burned Area. No Landsat, VIIRS, NIFC, panel, exposure, GT, or trajectory file is exposed. All 90 path references across these records are relative, unique, and present.

Required causal tool sequence

Step	Call	What it validates or computes	Output used by
1	`ReadData(question_1, hazard_stack_1, role=wildfire, time=event)`	Exact dNBR/MCD64 paths and metadata	`scene_1`
2	`PrepareScene(inputs={hazard:scene_1}, mode=wildfire_burn)`	Wildfire context and declared-file lineage	`work_1`
3	`ScoreHazard(work_1, hazard=wildfire, scale=question_defined_class)`	Masked class table, dominant class/fraction, counts, areas, and means	`score_1` and final answer
4	`TaskComplete(status=TASK_COMPLETE)`	Closes the specialist subgoal; carries no terminal answer	ORC termination by published answer reference

ScoreHazard binds question_defined_class to B questions. It rejects the D task's impact_footprint scale, another domain, an invented scale, or an empty analysis population. The class worker verifies that the class counts exactly partition every finite dNBR pixel inside the MCD64 footprint.

Classification contract

Select finite dNBR pixels where the finite binary MCD64 mask equals 1.
Assign unburned for dNBR < 0.10, low for 0.10 <= dNBR < 0.27, moderate for 0.27 <= dNBR < 0.44, and high for dNBR >= 0.44.
For each class, report pixel count, six-decimal mean dNBR (or null for a possible empty class), and three-decimal grid-aware area in hectares.
Choose the largest pixel count. An exact tie selects the first class in order: unburned, low, moderate, high.
dominant_fraction = dominant_count / all classified valid pixels, rounded to six decimals. mask_pixel_count separately describes the MCD64 mask.

Final answer contract

{
  "status": "TASK_COMPLETE",
  "answer": {
    "class": "unburned | low | moderate | high",
    "dominant_fraction": 0.0
  },
  "evidence": {
    "dominant_class": "unburned | low | moderate | high",
    "by_class": {
      "unburned": {"area_ha": 0.0, "pixel_count": 0, "mean_value": null},
      "low": {"area_ha": 0.0, "pixel_count": 0, "mean_value": null},
      "moderate": {"area_ha": 0.0, "pixel_count": 0, "mean_value": null},
      "high": {"area_ha": 0.0, "pixel_count": 0, "mean_value": null}
    },
    "analysis_scope": "mcd64_burn_footprint",
    "mask_pixel_count": 1
  }
}

The numbers illustrate types only. The class strings are enum-constrained, fraction is in [0,1], counts/areas are non-negative, class and dominant_class must agree, and extra keys are rejected.

Edge cases and verified result

Fresh worker/source computation exactly reproduces all 15 complete answers. Dominant labels are 2 unburned, 2 low, 0 moderate, and 11 high; the absent dominant-moderate outcome is population evidence, not an unreachable class.
Boundary probes immediately below and at 0.10, 0.27, and 0.44 produce the declared bins. A four-way equal-count probe selects unburned.
Every current class row is nonempty. Footprints contain 385–15,679 pixels and dominant fractions range from 0.285139 to 0.538427.
All 108 multi-hazard B records, the 2,921-record regression, and all 43 field-aware type checks pass with zero worker or input-path errors.

Validated contract. Inputs, masking, thresholds, denominator, tie behavior, area/mean evidence, schema, and targets are explicit and aligned.

Question type 11 of 43 — multi-hazard wildfire `C`

Population: 15 records — 13 train and 2 public-reference test. Each record has training weight 1.0.

What the task asks and why

Do. Open three declared, co-registered binary footprints—MCD64A1 burned area, fixed-threshold Sentinel-2 dNBR support, and the NIFC incident perimeter—measure each footprint, and decide whether their areas agree.

Infer. Compute all three symmetric pairwise area differences, take their median, and compare it with the fixed 20% limit. Return the source areas and all pairwise evidence needed to inspect that decision.

Meaning in the benchmark. This tests multi-source reconciliation. The agent must distinguish source agreement from source truth: agreement means that three operational footprints have similar mapped areas under the stated rule, not that they are spatially identical or that any product is perfect.

Declared sources and files

Declared file	Purpose
`available_handles.json`	Verifies the exact cached comparison handles.
`rasters/mcd64a1__event.tif` and `.json`	Binary MODIS burned-area footprint and provenance/grid metadata.
`rasters/s2_burn_support.tif` and `.json`	Binary support derived by applying dNBR `>=0.10`; the sidecar records the derivation.
`rasters/nifc_perimeter__event.tif` and `.json`	Exact incident polygon rasterized on the common grid, plus provider/selection provenance.

The first ReadData handle carries only the inventory; the second carries the six analytical files. No RGB, Landsat, VIIRS, panel, AOI, exposure, GT, or trajectory artifact is exposed. All 120 references are relative, unique, present, and inside the repository. The upstream products are NASA MODIS MCD64A1, Sentinel-2 L2A SR Harmonized, and the NIFC WFIGS Interagency Fire Perimeters cited in ATTRIBUTIONS.md.

Required causal tool sequence

Step	Call	What it validates or computes	Output used by
1	`ReadData(question_1, hazard_stack_1, role=wildfire, time=event)`	Handle inventory and wildfire context	`scene_1`
2	`ReadData(question_1, agreement_sources_1, role=wildfire, sources=all)`	Three masks and sidecars	`scene_2`
3	`PrepareScene(inputs={hazard:scene_1,sources:scene_2}, mode=multisource_hazard)`	Exact causal lineage and input readiness	`work_1`
4	`CompareSources(work_1, sources=all, criterion=area_agreement, maximum_median_pct_delta=20.0)`	Areas, decision, score, and spatial support statistics	`compare_1` and final answer
5	`TaskComplete(status=TASK_COMPLETE)`	Closes the specialist subgoal; carries no terminal answer	ORC termination by published answer reference

CompareSources accepts only a type-C multisource_hazard work artifact with the three declared raster/sidecar pairs. It rejects another criterion, threshold, work type, missing lineage, non-binary wildfire mask, empty common valid population, or mismatched grid.

Agreement contract

Validate finite wildfire mask values as binary and require identical shape, CRS, and affine transform.
Sum actual grid-cell areas where each mask is 1. Geographic grids use WGS84 ellipsoidal cell areas; projected grids use the affine determinant.
For every source pair compute abs(A-B) / ((A+B)/2) * 100; two zero areas have difference 0.
The decision statistic is the median of the three pairwise percentages. median_pct_delta <= 20.0 means sources_agree; larger values mean sources_disagree.
agreement_score_0_100 = max(0, 100 - median_pct_delta). Compute from unrounded source areas and round emitted area/percentage/score values to three decimals.
IoU, full binary-state agreement, and Cohen's kappa use common valid cells only. They are support statistics and do not silently alter the area verdict.

The final JSON requires the decision, metric, median area/difference, bounded score, all three source areas, all three pairwise area differences, and all three sets of spatial support statistics. Metric, verdict, score contract, status, numeric ranges, and object keys are schema-constrained.

Edge cases and verified result

All 45 TIFF/sidecar checksum pairs match; every mask is nonempty, binary, AOI-aligned, and co-registered.
Fresh computation reproduces all 15 complete targets: 11 agree and 4 disagree. Median differences span 2.858–63.742%; scores span 36.258–97.142.
A boundary probe agrees at exactly 15.317% and disagrees when the limit is lowered to 15.316%, confirming inclusive threshold behavior.
The model-facing observation exposes the score contract, three source areas, and every pairwise field required by the final schema.
Focused execution and field-aware checks pass 15/15; all 40 adversarial argument/causal-input mutations are rejected for the intended reason.

Validated contract. The inputs, georeferencing, binary-mask semantics, exact area calculation, decision rule, support statistics, worker lineage, schema, and stored answers are aligned.

Question type 12 of 43 — multi-hazard wildfire `D`

Population: 15 records — 13 train and 2 public-reference test.

Agent task. Count distinct cached OSM highway-feature representative points within an inclusive 5 km of the binary MCD64A1 burn footprint. Use the actual GeoJSON geometry, transform it and the burn-cell centers into a local azimuthal-equidistant metric CRS, and return the count with footprint, population, in-footprint, nearest-feature, and method evidence.

Agent inference. Treat each declared OSM feature as one representative point and use authoritative GeoJSON coordinates. The output is determined by the inclusive metric distance to positive burn-cell centers, not cached row/column hints, road length, or network connectivity.

Benchmark meaning. This tests raster/vector integration and metric distance reasoning. The output counts one representative point per cached OSM feature; it is not road length, intersection count, accessibility, or damage.

The seven declared paths are the question JSON, available_handles.json, the MCD64A1 TIFF/sidecar, WorldPop TIFF/sidecar, and osm/pois.geojson. No panel, unrelated wildfire raster, AOI, GT, or trajectory is exposed. The chain is ReadData (wildfire), ReadData (highways), PrepareScene, ScoreHazard(scale=impact_footprint), EstimateExposure(within_km=5), and TaskComplete.

The GeoJSON coordinates are authoritative; cached row/col properties are not. in_footprint is calculated by transforming the representative geometry through the raster CRS and affine transform. The 5,000 m comparison is inclusive and uses the unrounded nearest cell-center distance. Points outside the AOI can count only when they exist in the declared cache and fall inside that buffer. WorldPop must match the footprint's shape, CRS, and transform and contain no negative values; it supplies supporting population evidence but does not affect the highway count.

The 15 primary counts are 48–249. Authoritative GeoJSON geometry determines all auxiliary in_footprint values; cached row/column hints are non-authoritative. Nearest distances span 1.951–21.115 m. Focused execution and field-aware validation pass 15/15, all outputs are finite and nonempty, and 43/43 adversarial calls are rejected. The worker also requires a nonempty wildfire type-D impact score and the complete declared raster/population/OSM lineage. The shared geometry audit independently checks all 20,927 cached OSM geometries and all 157 D-task in-footprint evidence values with zero mismatches. Valid zero-pixel drought and air footprints are accepted as measured zero-exposure outcomes rather than mistaken for failures.

Validated contract. Geometry authority, grid alignment, distance boundary, proxy interpretation, causal chain, schema, and answers are explicit and consistent.

Question type 13 of 43 — multi-hazard deforestation `A`

Population: 23 records — 21 train and 2 public-reference test.

Agent task. Open the declared Hansen event mask, prepare the forest-loss scene, compute the area of pixels whose loss year equals 2022, and return the three-decimal hectare value with the source handle.

Agent inference. The agent must infer the AOI's 2022 loss area from the worker's measured binary footprint; it must not substitute all-years loss, canopy area, an alert product, or an image-panel estimate.

Benchmark meaning. This tests source/temporal selection, causal tool use, binary-mask interpretation, nodata handling, and grid-aware geospatial area measurement on a second provider/domain contract.

The four declared paths are the question JSON, available_handles.json, and the hansen_loss__event TIFF/sidecar. The sidecar identifies Hansen Global Forest Change 2025 v1.13 with lossyear=2022; its checksum matches the TIFF in all 23 samples. Unused Sentinel-2, RADD, TMF, panel, and AOI paths are not exposed to this task.

The fixed chain is ReadData(role=deforestation,time=event), PrepareScene(mode=forest_loss), ComputeStats(metrics={forest_loss:true},region=aoi), and TaskComplete. ComputeStats accepts only the causal type-A prepared-work handle, the exact domain metric, and the declared primary TIFF, sidecar, and inventory lineage. It rejects a scene handle or cross-domain metric.

Finite mask value 1 means Hansen loss year 2022; 0 means valid no-2022-loss. Declared nodata/non-finite cells are excluded. Geographic cell areas are summed on the WGS84 ellipsoid (projected grids use the affine determinant), then rounded once to 0.001 ha. The answer schema fixes units="ha", primary_handle="hansen_loss__event", nonnegative areas, and status="TASK_COMPLETE". The build sidecar discloses that source-masked binary cells were filled with 0. This cannot create positive loss and gives the same positive-mask numerator as excluding those cells.

Independent raster/checksum recomputation matches every stored target. Each mask has 16,384 valid pixels and 56–767 positive pixels; answers span 107.692–1,449.402 ha. All 23 outputs are positive and finite. Cross-domain type-A replay and field-aware validation pass 108/108, and all 43 adversarial argument/causal-input probes are rejected for the intended reason.

Validated contract. The event year, provider handle, declared inputs, causal sequence, binary/nodata semantics, exact area method, schema, and all 23 answers are aligned.

Question type 14 of 43 — multi-hazard deforestation `B`

Population: 23 records — 21 train and 2 public-reference test.

Agent task. Open the Hansen 2022 loss mask and Sentinel-2 NDVI-change raster, prepare the forest-loss scene, classify every finite NDVI-drop pixel inside the loss footprint, and return the pixel-count-dominant class and its fraction with complete per-class evidence.

Agent inference. The agent must distinguish the dominant NDVI-drop bin inside a known loss footprint. A result of none means the majority has a drop below 0.15; it does not contradict the Hansen loss observation.

Benchmark meaning. This tests masked raster classification, temporal-index sign interpretation, boundary rules, grid co-registration, causal tool use, and evidence-based answer assembly.

The six declared paths are the question JSON, handle inventory, and the hansen_loss__event and ndvi_change TIFF/sidecar pairs. The change sidecar defines the value as Sentinel-2 NDVI pre minus post. Unused RADD, TMF, pre/post component rasters, panels, and AOI files are not exposed.

The chain is ReadData, PrepareScene(mode=forest_loss), ScoreHazard(hazard=deforestation,scale=question_defined_class), and TaskComplete. The scorer accepts only the causal type-B prepared work and declared files. It requires the two rasters to have identical shape, CRS, and affine transform and rejects a scene handle, wrong hazard, scale, or lineage.

Classification uses lower-inclusive, upper-exclusive bins: none [-10,0.15), minor [0.15,0.30), major [0.30,0.50), and severe [0.50,10). Every finite NDVI value at a finite Hansen-1 pixel must fall in exactly one bin. Dominance uses pixel count; an exact tie follows the declared none/minor/major/severe order. Per-class areas use exact grid-cell areas, and rounding occurs only on emitted fields.

Independent raster and checksum recomputation reproduces every answer and all evidence fields. All 23 dominant classes are none, with fractions 0.670360–1.000000. This is a real threshold outcome, not an empty-mask fallback: masks contain 56–767 pixels, 20 samples contain minor pixels, 17 contain major pixels, and 6 contain severe pixels. No dominant tie occurs. All 432 multi-hazard chains and 20/20 field-aware types pass, and 43/43 adversarial probes are rejected for the intended reason.

Validated contract. The NDVI sign, Hansen mask, thresholds, boundary/tie rules, grid alignment, causal chain, schema, and all 23 outputs are explicit and independently consistent.

Question type 15 of 43 — multi-hazard deforestation `C`

Population: 23 records — 21 train and 2 public-reference test.

Agent task. Open three declared, co-registered 2022 forest-disturbance masks; compute each footprint's exact area and the three symmetric pairwise area differences; apply the declared 25% median-difference threshold; and return the decision, bounded score, source areas, and spatial support statistics.

Agent inference. The agent must infer whether Hansen annual loss, JRC TMF annual loss, and RADD event-window alerts agree closely enough in footprint area. It must not treat zero alerts as missing data, compare rounded display areas, or substitute spatial overlap for the declared area verdict.

Benchmark meaning. This tests explicit temporal/source semantics, multi-source raster comparison, exact geospatial area calculation, robust aggregation, and structured evidence assembly. Agreement is an operational benchmark rule, not a claim that one product is ground truth.

The eight declared paths are the question JSON, available_handles.json, and the TIFF/sidecar pairs for hansen_loss__event, tmf_loss__event, and radd_alerts__event. Their checksums and temporal metadata are verified. The first mask is finite binary Hansen lossyear=2022; the second is finite binary JRC TMF change where December 2022 is deforested and December 2021 is not; the third uses 0=no alert, 1=low confidence, and 2=high confidence for RADD alerts dated 2022-06-01 through 2022-09-30. Values above zero form the RADD footprint. Unused panels, AOI files, and unrelated rasters are not exposed.

The chain is two ReadData calls, PrepareScene(mode=multisource), CompareSources(metric=area_agreement,maximum_median_pct_delta=25.0), and TaskComplete. The comparison worker reads only declared lineage, validates the allowed mask values and exact shape/CRS/affine co-registration, and fails on an all-nodata source or grid mismatch.

For source areas A and B, each pairwise percentage is abs(A-B)/((A+B)/2)*100; it is zero when both areas are zero. The verdict uses the median of the three unrounded percentages: at or below 25% means sources_agree; above 25% means sources_disagree. The bounded score is max(0,100-median_pct_delta). Emitted areas and percentages are rounded to three decimals. IoU, binary-state agreement, and Cohen's kappa use common valid cells and are support statistics; they do not alter the area verdict. TaskComplete.status is TASK_COMPLETE for either verdict because source disagreement is a successfully computed answer, not a contract failure.

Independent rasterio/PyProj recomputation reproduces all source areas, pairwise fields, support statistics, scores, and verdicts for 23/23 records. All 23 are sources_disagree; median differences span 48.008–200.000% and scores 0.000–51.992. Hansen and TMF areas are positive in every record. Nine RADD masks have a valid measured zero-alert footprint; the files still contain a valid grid and finite observations, so this is neither empty output nor a fallback. All 23 focused outputs have complete direct answer support, the full 2,921-record replay and field-aware audit pass, and no declared input is missing or outside the repository.

Validated contract. Temporal windows, source encodings, georeferencing, co-registration, zero-event semantics, exact area formula, threshold, support statistics, causal chain, schema, and stored answers are aligned.

Question type 16 of 43 — multi-hazard deforestation `D`

Population: 23 records — 21 train and 2 public-reference test.

Agent task. Measure the distance from the Hansen 2022 loss footprint to the nearest distinct cached OSM highway-feature representative point, or return an explicit unavailable state when no highway representative is cached. Also return the fixed 5 km context count, in-footprint count, nearest name, footprint area, and WorldPop footprint sum.

Agent inference. The agent must distinguish a measured distance from the valid no-feature state. It must use authoritative GeoJSON geometry and the current loss mask, not cached row/column hints, an AOI centroid, a rendered panel, or a road-line distance that the point-based cache cannot support.

Benchmark meaning. This tests causal raster/vector exposure analysis, coordinate transformation, metric nearest-neighbour reasoning, null-state handling, and honest interpretation of cached OSM representatives. It does not claim road-network access, road length, or field-verified infrastructure.

The seven declared paths are the question JSON, available_handles.json, the Hansen hansen_loss__event TIFF/sidecar, WorldPop TIFF/sidecar, and osm/pois.geojson. Panels, AOI files, alternate loss products, and unrelated rasters are not exposed. The chain is two ReadData calls, PrepareScene(mode=forest_loss_exposure), ScoreHazard(scale=impact_footprint), EstimateExposure(exposure=nearest_highway), and TaskComplete.

Finite Hansen value 1 is a 2022 loss cell and 0 is valid no loss. The worker requires a binary raster and exact WorldPop shape/CRS/affine co-registration. It transforms every raster-cell center and authoritative GeoJSON representative point into an AOI-centered azimuthal-equidistant CRS. Distance is the minimum point-to-positive-cell-center distance. Therefore a point located inside a positive raster cell can have a small nonzero cell-center distance; in_footprint separately records affine cell membership. The 5,000 m context boundary is inclusive. Counts are per distinct cached OSM feature.

Twenty records contain a cached highway representative and return finite distances from 14.648 to 4,136.076 m. Three contain none and consistently return availability_status=no_cached_highway_representative, null answer distance, null nearest name/distance, and zero highway counts. Six samples have at least one representative in a positive loss cell. These are valid source outcomes, not empty outputs.

An independent geometry audit reproduces all 157 multi-hazard D-task in-footprint counts, 5 km counts, nearest names, and distances from the source rasters and GeoJSON. It also verifies all 1,063 multi-hazard GeoTIFF grids and 20,927 cached OSM features. Focused and repository-wide replay use only the seven declared inputs. EstimateExposure emits the generically derived availability status alongside its nearest-feature evidence, and all 23 records have complete field-level worker support.

Validated contract. Source semantics, binary/nodata handling, geometry authority, projection, cell-center distance interpretation, buffer boundary, null state, causal lineage, schema, and all 23 answers are aligned.

Question type 17 of 43 — multi-hazard urban expansion `A`

Population: 29 records — 27 train and 2 public-reference test.

Agent task. Open the declared Dynamic World built-gain mask, prepare the urban-change scene, measure its exact area, and return hectares with the source handle.

Agent inference. The agent must identify new built probability between the 2019 and 2024 annual composites, rather than reporting total built cover, negative change, an NDBI proxy, or another provider's time window.

Benchmark meaning. This tests source and temporal selection, derived-mask interpretation, exact geospatial area measurement, causal tool use, and strict answer assembly.

The four declared paths are the question JSON, available_handles.json, and the built_gain TIFF/sidecar. The mask is derived from Dynamic World GOOGLE/DYNAMICWORLD/V1: post-2024 built probability above 0.55 and pre-2019 built probability at or below 0.55. The annual composite windows are 2019-01-01–2020-01-01 and 2024-01-01–2025-01-01. Unused NDBI, GHSL, WorldCover, component rasters, panels, and AOI files are not exposed.

The chain is ReadData(role=urban_expansion,time=event), PrepareScene(mode=built_up_gain), ComputeStats(metrics={built_up_gain:true},region=aoi), and TaskComplete. The statistics worker accepts only the matching causal Type-A work handle, metric, and declared TIFF/sidecar/inventory lineage. It requires a finite binary 0/1 mask. Geographic grids use WGS84 ellipsoidal cell areas; projected grids use their affine determinant; final hectares are rounded to three decimals.

Independent reconstruction from every pre/post Dynamic World pair exactly matches all 29 packaged gain masks, grids, and checksums. Independent area summation reproduces every stored answer, spanning 177.428–5,104.218 ha. All 29 outputs are finite and positive, focused evidence support is 29/29, and the full replay plus field-aware audit pass.

Validated contract. Provider, composite windows, threshold direction, derived pixels, binary validation, area method, causal lineage, schema, declared inputs, and all 29 answers are aligned.

Question type 18 of 43 — multi-hazard urban expansion `B`

Population: 29 records — 27 train and 2 public-reference test.

Agent task. Classify every finite whole-AOI Sentinel-2 NDBI-change pixel, then return the pixel-count-dominant change class, its fraction, and complete per-class counts, exact areas, and means.

Agent inference. The agent must preserve the signed NDBI_2024 − NDBI_2019 direction. Negative change represents contraction and must never be reported as positive urban growth.

Benchmark meaning. This tests temporal-index sign reasoning, exhaustive raster classification, interval boundaries, dominance/tie handling, causal tool use, and structured evidence assembly.

The four declared paths are the question JSON, handle inventory, and ndbi_change TIFF/sidecar. Component NDBI rasters, built-gain products, panels, AOI, GHSL, and WorldCover files are not runtime inputs. The chain is ReadData, PrepareScene(mode=built_up_gain), ScoreHazard(hazard=urban_expansion,scale=question_defined_class), and TaskComplete.

The complete lower-inclusive, upper-exclusive bins are contraction [-10,-0.02), stable [-0.02,0.02), minor [0.02,0.18), and major [0.18,10). Dominance uses pixel count; exact ties follow that declared order. Per-class areas use exact grid-cell areas. Every finite AOI pixel must fall in exactly one class.

Independent reconstruction verifies all 29 change rasters equal their Sentinel-2 post-minus-pre inputs. The stored results reproduce exactly: 15 contraction, 12 minor, and 2 stable dominant cases; no major-dominant case is manufactured. Focused execution and complete evidence support pass 29/29.

Validated contract. Change direction, scope, thresholds, coverage, area method, tie rule, causal lineage, schema, files, and answers are aligned.

Question type 19 of 43 — multi-hazard urban expansion `C`

Population: 29 records — 27 train and 2 public-reference test.

Agent task. Compare three source-native built-gain masks, calculate exact areas and symmetric pairwise differences, apply the fixed 30% median rule, and return the decision, bounded score, source areas, and spatial support statistics.

Agent inference. The agent must decide operational area agreement while recognizing that Dynamic World 2019–2024, GHSL 2020–2025, and WorldCover 2020–2021 measure different epochs and product constructs. Disagreement is a valid scientific outcome, not failed execution.

Benchmark meaning. This tests heterogeneous-source semantics, robust three-source comparison, exact area calculation, support-statistic separation, and complete structured synthesis.

The eight declared paths are the question, handle inventory, and three TIFF/sidecar pairs: built_gain, ghsl_built_gain, and worldcover_built_gain. Every mask is independently reconstructed from its provider-native pre/post inputs and is binary, checksum-valid, and exactly co-registered. Unrelated indices, panels, and AOI files are excluded.

The tool sequence is ReadData for the hazard inventory, ReadData for all three agreement sources, PrepareScene(mode=multisource_hazard), CompareSources(criterion=area_agreement, maximum_median_pct_delta=30.0), and TaskComplete.

The three symmetric differences use abs(A-B)/((A+B)/2)*100, with zero for two zero areas. The median of the unrounded differences is compared inclusively with 30%; the score is max(0,100-median). IoU, state agreement, and kappa use common valid pixels but do not alter the area verdict.

All stored evidence recomputes exactly. There are 28 disagreements and one agreement; median differences span 18.893–174.701% and bounded scores 0–81.107. The published population therefore contains 28 disagreements and one agreement under the declared source products and threshold. Focused execution and full evidence support pass 29/29.

Validated contract. Source epochs, gain definitions, co-registration, formulas, decision threshold, support statistics, chain, schema, and all answers are aligned.

Question type 20 of 43 — multi-hazard urban expansion `D`

Population: 29 records — 27 train and 2 public-reference test.

Agent task. Use the binary Dynamic World 2019–2024 built-gain footprint to sum co-registered WorldPop 2020 people-per-pixel values, then return that population with footprint area and cached hospital/school representative-point context.

Agent inference. The returned number is population located inside newly built cells at the WorldPop reference epoch. It is not a causal estimate of population growth or people added by urban expansion. The agent must preserve that distinction when selecting the exposure operation and formatting the answer.

Benchmark meaning. This tests raster-mask exposure estimation, temporal and source interpretation, exact grid alignment, vector/raster spatial context, causal tool routing, and strict structured synthesis.

The seven declared paths are the question, handle inventory, built_gain TIFF/sidecar, WorldPop 2020 TIFF/metadata, and one authoritative cached OSM GeoJSON. Unused NDBI, GHSL, WorldCover, panels, AOI files, and intermediate rasters are excluded. The chain is two ReadData calls, PrepareScene(mode=built_up_gain_exposure), ScoreHazard(hazard=urban_expansion,scale=impact_footprint), EstimateExposure(exposure=population_in_footprint), and TaskComplete.

The worker requires exact shape, CRS, and affine co-registration; validates a binary gain mask; sums only non-negative WorldPop count values where the mask equals 1; and computes footprint area from exact grid-cell areas. Hospital and school membership uses authoritative GeoJSON representative points mapped to positive cells. Inclusive 5 km context and nearest cell-center distances use a local azimuthal-equidistant CRS. Cached row/column hints are never geometry authority.

Independent source recomputation exactly reproduces all 29 population values, footprint areas, in-footprint counts, 5 km counts, nearest names, and distances. Population spans 51.108–612,394.383 people; zero hospital or school counts are valid measured outcomes, not missing data. Focused execution and complete evidence support pass 29/29, and all required artifacts are nonempty.

Validated contract. Source meaning, epoch interpretation, mask semantics, grid alignment, geometry, exposure sum, causal lineage, schema, declared inputs, and all answers are aligned.

Question type 21 of 43 — multi-hazard drought `A`

Population: 21 records — 18 train and 3 public-reference test.

Agent task. Open the declared MODIS NDVI target raster and six-observation same-season baseline, compute the target AOI mean and its z-score, and return that value with units and baseline mean.

Agent inference. The agent must distinguish a spatial target statistic from its temporal reference distribution. The target is the arithmetic mean of finite pixels in the actual modis_ndvi__target GeoTIFF; the comparison mean and sample standard deviation come only from the six non-target baseline rows.

Benchmark meaning. This tests source selection, spatial aggregation, temporal-standardization reasoning, causal tool use, and exact numeric answer assembly without treating a cached scalar as a substitute for the raster.

The five declared paths are the question, handle inventory, MODIS NDVI target TIFF/sidecar, and baselines/modis_ndvi_anomaly.csv. Other drought sources, derived masks, panels, and AOI files are not inputs to this task. The chain is ReadData(role=drought,time=event), PrepareScene(mode=drought_anomaly), ComputeStats(metrics={ndvi_z_score:true},region=aoi), and TaskComplete.

The worker requires a single-band target with at least one finite pixel and NDVI in [-1,1], exactly one dated target marker and at least two historical observations, and a finite positive baseline sample standard deviation. With six baseline values, it computes mean, std(ddof=1), and z=(target_AOI_mean-baseline_mean)/std; final values are rounded to six decimals. The CSV target row mirrors the raster mean for auditability, but the worker calculates the target from the GeoTIFF.

Independent Rasterio/NumPy recomputation reproduces all 21 answers. Final z-scores span -0.226802–3.313451, all target rasters have finite data, and focused execution plus complete evidence support pass 21/21.

Validated contract. The target raster is computational, baseline membership is explicit, the formula and ddof are fixed, validation rejects empty/nonfinite inputs, the chain is causal, and all answers match independent source computation.

Question type 22 of 43 — multi-hazard drought `B`

Population: 21 records — 18 train and 3 public-reference test.

Agent task. Classify every finite CHIRPS same-window SPI pixel into the declared train-relative quartile intervals, then return the pixel-count dominant class, fraction, and complete per-class counts, areas, and means.

Agent inference. The four labels describe the position of a pixel in this benchmark's training distribution. q1_low_spi is relatively low within that fit; none of the labels is a universal drought-severity category.

Benchmark meaning. This tests leakage-safe normalization selection, continuous-raster classification, interval boundaries, dominance/tie logic, exact geospatial area evidence, and structured synthesis.

The five declared paths are the question, handle inventory, CHIRPS SPI TIFF/sidecar, and drought_normalization.json. Other drought rasters, masks, panels, and AOI files are excluded. The chain is ReadData, PrepareScene(mode=drought_anomaly), ScoreHazard(hazard=drought,scale=question_defined_class), and TaskComplete.

The normalization fit contains all 294,912 finite pixels from the 18 training AOIs and excludes the three test AOIs. Recomputed quartile boundaries are exactly -0.1381913721561432, 0.6701323390007019, and 1.3935534954071045. The four intervals are lower-inclusive and upper-exclusive with open numerical coverage extended to ±1e30. Dominance uses pixel count and exact ties use q1-to-q4 declaration order. Areas use WGS84 ellipsoidal cells.

Independent split/quantile reconstruction and raster classification reproduce every target: 3 q1, 4 q2, 8 q3, and 6 q4 dominant cases. Every record classifies all 16,384 finite pixels; focused execution and complete evidence support pass 21/21.

Validated contract. Fit membership, test isolation, thresholds, interval coverage, interpretation, dominance, exact areas, lineage, schema, files, and answers are aligned.

Question type 23 of 43 — multi-hazard drought `C`

Population: 21 records — 18 train and 3 public-reference test.

Agent task. Derive one low-state mask from each co-registered CHIRPS SPI, MODIS NDVI, and ERA5 soil-moisture raster using its training-pixel q25; compute all three pairwise IoUs; and decide whether their median reaches 0.30.

Agent inference. Agreement concerns spatial overlap between three source-specific, benchmark-relative low states. It does not compare raw units, claim that q25 means universal severe drought, or treat a scientifically valid disagreement as tool failure.

Benchmark meaning. This tests multi-source normalization, thresholded mask construction, co-registration, spatial agreement, leakage-safe evidence use, and complete structured decision synthesis.

The nine declared paths are the question, handle inventory, train-only normalization contract, and three TIFF/sidecar pairs. The hazard handle carries the fit contract; the source handle carries the continuous rasters, avoiding duplicate declarations. The chain is two ReadData calls, PrepareScene(mode=multisource_hazard), CompareSources(criterion=iou,minimum_iou=0.3, threshold_contract=per_source_train_pixel_q25_v1, normalization_fit_split=train), and TaskComplete.

Each threshold exactly reproduces from 294,912 finite training pixels per source. A source mask is value < q25; pairwise IoU uses common valid, co-registered cells. Two empty masks have IoU 1 by the declared set rule. The decision is the median of the three six-decimal pairwise IoUs, inclusive at 0.30. Exact source-mask areas are evidence but do not change the decision.

Independent Rasterio/NumPy/PyProj reconstruction reproduces every threshold, mask area, pairwise IoU, score, and verdict. All 21 medians are honestly 0 and all 21 verdicts are sources_disagree; nine CHIRPS and twelve ERA5 AOI masks are valid measured-zero masks, while no MODIS mask is empty. The class balance is therefore 21 sources_disagree cases under the declared sources and rule. The worker emits all nine contract fields, and focused full-answer support passes 21/21.

Validated contract. Training fit, source thresholds, grid alignment, empty-mask rule, IoU aggregation, interpretation, worker evidence, lineage, schema, files, and answers are aligned.

Question type 24 of 43 — multi-hazard drought `D`

Population: 21 records — 18 train and 3 public-reference test.

Agent task. Count cached OSM school representative points whose authoritative geometry falls in a positive cell of the declared low-CHIRPS-SPI footprint, and return the count with footprint and exposure context.

Agent inference. The footprint is a prepared binary representation of CHIRPS SPI < training-pixel q25. The primary answer is a representative-point count, not a count of school polygons, students, buildings, or nearby schools.

Benchmark meaning. This tests derived-footprint interpretation, vector/raster spatial membership, empty-footprint handling, causal exposure routing, and exact structured synthesis.

The seven declared paths are the question, handle inventory, low_chirps_spi TIFF/sidecar, WorldPop TIFF/metadata, and authoritative cached OSM GeoJSON. The binary sidecar records the train-only threshold provenance; the runtime task consumes the prepared mask rather than redundantly rereading raw CHIRPS. The chain is two ReadData calls, PrepareScene(mode=drought_anomaly_exposure), ScoreHazard(hazard=drought,scale=impact_footprint), EstimateExposure(exposure=schools), and TaskComplete.

Independent reconstruction verifies every binary mask against its raw CHIRPS raster and stored training q25. School membership uses authoritative GeoJSON representative points transformed into the footprint grid. The 5 km context and nearest distance use a local azimuthal-equidistant CRS and positive cell centers. WorldPop is exactly co-registered supporting evidence. Cached row/column hints are not geometry authority.

Nine footprints are valid measured-zero masks, yielding zero area, population, school count, buffer count, and null nearest context as appropriate. Nineteen records contain zero schools in the footprint and two contain three; this source-driven imbalance is retained. All 21 answers reproduce, focused execution/support pass 21/21, and the independent 157-case multi-hazard D geometry audit reproduces every count and distance.

Validated contract. Mask derivation, zero-mask semantics, geometry authority, membership, distance, population context, causal lineage, schema, files, and all answers are aligned. The binary-mask chain reads no undeclared normalization file.

Question type 25 of 43 — multi-hazard air pollution `A`

Population: 20 records — 18 train and 2 public-reference test.

Agent task. Compute the maximum and 95th percentile of finite pixels in the declared Sentinel-5P episode-peak tropospheric NO2 raster, then return the maximum in µmol/m² with both statistics as evidence.

Agent inference. The raster already stores each pixel's maximum over the declared episode. The requested answer is the spatial maximum of that temporal peak composite—not a surface concentration, health category, or live reading.

Benchmark meaning. This tests product/unit interpretation, two-stage spatiotemporal statistic reasoning, declared-raster computation, causal tool use, and precise numeric synthesis.

The four declared paths are the question, handle inventory, and s5p_no2__peak TIFF/sidecar. Other pollutants, derived masks, normalization, panels, and AOI files are excluded. The chain is ReadData, PrepareScene(mode=pollution_peak), ComputeStats(metrics={peak_no2_column:true},region=aoi), and TaskComplete.

The source is Sentinel-5P OFFL/L3 tropospheric NO2 column number density, scaled by 1e6 from mol/m² to µmol/m². The worker rejects missing declared dependencies and an empty finite population. It computes NumPy-compatible maximum and percentile statistics and rounds to six decimals.

Independent hash, Rasterio, and NumPy checks reproduce all 20 answers and p95 values exactly. Peaks span 113.555962–412.991425 µmol/m²; every raster is finite and nonempty. Focused execution plus complete evidence support pass 20/20.

Validated contract. Product identity, temporal/spatial meaning, units, finite-data handling, statistics, causal lineage, schema, declared inputs, and all answers are aligned.

Question type 26 of 43 — multi-hazard air pollution `B`

Population: 20 records — 18 train and 2 public-reference test.

Agent task. Classify every finite pixel of the Sentinel-5P episode-peak tropospheric NO2-column raster into the declared training-relative quartile intervals, then return the pixel-count dominant class, fraction, and complete per-class counts, areas, and means.

Agent inference. The four labels locate pixels within this benchmark's training distribution. They are not WHO categories, surface concentrations, health-risk classes, or live pollution assessments.

Benchmark meaning. This tests leakage-safe normalization selection, continuous-raster classification, interval boundaries, dominance/tie logic, geospatial area evidence, and structured synthesis.

The five declared paths are the question, handle inventory, Sentinel-5P TIFF/sidecar, and normalization.json. Other pollutants, derived masks, panels, and AOI files are excluded. The chain is ReadData, PrepareScene(mode=pollution_peak), ScoreHazard(hazard=air_pollution,scale=benchmark_relative_column_quartile), and TaskComplete.

The normalization was fitted only on all 294,912 finite pixels from the 18 training AOIs; neither test AOI contributed. Recomputed quartile boundaries are exactly 88.19622039794922, 121.34806823730469, and 179.93792724609375 µmol/m². Intervals are lower-inclusive and upper-exclusive with numerical coverage extended to ±1e30. Dominance uses pixel count and an exact tie uses q1-to-q4 declaration order. Areas use WGS84 ellipsoidal cells.

Independent Rasterio/NumPy reconstruction reproduces all thresholds and all 20 expected answers. Dominant results are 5 q1, 5 q2, 6 q3, and 4 q4; all 16,384 pixels in every AOI are finite and classified. The worker emits the benchmark-specific train-only fit contract, and complete evidence support passes 20/20.

Validated contract. Training membership, test isolation, thresholds, units, interval coverage, interpretation, dominance, areas, lineage, schema, files, worker evidence, and all answers are aligned.

Question type 27 of 43 — multi-hazard air pollution `C`

Population: 20 records — 18 train and 2 public-reference test.

Agent task. Compute each source's finite-pixel AOI mean for Sentinel-5P NO2 column, CAMS NO2 column, and MODIS AOD; standardize each mean against that source's training-AOI distribution; assign one of three relative states; and report whether all three sources occupy the same state.

Agent inference. Agreement means agreement in within-product relative episode intensity. It does not mean equal physical values, interchangeable pollutants, pixelwise spatial correlation, or health-risk agreement.

Benchmark meaning. This tests heterogeneous-source interpretation, leakage-safe per-source normalization, unit-aware evidence, categorical aggregation, causal tool use, and complete structured synthesis.

The nine declared paths are the question, handle inventory, train-only normalization contract, and three TIFF/sidecar pairs. Panels and unrelated pollutants are excluded. The chain is two ReadData calls, PrepareScene(mode=multisource_hazard), CompareSources(criterion=benchmark_relative_intensity_agreement, required_agreement_fraction=1.0,normalization_fit_split=train), and TaskComplete.

For each source, z is (AOI mean - mean of 18 training AOI means) / population standard deviation of those 18 means. Values below -0.5 are below_training_typical, values above 0.5 are above_training_typical, and the inclusive middle is training_typical. Agreement fraction is the largest state count divided by three; the required 1.0 means all sources must match. Test AOIs never enter the fit.

Independent Rasterio/NumPy reconstruction reproduces all training means and population standard deviations, then every source mean, z-score, state, fraction, and verdict. Four records agree and 16 disagree; fractions are 1.0 for 4, 0.666667 for 7, and 0.333333 for 9. The worker rejects a non-train fit or agreement threshold other than 1.0. Focused execution/support pass 20/20.

Validated contract. Product meanings, differing units, finite-pixel summaries, training-only normalization, state boundaries, decision rule, worker arguments, lineage, schema, files, and all answers are aligned.

Question type 28 of 43 — multi-hazard air pollution `D`

Population: 20 records — 18 train and 2 public-reference test.

Agent task. Use the declared binary top-quartile Sentinel-5P NO2-column footprint, sum co-registered WorldPop people-per-pixel values inside it, and return that population with footprint area and cached hospital/school context.

Agent inference. The answer is population geographically located in a benchmark-relative high-column footprint at the WorldPop 2020 reference epoch. It is not exposed-person causality, surface concentration, health risk, or a live population estimate.

Benchmark meaning. This tests prepared-mask provenance, inclusive threshold boundaries, raster/raster aggregation, vector/raster membership, valid zero-footprint handling, and causal exposure routing.

The seven declared paths are the question, handle inventory, high_no2_column TIFF/sidecar, WorldPop TIFF/metadata, and authoritative cached OSM GeoJSON. The sidecar carries the training threshold provenance, so the runtime consumes the prepared footprint without an undeclared raw-raster read. The chain is two ReadData calls, PrepareScene(mode=pollution_peak_exposure), ScoreHazard(hazard=air_pollution,scale=impact_footprint), EstimateExposure(exposure=high_no2_column_population), and TaskComplete.

Independent reconstruction verifies every mask as finite s5p_no2__peak >= training-pixel q75, where q75 is 179.93792724609375 µmol/m² and the lower q4 boundary is inclusive. WorldPop must match shape, CRS, and affine transform and contain no negative values. OSM membership uses each authoritative geometry's representative point; 5 km context and nearest distance use a local azimuthal-equidistant CRS and positive cell centers.

Twelve footprints are valid measured-zero masks, so their population, area, asset counts, and nearest values are zero/null as defined—not missing output. Across the other eight cases, population reaches 711,292.703 and footprint area reaches 26,228.546 ha. All expected answers reproduce. The exposure worker emits the required train_only_s5p_pixel_q75_v1 provenance field, and binary-mask scoring performs no undeclared normalization read. Focused execution/support pass 20/20; the independent 157-case multi-hazard D geometry audit passes all counts and distances.

Validated contract. Train-only threshold provenance, inclusive boundary, zero-mask semantics, grid alignment, population sum, geometry authority, distances, worker output, lineage, schema, files, and all answers are aligned.

Question type 29 of 43 — xBD damage `XBD-Q1`

Population: 177 records — 160 train and 17 public-reference test.

Agent task. Open the post-disaster scene and its declared xBD building polygons, compute counts by damage class, classified-only damaged and severe percentages, and total severe-building footprint area.

Agent inference. Severe means major-damage or destroyed. Unclassified polygons count toward total_buildings but not percentage denominators; polygon labels and geometry—not visual guessing from the PNG—are authoritative.

Benchmark meaning. This tests causal multimodal routing, categorical aggregation, denominator discipline, polygon-area computation, and exact structured reporting.

The two declared paths are the post-disaster PNG and building_damage_xy.geojson. The chain is two ReadData calls, PrepareScene(mode=xbd_damage), AssessDamage(aggregation=scene), ComputeStats on damage_1, and TaskComplete. ComputeStats consumes the causal damage artifact; it does not reread a hidden target.

The five normalized classes are no-damage, minor-damage, major-damage, destroyed, and unclassified. Damaged count is minor+major+destroyed; severe count is major+destroyed. Both percentages divide by the four-class classified count and return 0 only if that denominator is zero. Severe area is polygon pixel-coordinate area multiplied by the scene GSD squared, summed and rounded to four decimals. Source coordinates are retained for scene-level area; the separate frame-clipping rule belongs only to quadrant tasks.

Independent source-first geometry recomputation matches all 177 answers. Sixty scenes contain unclassified buildings, confirming that the denominator rule is exercised; no scene has zero classified buildings. Severe counts range 0–271, and 56 scenes validly report zero severe area. Focused execution and complete evidence support pass 177/177. The schema bounds counts/area at zero and percentages at 0–100.

Validated contract. Source authority, label normalization, class definitions, denominators, area method, causal handles, schema, files, and every expected answer are aligned.

Question type 30 of 43 — xBD damage `XBD-Q2`

Population: 177 records — 160 train and 17 public-reference test.

Agent task. Rank the four 2×2 frame quadrants by severe polygon area clipped to each quadrant, select the ascending-ID representative of an exact maximum-area tie, and report that quadrant's centroid-assigned severe-building count and clipped severe area.

Agent inference. Area is spatially divisible across boundaries, while a building count is indivisible and belongs to its centroid quadrant. The count and area therefore need not describe the same set of whole polygons.

Benchmark meaning. This tests prepared spatial summaries, object-versus- area aggregation, clipping semantics, deterministic ranking, tie handling, causal artifact use, and exact synthesis.

The two declared paths are the post-disaster PNG and quadrants.geojson. That table is a source-derived cache independently checked against the xBD polygons. The chain is two ReadData calls, PrepareScene(mode=xbd_quadrants,grid=2x2), AssessDamage(aggregation=quadrant_2x2,severity=severe), RankRegions(by=severe_footprint_area_m2,top_k=1), and TaskComplete.

The frame is clipped at x/y 0, 512, and 1024. Severe classes are major and destroyed. Ranking uses unrounded clipped square-metre area; exact ties select the lexicographically ascending Q0–Q3 ID only as the reported representative. In a zero-severe scene all four areas tie at zero, so Q0 with count/area zero is the defined result.

Independent polygon clipping reproduces all 177 targets. It finds 146 severe polygons crossing quadrant boundaries in 53 scenes; using centroid area instead would change the winner in two scenes, confirming that the declared rule is material. Fifty-six zero-severe scenes exercise the all-zero tie. Focused execution and complete evidence support pass 177/177. Schema values are constrained to Q0–Q3 and nonnegative count/area.

Validated contract. Frame/clipping coordinates, severe-class definition, mixed count/area semantics, ranking, exact ties, zero state, causal handles, schema, files, and every answer are aligned.

Question type 31 of 43 — xBD damage `XBD-Q3`

Population: 177 records — 160 train and 17 public-reference test.

Agent task. Rank eligible cells in the prepared 4×4 grid by the exact classified-building severity fraction, return the ascending-ID representative of the exact top tie, all tied IDs, the top three eligible cells, and complete winner/count/status evidence.

Agent inference. The score is (minor + 2×major + 3×destroyed)/(3×classified). Unclassified buildings do not enter the denominator; a cell with no classified buildings has no score and cannot win, even if its cached display score is zero.

Benchmark meaning. This tests formula interpretation, undefined-versus- zero handling, exact rational ranking before display rounding, full tie reporting, top-k behavior, and structured synthesis from a causal artifact.

The two declared paths are the post-disaster PNG and grid_4x4.geojson. The prepared table is independently reconstructed from the polygon source. The chain is two ReadData calls, PrepareScene(mode=xbd_grid,grid=4x4), AssessDamage(aggregation=grid_4x4), RankRegions(by=normalized_damage_severity,top_k=3), and TaskComplete.

Ranking uses exact integer fractions, descending, then ascending cell ID. severity_score is only the six-decimal display value. top_tied_cell_ids contains every exact top tie, while top_k_cells contains the first three ranked eligible cells. A future scene with no eligible cell has null winner and score, an empty top list, and status no_classified_buildings.

Independent source-first recomputation matches all 177 answers, representative winners, tie sets, and top-three rows. Across the scenes, 1,089 cells are excluded for lacking classified buildings, including 30 all-unclassified cells; 90 records exercise an exact top tie, with as many as 16 tied cells. Every current scene still has at least one eligible cell. Focused execution and complete evidence support pass 177/177. Schema bounds preserve both current and defined future branches.

Validated contract. Formula, classified denominator, eligibility, exact ranking, display rounding, ties, top-k, no-eligible branch, causal handles, schema, files, and all answers are aligned.

Question type 32 of 43 — xBD damage `XBD-Q4`

Population: 55 records, each a unique same-event scene pair — 49 train and 6 public-reference test.

Agent task. Compute a complete polygon-derived damage summary for each scene, compare classified-only severe percentages and severe footprint areas separately, and report each metric's unique winner or exact tie set.

Agent inference. Percentage and area answer different questions and may have different winners. A tied metric has every tied scene ID, a null singular winner, and status tie; the agent must not choose an arbitrary representative.

Benchmark meaning. This tests multi-scene handle routing, denominator discipline, two-criterion comparison, exact tie/null semantics, and structured synthesis across causal analytical stages.

The four declared paths are each scene's post-disaster PNG and xBD polygon GeoJSON. All pairs are unique, unordered same-event pairs and remain wholly in one split. The chain is two ReadData calls, PrepareScene(mode=scene_compare), AssessDamage(aggregation=pair_compare,severity=severe), CompareScenes(by=severe_damage_ratio_and_footprint), and TaskComplete.

For each scene, severe percentage is 100 × (major+destroyed) / classified, rounded to two decimals; unclassified polygons remain in totals but not the denominator. A future zero-classified scene emits 0.0 with status no_classified_buildings and is excluded from percentage ranking. Area uses the unrounded sum of severe polygon areas. Comparisons use emitted numeric values and preserve all exact top ties.

Independent polygon/GSD recomputation matches both summaries, both winners, tie lists, and statuses for all 55 pairs. Every current scene is percentage- eligible. Forty-nine pairs have unique winners for each metric; six pairs tie on both and correctly return null winner fields. Focused execution and complete evidence support pass 55/55. Schema bounds cover counts, percentages, areas, and 0–2-item tie lists.

Validated contract. Pair identity/split isolation, class and denominator rules, polygon areas, independent comparison axes, zero- denominator policy, exact ties/nulls, causal handles, schema, files, and all answers are aligned.

Question type 33 of 43 — xBD damage `XBD-Q5`

Population: 121 records — 109 train and 12 public-reference test.

Agent task. Extract exactly the major-damage and destroyed building polygons, then report their count, combined footprint area, separate major/destroyed areas, source, filter, and area method.

Agent inference. “Severe” is the union of two explicit xBD label classes, not a new visual classification. A scene with no qualifying polygon returns a complete zero-valued result, not an empty object or failed extraction.

Benchmark meaning. This tests metadata-bound question interpretation, object filtering, causal handle passage from extraction to assessment, polygon/GSD area aggregation, zero-object handling, and exact synthesis.

The two declared paths are the post-disaster PNG and building_damage_xy.geojson. The chain is two ReadData calls, PrepareScene(mode=xbd_damage), ExtractObjects(target=severe_buildings,severity=major_or_destroyed), AssessDamage(work=objects_1,aggregation=severe_only), and TaskComplete. The final assessment consumes the extraction artifact rather than bypassing the declared causal step.

Each polygon area is its pixel-coordinate geometry area multiplied by the scene GSD squared. Major and destroyed areas are summed independently and their sum is the combined severe area, rounded to four decimals. Counts are whole qualifying polygons; unclassified, no-damage, and minor-damage polygons are excluded.

Independent source-first recomputation matches all 121 counts and all three areas. Thirty-nine records validly contain zero severe polygons and return a complete seven-field result with zero count/areas. Focused execution and full evidence support pass 121/121. Every Q5 prompt derives the disaster type and event from its declared scene metadata.

Validated contract. Prompt/scene identity, severe filter, causal extraction, polygon source, GSD area calculation, per-class sum, zero-object behavior, schema, files, and all answers are aligned.

Question type 34 of 43 — xBD damage `XBD-Q7`

Population: 121 records — 109 train and 12 public-reference test.

Agent task. Count severe-building centroids in each 2×2 quadrant, rank all four quadrants, compute maximum-count concentration, classify its spatial pattern, and return the full distribution and exact maximum tie set.

Agent inference. Concentration is max quadrant severe count / total severe count. When total severe is zero the ratio is undefined, so the score and worst quadrant are null and the top-tie list is empty—not four tied zeros.

Benchmark meaning. This tests causal multi-stage aggregation, spatial count concentration, threshold classification, exact ties, null/zero distinction, and structured answer synthesis.

The two declared paths are the post-disaster PNG and quadrants.geojson. The chain is two ReadData calls, PrepareScene(mode=xbd_quadrants,grid=2x2), AssessDamage(aggregation=quadrant_2x2,severity=major_or_destroyed), RankRegions(by=severe_damage_count,top_k=4), ScoreHazard(work=rank_1,scale=spatial_concentration), and TaskComplete. The scorer consumes the ranked artifact and cannot bypass it with work_1.

For a positive total, scores at least 0.70 are highly concentrated, scores at least 0.45 and below 0.70 are moderately concentrated, and lower scores are spatially dispersed. Exact maximum ties emit every tied ID; the ascending ID is only the representative worst_quadrant, with tie_break_applied=true.

Independent centroid-count reconstruction reproduces all 121 distributions, scores, classes, ties, and representatives. Coverage is 36 highly concentrated, 26 moderately concentrated, 20 dispersed, and 39 no_severe_damage; four scored records exercise a maximum tie. All 39 zero cases have null score/winner and empty tie list. Focused execution and complete evidence support pass 121/121. Schema constraints bind grid, counts, score range, quadrants, and the null branch.

Validated contract. Severe definition, centroid assignment, concentration formula, thresholds, exact ties, zero/null state, causal handles, schema, files, and every answer are aligned.

Question type 35 of 43 — xBD damage `XBD-Q8`

Population: 177 records — 160 train and 17 public-reference test.

Agent task. Produce one comprehensive scene report from the declared xBD building polygons, 2×2 quadrant table, and target mask: damage counts and percentages, footprint areas, severe-damage concentration, and raster/polygon class-presence validation.

Agent inference. Polygon labels are authoritative for building counts and areas. The target mask is an independent rasterized validation view: mask classes 3+4 count severe pixels, not buildings. Therefore the valid cross-source check compares whether each damage class is present, while exact pixel histograms and polygon counts remain separate quantities.

Benchmark meaning. This tests a multimodal four-input join, nested answer synthesis, classified-only denominator policy, polygon/GSD area computation, spatial concentration, raster-code interpretation, and explicit cross-source validation without conflating objects with pixels.

The four declared paths are the post-disaster PNG, building_damage_xy.geojson, quadrants.geojson, and post_target.png. The chain calls ReadData for image_1, damage_labels_1, and mask_layers_1; the last handle contains both quadrant and target-mask paths. It then calls PrepareScene(mode=xbd_scene_summary), AssessDamage(aggregation=scene), and ComputeStats(metrics=damage_summary,target_mask_distribution,severe_footprint_area), before TaskComplete. Every runtime read resolves from those opened handles.

Severe percentage uses classified buildings only. Severe polygon area uses pixel-coordinate geometry multiplied by GSD squared. Concentration uses the largest quadrant severe-centroid count divided by total severe count; its zero-severe branch returns null score, null worst quadrant, and no tied IDs. The 1024×1024 mask permits only codes 0–4, so every stored histogram sums to 1,048,576 pixels. Unclassified polygons are intentionally excluded because the target encoding has no unclassified class. class_presence_consistent means polygon and raster sources agree on whether each class occurs; it does not assert equal object and pixel counts.

Independent reads of the declared assets reproduce all 177 polygon summaries, all 177 mask histograms, and every severe-pixel total; all class-presence checks are consistent. Fifty-six scenes validly exercise the complete zero-severe branch. Focused execution and complete answer evidence pass 177/177, with 708 declared path references present and repository-relative. Capture date is not reported because the question does not ask for it and no declared input is an authoritative capture-date source. The public schema binds methods, encodings, ranges, class names, and null states explicitly.

Validated contract. Input roles, denominators, areas, concentration, zero/null behavior, mask encoding, polygon/raster semantics, schema, paths, and all answers are aligned.

Question type 36 of 43 — routing `nearest_shelter`

Population: 200 records — 180 train and 20 public-reference test.

Agent task. Validate the supplied flood-screened route candidates against the declared road-edge costs, select the minimum fastest-objective candidate, and report its attached shelter, confidence, route ID, distance, ETA, length-weighted flood risk, objective cost, and blocked-edge count.

Agent inference. nearest_shelter is a backward-compatible type name. The agent is not given a new universe of shelters and must not claim a fresh spatial nearest-neighbour or road-graph search. It validates the packaged F candidate group and reports the shelter attached to its computed winner.

Benchmark meaning. This tests precise scope reading, causal use of flood-exposure evidence, independent validation of a cached geospatial product, constrained candidate selection, metadata reporting, and resistance to overclaiming what the inputs prove.

The seven declared paths are Sentinel-2 RGB GeoTIFF/preview, metadata, Copernicus flood extent, OSM-derived flood-risk roads, candidate-route geometries, and route scores. The chain is ReadData, PrepareScene(mode=flood_safe_routing), EstimateExposure(exposure=blocked_roads), then PlanRoutes(objective=nearest_reachable_shelter, avoid.flood_affected_roads=true) consuming exposure_1, followed by TaskComplete.

The worker decomposes each route geometry into its declared road edges and recomputes distance, ETA, node count, blocked-edge count, length-weighted risk, and objective cost. Fastest cost is summed travel_time_min; only candidates with zero blocked edges are eligible, with ETA, distance, and route ID as deterministic tie-breakers. The selected shelter fields must match metadata. Confidence is copied from declared metadata: 187/200 scenes carry the low_fallback_public_facility label. This is a benchmark candidate-validation task, not official evacuation guidance.

Independent source-first audit reproduces all 1,800 route metric sets from 454,968 road edges and verifies the fastest winner, unblocked state, origin, common endpoint, shelter metadata, and expected Type 36 answer in all 200 scenes. Focused execution and complete answer evidence pass 200/200; all 1,400 path references are present and repository-relative. Six argument-sensitivity tests confirm that invalid objectives, unscreened requests, and invalid render layer combinations are rejected. The schema requires a nonempty answer and terminated=true.

Validated contract. Inputs, blocked-road evidence, metric formulas, selection, causal handles, shelter metadata, confidence disclosure, answer synthesis, schema, and paths align.

Question type 37 of 43 — routing `fastest_safest`

Population: 200 records — 180 train and 20 public-reference test.

Agent task. Validate all supplied route candidates, select the minimum-cost candidate within the fastest, safest-weighted, and balanced groups, report all three, and quantify the safest-versus-fastest distance, ETA, and mean-risk trade-off without inventing an evacuation recommendation.

Agent inference. “Safest-weighted” is a named composite objective, not a claim that the selected route has the smallest displayed mean risk. The agent must apply the declared edge-cost formula and distinguish selection cost from the descriptive length-weighted risk statistic.

Benchmark meaning. This tests validation of cached route products, multi-objective reasoning, formula fidelity, comparison-handle dataflow, zero-versus-positive trade-offs, and disciplined reporting under a declared safety scope.

The same seven repository-relative routing inputs are used. The chain is ReadData, PrepareScene(mode=flood_safe_routing), PlanRoutes(objective=candidate_routes, avoid.flood_affected_roads=true), and CompareRoutes consuming routes_1 with fastest, safest, and balanced preferences enabled, then TaskComplete.

For each route geometry, the worker reconstructs source road edges and checks all metrics. Fastest cost is sum(travel_time_min), safest-weighted cost is sum(travel_time_min * (1 + 10*flood_risk)), and balanced cost uses multiplier 4. Eligibility requires zero blocked edges; objective cost, ETA, distance, and route ID define deterministic selection. Deltas are S1 minus F1 for distance and ETA and F1 minus S1 for displayed mean risk.

Independent audit validates every F1/S1/B1 winner and all 1,800 candidate metric sets. S1 is also the minimum raw-risk S candidate in only 27/200 scenes, demonstrating why the composite-objective wording matters. In 144 scenes the rank-one strategies share a geometry and correctly report a zero trade-off; the remaining 56 have larger S1 distance/ETA and lower displayed mean risk. Focused execution and complete answer evidence pass 200/200, with all 1,400 path references present. The worker emits no recommendation, performs no runtime graph search, and rejects partial/invalid preference contracts. The schema requires a nonempty answer and terminated=true.

Validated contract. Candidate validation, objective formulas, eligibility, selections, deltas, zero cases, causal handles, safety wording, schema, files, and all answers align.

Question type 38 of 43 — routing `visual_overlay`

Population: 200 records — 180 train and 20 public-reference test.

Agent task. Validate and compare the packaged route candidates, render a fresh overlay from the declared RGB, road, flood, route, and metadata inputs, then return the image marker and schema-fixed render-validation receipt.

Agent inference. The requested image must be newly derived from the current route_compare_1 handle. A prepared overlay is not an allowed shortcut. A zero visible count for a source layer is valid when the declared feature set has no in-chip member; the agent must not manufacture content.

Benchmark meaning. This tests causal artifact generation, multi-layer CRS alignment, deterministic visualization, image-plus-JSON output, provenance, and the distinction between actually rendering evidence and retrieving a cached final answer.

The same seven routing inputs are declared; no prepared overlay PNG is on the input card. The chain is ReadData, PrepareScene, PlanRoutes(objective=overlay_routes), CompareRoutes over all three objectives, then RenderMap(work=route_compare_1,style=route_overlay) with RGB, roads, flood, blocked roads, origin, shelter, routes, and labels enabled, followed by TaskComplete.

The renderer reads the GeoTIFF CRS and affine transform, converts WGS84 vector coordinates into that CRS, and applies the inverse raster transform to obtain pixels. It does not use a synthetic from_origin(0, ...) transform. It validates the current F1/S1/B1 selections, all nine route geometries and labels, both point markers, nonempty bytes, and changed RGB pixels, and writes a PNG plus JSON manifest inside the current run artifact directory.

The final answer is a structured receipt directly backed by worker fields: <image>, existence, render contract, prepared_overlay_used=false, current handle, selected IDs, requested layers, rendered route/label IDs, origin and shelter flags, 912×512 dimensions, and termination.

Focused execution has complete answer evidence for 200/200. Independent artifact inspection opens all 200 PNGs/manifests with zero errors; byte sizes range from 499,654 to 696,682 and changed chip pixels from 4,452 to 131,737. All 200 image hashes reproduce exactly across separate runs. Five scenes validly render zero in-chip risky roads and 20 render zero blocked roads; all still contain flood pixels, ordinary roads, nine routes/labels, and markers. All 1,400 declared path references exist and remain repository-relative.

Validated contract. Current-handle causality, real georeferencing, declared inputs, layer toggles, rendered content, determinism, artifact storage, structured gold, schema, files, and all outputs align.

Question type 39 of 43 — Sen1Floods11 `A`

Population: 100 records — 90 train and 10 public-reference test.

Agent task. Derive a flood-prediction mask from the declared Sentinel-1 VV/VH raster and report predicted flood percentage and exact-grid area, 8-neighbour component count, largest-component area/centroid, and the complete 4×4 spatial inventory.

Agent inference. These are properties of a deterministic SAR-derived prediction, not claims that every predicted pixel is true flood. Hand labels are evaluator-only and must not be used to create the answer. The valid empty-prediction branch returns zero measurements and no largest centroid; it must never fall back to a label or cached target.

Benchmark meaning. This tests raster loading, finite/nodata handling, deterministic thresholding, generated-artifact lineage, connected components, geodesic measurement, spatial aggregation, and honest prediction semantics.

The sole declared data path is the chip's S1Portable.tif, a finite-nodata derivative of the official Sen1Floods11 two-band Sentinel-1 input. The chain is ReadData, PrepareScene(mode=sar_flood), MapFlood(method=s1_min_backscatter_otsu_v1), ExtractObjects(target=flood_components), DescribeFlood consuming objects_1 with percentage/area/components/4×4 metrics, then TaskComplete.

The predictor uses pixels finite and non-nodata in both VV and VH, calculates min(VV,VH), applies a 256-bin Otsu threshold, and marks scores at or below the threshold as predicted flood. It writes a run-scoped uint8 mask with 0 valid non-flood, 1 predicted flood, and 255 nodata. Components use 8-neighbour connectivity. Area is summed from WGS84 ellipsoidal cell areas; centroids are means of positive pixel-center coordinates. The 4×4 grid reports valid/flood pixels, within-cell percentage, and share of total predicted flood.

Independent source and execution audit covers all 100 rasters: they are 512×512, two-band VV/VH, EPSG:4326, float32, nodata -9999, with no duplicate source hashes and no material train/test spatial overlap. Recomputed 8-neighbour counts match 100/100; ellipsoidal area differs from worker output by at most 0.000493 m² before three-decimal rounding. No prediction in the release is empty; flood coverage spans 0.3638–67.83638% and component counts 31–4,138. Focused execution and complete answer evidence pass 100/100, all 100 input paths exist, and nine invalid/sensitivity tests pass. The agent-facing records declare zero LabelHand, cached-target, or hidden-answer paths.

The separate label-based accuracy audit is construct validation, not part of this task's answer and not an input available to the worker. The public thin release requires users to hydrate the official Sen1Floods11 subset with the checksum-verified installer documented in ATTRIBUTIONS.md.

Validated contract. Input isolation, threshold method, nodata encoding, connectivity, areas, centroids, grid, empty-mask policy, causal handles, schema, files, and all answers align.

Question type 40 of 43 — Sen1Floods11 `B`

Population: 100 records — 90 train and 10 public-reference test.

Agent task. Generate a run-scoped polygon GeoJSON from the SAR-derived prediction and report generation status, polygon count, total and largest component areas, largest bounding box/centroid, and the three most flooded 4×4 cells.

Agent inference. Polygons represent connected regions in the generated prediction, not hand-label objects. Connectivity is 8-neighbour throughout; switching to 4-neighbour would change every current scene. Reported area is the exact-grid cell-area sum associated with the component, not an unqualified planar area computed from longitude/latitude polygon coordinates.

Benchmark meaning. This tests causal raster-to-vector conversion, connectivity semantics, run-scoped artifact creation, geometry validation, largest-object characterization, deterministic spatial ranking, and combined artifact-plus-numeric synthesis.

The only declared data path is S1Portable.tif. After ReadData, PrepareScene, and MapFlood, ExtractObjects(target=flood_polygons) consumes flood_1, writes the GeoJSON, and validates the feature count against the mask. DescribeFlood consumes objects_1; RankRegions consumes flood_report_1 with by=flood_percentage, candidates=grid_4x4, and top_k=3, followed by TaskComplete.

Polygonization uses rasterio feature extraction with connectivity 8 and source georeferencing. Each feature receives a component ID. The largest component is selected by positive pixel count; its bounding box uses outer pixel edges and its centroid averages pixel centers. Grid ranking is descending flood percentage, then flood-pixel count, then descending cell ID. The final key is only a stable serialization rule, and numerical ties remain visible.

Independent artifact audit opens all B and E polygon outputs with zero failures: feature counts match, every coordinate is finite, every geometry is Polygon/MultiPolygon, and all coordinates lie within the source raster bounds. The current Type B feature count ranges from 31 to 4,138 and GeoJSON size from 64,722 to 9,014,842 bytes. All 300 family ranking calls reproduce the expected order; one Type B record exercises an equal percentage/pixel tie within its top three. Four- and eight-neighbour counts differ in all 100 scenes, while the reported counts match eight-neighbour results in 100/100. Focused execution and complete answer evidence pass 100/100, and all 100 declared paths exist.

Validated contract. Prediction lineage, 8-neighbour polygonization, geometry bounds, feature counts, exact-grid area semantics, largest-object fields, ranking/ties, schema, files, and all answers align.

Question type 41 of 43 — Sen1Floods11 `C`

Population: 100 records — 90 train and 10 public-reference test.

Agent task. Derive the predicted mask, rank the three most flooded 4×4 cells, calculate maximum-cell concentration, and classify the prediction as localized or dispersed under the declared threshold.

Agent inference. Concentration is a share of total predicted-flood pixels, not the winning cell's within-cell flood percentage. localized means that a single grid cell holds at least half of all predicted-flood pixels. An empty prediction has concentration zero and is dispersed.

Benchmark meaning. This tests exact denominator selection, deterministic grid aggregation and ranking, causal report-to-ranking dataflow, threshold classification, visible tie handling, and zero-state interpretation.

The sole input is S1Portable.tif. ReadData, PrepareScene with a 4×4 grid check, and MapFlood produce the current mask; DescribeFlood consumes it for grid, concentration, and pattern metrics; RankRegions consumes flood_report_1 with by=flood_percentage, candidates=grid_4x4, and top_k=3; then TaskComplete closes the specialist subgoal and ORC terminates with the published answer reference.

Each deterministic index partition reports within-cell flood percentage and share of total flood. Ranking is descending percentage, then flood pixels, then descending cell ID; the ID key is stable serialization only, and tied numbers remain in the answer. Concentration is max(cell flood pixels) / total predicted flood pixels, with localized at >=0.5 and dispersed below it.

Independent checks reproduce all 100 full-grid totals, top-three lists, concentrations, and classes with zero failures. Scores span 0.069474–0.673538. Ninety-nine scenes are dispersed; the single localized scene belongs to the test split. Thus training does not demonstrate the localized or empty-mask branches, an intentional observed coverage limitation rather than a reason to retune the threshold or split. Focused execution and complete answer evidence pass 100/100, all 100 paths exist, and the worker rejects unsupported grid, metric, candidate, and top_k contracts.

Validated contract. Grid accounting, denominator, ranking/ties, concentration, threshold, zero policy, causal handles, schema, files, and all answers align.

Question type 42 of 43 — Sen1Floods11 `D`

Population: 100 records — 90 train and 10 public-reference test.

Agent task. Generate the SAR prediction, compare its positive pixels with the required JRC known-water raster, and report total predicted flood, overlap and outside-reference counts/percentages, and largest-component area/centroid.

Agent inference. JRC is contextual known-water coverage, not flood ground truth. A predicted pixel outside JRC is outside the known-water reference; it must not automatically be called new or erroneous flood. Both percentages use total predicted-flood pixels as denominator.

Benchmark meaning. This tests a causal two-source join, raster-grid compatibility, boolean intersection, denominator discipline, zero-reference handling, careful semantic wording, and structured multi-source synthesis.

Two required paths are declared: S1Portable.tif and JRCWaterHand.tif. Separate ReadData calls feed PrepareScene(mode=sar_flood_reference), then MapFlood creates the current prediction. CompareSources(work=flood_1,reference=known_water, criterion=overlap_vs_outside_reference) validates and intersects the grids. DescribeFlood consumes compare_1 for prediction, component, overlap, and outside-reference metrics, then TaskComplete closes the specialist subgoal and ORC terminates with the published answer reference.

The worker requires identical array shape, CRS, and affine transform and accepts only JRC values 0/1 (plus declared nodata if present). Overlap is prediction & JRC==1; outside is prediction & JRC!=1. Counts partition the predicted-flood pixels, and their percentages sum to 100 for nonempty masks. A valid all-zero JRC raster is measured as zero known-water overlap, not treated as absent data.

Independent reads reproduce every comparison with zero failures. All 100 grid, CRS, transform, and encoding checks pass. Forty-two references contain known water and yield positive overlap; 58 contain only valid zeros and yield exact zero overlap. Overlap ranges from 0 to 76,295 pixels and 0–79.140461% of the prediction; every overlap/outside count and percentage partitions correctly. Focused execution and complete answer evidence pass 100/100, all 200 required path references exist, and invalid/missing reference contracts are rejected.

Validated contract. Required inputs, grid alignment, encoding, intersections, denominators, zero-reference state, largest-component fields, cautious semantics, causal handles, schema, paths, and all answers align.

Question type 43 of 43 — Sen1Floods11 `E`

Population: 100 records — 90 train and 10 public-reference test.

Agent task. Execute the complete SAR-flood analysis: generate the mask and polygon artifacts, measure components/area/grid, rank cells, compare JRC known-water context, and render a fresh analysis image. The runtime projects the complete worker evidence into the fixed-schema result.

Agent inference. Each step must consume the current preceding handle; the answer is not a static report lookup. The map visualizes the requested SAR, prediction, component-boundary, and grid layers. JRC informs the numeric comparison but is intentionally not drawn because the render request does not enable that layer.

Benchmark meaning. This is the end-to-end capstone for raster processing, raster-to-vector conversion, quantitative/spatial analysis, multi-source comparison, ranking, deterministic visualization, artifact provenance, causal tool use, and large structured-answer synthesis.

The two required paths are S1Portable.tif and JRCWaterHand.tif. The chain calls ReadData for each input, PrepareScene(mode=sar_flood_diagnostic), MapFlood, ExtractObjects(target=flood_polygons), DescribeFlood, RankRegions, CompareSources from rank_1, and RenderMap from compare_1 with style flood_diagnostic; only then does TaskComplete close the specialist subgoal. ORC terminates with the published reference to the runtime-projected 19-field summary. The three generated artifacts are the prediction GeoTIFF, polygon GeoJSON, and PNG/JSON render pair.

The returned summary combines all final contracts from Types A–D: predicted coverage/area, 8-neighbour component and polygon counts, largest geometry, full grid and top three, concentration/pattern, JRC overlap/outside partition, method, and three generation flags. The renderer verifies its current handle, requested/rendered layers, mask hash, component boundaries, changed pixels, nonempty output, and prepared_overlay_used=false.

Persistent independent audit reports zero mask, geometry, grid, ranking, known-water, render-contract, or render-artifact failures. All 100 PNGs and manifests open; images are 512×512, 69,088–636,031 bytes, change 3,714–178,679 pixels from the SAR base, and contain 682–79,742 component-boundary pixels. Both prediction-mask and image SHA-256 values reproduce in 100/100 separate executions. Focused execution and complete answer evidence pass 100/100, all 200 required path references exist, and all nine worker argument/sensitivity tests pass.

Validated contract. Every input, intermediate handle, measurement, polygon, grid/rank, contextual comparison, render, artifact, schema field, path, and expected answer is aligned end to end.

All 43 of 43 question types are defined by the contracts above. The release baseline includes the repository-wide replay, deep asset inspection, derivative consistency checks, and documentation consistency checks.

GeoDisaster 2,921 — final user guide

1. Benchmark purpose

2. Authoritative dataset and evaluation contract

3. How an agent solves a record

4. Data, provenance, and release policy

5. Final validation baseline

6. Dataset construction, split, and answer registry

7. Reproducible runtime and file authorization

8. Shared geospatial and source-data contracts

9. Honest scope and interpretation limits

10. Public-release procedure

Question type 1 of 43 — deforestation forest_loss_detection

Task and coverage

What the agent must do and infer

Declared sources and files

Required causal tool sequence

Calculation and decision contract

Final answer contract

Edge cases and verified result

Question type 2 of 43 — deforestation area_estimation

Task and coverage

What the agent must do and infer

Declared sources and files

Required causal tool sequence

Calculation contract

Final answer contract

Edge cases and verified result

Question type 3 of 43 — deforestation change_window_estimation

Task and coverage

What the agent must do and infer

Declared sources and files

Required causal tool sequence

Calculation and decision contract

Final answer contract

Edge cases and verified result

Question type 4 of 43 — deforestation false_positive_rejection

Task and coverage

What the agent must do and infer

Declared sources and files

Required causal tool sequence

Ordered proxy-rule contract

Final answer contract

Edge cases and verified result

Question type 5 of 43 — deforestation baseline_canopy_assessment

Task and coverage

What the agent must do and infer

Declared source and file

Required causal tool sequence

Calculation and decision contract

Final answer contract

Edge cases and verified result

Question type 6 of 43 — deforestation multi_year_loss_attribution

Task and coverage

What the agent must do and infer

Declared source and file

Required causal tool sequence

Calculation and decision contract

Final answer contract

Edge cases and verified result

Question type 7 of 43 — deforestation cross_modal_concordance

Task and coverage

What the agent must do and infer

Declared sources and files

Required causal tool sequence

Calculation and decision contract

Final answer contract

Edge cases and verified result

Question type 8 of 43 — deforestation severity_proxy_numeric

What the task asks and why

Declared sources and files

Required causal tool sequence

Calculation and decision contract

Final answer contract

Edge cases and verified result

Question type 9 of 43 — multi-hazard wildfire A

What the task asks and why

Declared sources and files

Required causal tool sequence

Calculation contract

Final answer contract

Question type 1 of 43 — deforestation `forest_loss_detection`

Question type 2 of 43 — deforestation `area_estimation`

Question type 3 of 43 — deforestation `change_window_estimation`

Question type 4 of 43 — deforestation `false_positive_rejection`

Question type 5 of 43 — deforestation `baseline_canopy_assessment`

Question type 6 of 43 — deforestation `multi_year_loss_attribution`

Question type 7 of 43 — deforestation `cross_modal_concordance`

Question type 8 of 43 — deforestation `severity_proxy_numeric`

Question type 9 of 43 — multi-hazard wildfire `A`

Question type 10 of 43 — multi-hazard wildfire `B`

Question type 11 of 43 — multi-hazard wildfire `C`

Question type 12 of 43 — multi-hazard wildfire `D`

Question type 13 of 43 — multi-hazard deforestation `A`

Question type 14 of 43 — multi-hazard deforestation `B`

Question type 15 of 43 — multi-hazard deforestation `C`

Question type 16 of 43 — multi-hazard deforestation `D`

Question type 17 of 43 — multi-hazard urban expansion `A`

Question type 18 of 43 — multi-hazard urban expansion `B`

Question type 19 of 43 — multi-hazard urban expansion `C`

Question type 20 of 43 — multi-hazard urban expansion `D`

Question type 21 of 43 — multi-hazard drought `A`

Question type 22 of 43 — multi-hazard drought `B`

Question type 23 of 43 — multi-hazard drought `C`

Question type 24 of 43 — multi-hazard drought `D`

Question type 25 of 43 — multi-hazard air pollution `A`

Question type 26 of 43 — multi-hazard air pollution `B`

Question type 27 of 43 — multi-hazard air pollution `C`

Question type 28 of 43 — multi-hazard air pollution `D`

Question type 29 of 43 — xBD damage `XBD-Q1`

Question type 30 of 43 — xBD damage `XBD-Q2`

Question type 31 of 43 — xBD damage `XBD-Q3`

Question type 32 of 43 — xBD damage `XBD-Q4`

Question type 33 of 43 — xBD damage `XBD-Q5`

Question type 34 of 43 — xBD damage `XBD-Q7`

Question type 35 of 43 — xBD damage `XBD-Q8`

Question type 36 of 43 — routing `nearest_shelter`

Question type 37 of 43 — routing `fastest_safest`

Question type 38 of 43 — routing `visual_overlay`

Question type 39 of 43 — Sen1Floods11 `A`

Question type 40 of 43 — Sen1Floods11 `B`

Question type 41 of 43 — Sen1Floods11 `C`

Question type 42 of 43 — Sen1Floods11 `D`

Question type 43 of 43 — Sen1Floods11 `E`