task_charts__heatmap__condition_run_extremum_label

Contract

1. Domain: charts
2. Scene id: heatmap
3. Source implementation: src/trace_tasks/tasks/charts/heatmap/condition_run_extremum_label.py
4. Query id: single
5. Condition kind is a sampled generation parameter; it does not change the public query branch.
6. Condition kinds match only the single extremal color level: highest/lowest intensity, highest/lowest activity, or strongest increase/decrease.

Implementation

1. Registered class: trace_tasks.tasks.charts.heatmap.condition_run_extremum_label.ChartsHeatmapConditionRunExtremumLabelTask
2. Prompt lookup: src/trace_tasks/resources/prompts/charts/heatmap/charts_heatmap_v1.json
3. Generation is deterministic from instance_seed, explicit params, prompt bundle, renderer config, and code versions.
4. Answers and annotation are produced from the same metadata execution trace.

Program Contract

Program: arg_extreme(row_label, longest_run(filter(row_cells, exact_extremal_condition_kind)), direction=largest); output=string_label; annotation=bbox_set(winning_consecutive_run_cells); scene=heatmap; scope=condition_run_extremum_label

Candidate set: the visible heatmap cells with row and column labels inside the condition_run_extremum_label objective scope. Operands: prompt-bound labels, categories, series names, thresholds, intervals, references, and encoded chart values, plus the task's prompt-bound target operands when present. Operation: evaluate arg_extreme over the candidate set using the filters, comparisons, aggregations, rankings, projections, or counterfactual edits named in the program expression; generation enforces a unique final answer. Output binding: answer is the string_label value bound by string_label. Annotation witnesses: bbox_set witnesses bound by bbox_set(winning_consecutive_run_cells). Annotation marks rendered bboxes for the winning consecutive run cells from left to right. Axes, legend, title, and distractor text are context unless the task explicitly asks for them as annotation. Query ids: single.

Reasoning Operations

Families: filtering, ranking

Annotation Contract

1. Answer schema: string_label.
2. Annotation schema: bbox_set.
3. Annotation marks rendered bboxes for the winning consecutive run cells from left to right.
4. Axes, legend, title, and distractor text are context unless the task explicitly asks for them as annotation.

Query Details

Query id	Program signature	Answer schema	Annotation schema
single	arg_extreme.longest_exact_condition_run	string_label	bbox_set