gaia-compact-evidence-router-skill

name	gaia-compact-evidence-router-skill
description	Compact attachment-first router for GAIA-style tasks under a static ReAct loop with narrow evidence gathering, lightweight derivation, and strict answer-shape control.
allowed-tools	["list_dir","read_file","read_json_file","extract_pdf_text","read_table","image_metadata","audio_transcribe","ocr_image","image_qa","parse_docx","parse_pptx","extract_archive","web_search","fetch_url","html_extract","run_python"]
metadata	{"version":"boot-al-compact-evidence-router","split_source":"train","gold_input_quality":"constructed_trajectory_weak_route_signal","graph_policy":"internal_route_graph_global_static","skill_injection_mode":"global_static","stream":false,"max_executor_steps":24,"evidence_scope":"train_only","repeat_scope":"boot_from_train_gold","learning_scope_after_boot":"train_r1_only","rl_candidate_id":"X4","rl_candidate_slug":"answer-shape-compiler","external_helper_models":"forbidden","no_instance_memory":true}

Use Contract

• Start by reading task.json; if the workspace view is unclear, use list_dir once and then continue locally.
• Treat task wording, local attachments, and extracted local text as the first evidence layer.
• Browser automation is unavailable. For public web evidence, use web_search to locate, fetch_url to retrieve, and html_extract to isolate usable text.
• Use web only when the task truly needs external public facts, historical page states, source verification, or cited public records that local files do not contain.
• When media exists, stay local first: image_metadata, ocr_image, image_qa, or audio_transcribe.
• Use run_python for arithmetic, parsing, decoding, filtering, sorting, joins, unit conversion, coordinate math, and answer-shape checks.
• Each tool call should do one decisive job: locate a source, extract a field, compute a value, or verify final form.
• Prompt Contract Compiler: before routing or answering, bind four scratch fields—scope (entity/source/snapshot/inclusions/exclusions), evidence (minimal fields to extract), op (lookup/filter/compare/count/sort/diff/transform), and surface (required answer form). Use scope to lock the source and boundary, evidence plus op to decide whether you need one field or a full set, history, table, or paired snapshots before computing, and surface to shape the final string. If any decisive field is unbound, the next tool call must bind it; if all four are already clear from the prompt, answer or route directly.
• Final output must be a non-empty <ANSWER>...</ANSWER> and match the requested units, ordering, separators, casing, rounding, and cardinality.
• Before final, compile the answer surface from four fields when relevant: value, unit or label, order, and delimiter.

Training Signals

• Gold trajectory evidence here is weak and mostly constructed. Use it as distribution pressure only: attachment-first routing is common, short derivations are common, and exact answer shape is often decisive.
• Reusable path patterns are concentrated around local tables, local text extraction from documents or archives, OCR or audio transcription, narrow public-source verification, and lightweight Python synthesis.
• Common failure guards recur: wrong scope date, wrong subset, wrong unit, wrong sort order, wrong surface form copied from a nearby field, or an answer that omits a requested sign, separator, or format.
• Default bias: pick the smallest evidence chain that can close the task inside the step budget.

Route Table

Signal	Primary route	Recommended tools	Stop when
Spreadsheet, roster, workbook, matrix, inventory	parse then normalize	read_table, run_python	the target rows and requested field or aggregate are isolated
PDF, text file, rich document, archive, citation bundle	expand then extract	extract_archive, extract_pdf_text, read_file, parse_docx, parse_pptx, run_python	the decisive passage or table is in local text
Image, scan, screenshot, board, label, diagram, music sheet	observe then structure	image_metadata, ocr_image, image_qa, run_python	visible content becomes discrete facts, text, counts, or positions
Audio, spoken instructions, narrated clue	transcribe then normalize	audio_transcribe, run_python	the relevant spoken tokens or numbers are recovered cleanly
Pure logic, arithmetic, decoding, geometry, comparison, ordering	derive locally	run_python, read_file, read_json_file	the exact output value or set is computed from available facts
Public page, official record, archive snapshot, cited article	narrow lookup then verify	web_search, fetch_url, html_extract, run_python	a source-locked passage covers the exact entity, date, and scope

Phase: INIT

First action: read task.json. Route immediately from the question plus attachment suffixes to the shortest decisive local path.

• Route selection: choose local extraction if any attachment or embedded data appears; choose direct derivation if the prompt itself contains all facts; choose web fallback only when external public evidence is explicitly required.
• Recommended tools: read_json_file, then list_dir only if needed.
• Stop condition: you know the answer shape, the likely evidence location, and the next single decisive tool call.
• Failure guard: do not begin with web_search before confirming whether local files already contain the target evidence.
• Final: if the question already provides enough information and format is explicit, answer immediately.

Phase: LOCAL_LOCK

Classify all local materials before branching deeper.

• Route selection: use when attachments, archives, or local reference files are present; map each item to table, text, archive, image, audio, or mixed media.
• Recommended tools: list_dir, read_json_file, read_file, image_metadata, extract_archive.
• Stop condition: each relevant local artifact has a concrete extraction path and irrelevant files are ignored.
• Failure guard: filenames can mislead; rely on actual suffix, visible structure, and extracted contents rather than assumptions.
• Final: if a local artifact already exposes the exact requested field or string, answer immediately.

Phase: TEXT_TABLE_PATH

Turn local structured material into queryable facts.

• Route selection: use for spreadsheets, PDFs, docx, pptx, txt, extracted archives, rosters, qualifications, inventories, captions, and reference lists.
• Recommended tools: read_table, extract_pdf_text, parse_docx, parse_pptx, read_file, extract_archive, run_python.
• Stop condition: the target subset, cited passage, row group, or candidate field list is extracted into clean text or rows.
• Failure guard: watch header shifts, merged cells, duplicate labels, row-order traps, and requests that demand the source surface form exactly as written.
• Final: if the extracted rows or passage directly answer the question and the requested presentation is clear, answer immediately.

Phase: MEDIA_PATH

Convert local media into discrete evidence without leaving the workspace.

• Route selection: use when the task depends on visible text, object identity, board state, layout, note letters, labels, or spoken content.
• Recommended tools: image_metadata, ocr_image, image_qa, audio_transcribe, run_python.
• Stop condition: the media yields stable tokens, numbers, entities, positions, or transcript snippets sufficient for a direct answer or a short derivation.
• Failure guard: separate observation from inference; if OCR and image or audio interpretation disagree on a decisive token, spend one step resolving the conflict locally.
• Final: if the recovered media facts already match the requested answer form, answer immediately.

Phase: WEB_VERIFY

Use narrow public-source lookup only when the task needs evidence outside the workspace.

• Route selection: use for public facts, official databases, historical page states, cited papers, public logs, external source chains, or verification of a locally identified entity.
• Recommended tools: web_search, fetch_url, html_extract, run_python.
• Stop condition: one or two retrieved passages pin down the exact entity, date, edition, revision, or scope required by the prompt.
• Failure guard: search with tight anchors from the question, including date or source type when available. If results are noisy, reformulate once with narrower anchors and keep only source-supported claims.
• Final: if a retrieved passage cleanly supplies the answer in the requested scope and form, answer immediately.

Phase: DERIVE_CHECK

Compute, compare, intersect, and format exactly.

• Route selection: use when the answer requires arithmetic, filtering, chronology, set operations, sorting, ratio comparison, route counting, decoding, unit conversion, or answer-shape normalization.
• Recommended tools: run_python plus the extractor that produced raw facts.
• Stop condition: the requested value, string, or ordered list is reproduced from extracted evidence and survives a cheap consistency check.
• Failure guard: verify variable identity, endpoint inclusion, threshold direction, unit scale, rounding precision, alphabetical or source order, separator choice, and any requirement to keep or remove articles, punctuation, or signs.
• Answer-shape compiler: for transformed outputs, produce the value in scratch text first, then rewrite only the surface form to match prompt constraints.
• Final: if the computed result already matches the required surface form, answer immediately.

Answer Guard

• You may answer from any phase once the evidence chain supports the exact requested form.
• Keep the final payload minimal: only the requested number, string, list, date, or phrase inside <ANSWER>...</ANSWER>.
• If the prompt asks for a bare answer, omit explanations, source names, units, and labels unless they are part of the requested surface.
• Preserve source spelling and casing when asked for text as written; normalize only when the prompt explicitly asks for transformed case, ordering, singularization, or punctuation removal.
• For numeric answers, confirm unit, sign, decimal places, thousands separators, rounding rule, and whether the task wants a bare number or a labeled value.
• For lists, confirm item count, deduplication, ordering rule, and separator formatting.
• For date or chronology tasks, confirm the referenced snapshot, revision, census year, publication date, or archive date before finalizing.
• If two candidate answers remain, use one more discriminating tool call; guessing burns the step budget and damages answer precision.