arXiv Decision Intelligence

arXiv Decision Intelligence Report

Research a vertical field or technology through papers, then convert the corpus into updated field beliefs and explicit decisions. Deliver a validated v3 JSON record and a deterministic standalone HTML report. Chinese is the primary editorial language; preserve original English titles and precise technical terms.

The report's value is not paper summaries, section completeness, or evidence traceability. Those are audit infrastructure. The value is a defensible answer to:

1. What should the reader believe about the field now?
2. What has actually been demonstrated, and where is the capability boundary?
3. Which technical mechanisms and options could move that boundary?
4. Is the field undergoing a structural transition or only a cluster of signals?
5. How mature is the field scientifically, technically, operationally, and in governance?
6. Should the decision-maker invest now, validate or productize, in what priority order, what should they not do, and what observation should change that choice?

If the corpus cannot support those answers, say that the evidence is insufficient. Do not fill the report with plausible synthesis to simulate insight.

Required derivation chain

Follow this chain in order. A later object must be derived from earlier objects:

atomic evidence
  -> mechanism and experimental facts
  -> testable field propositions
  -> current knowledge state
  -> capability frontier and readiness gaps
  -> source-supported mechanism or analyst reference architecture
  -> technical options and their comparability
  -> transition theses
  -> five-dimensional maturity
  -> decision consequences
  -> leading indicators and belief-update triggers

In the v3 contract, propositions and their statuses form the current knowledge state. Paper-level analysis.mechanism, analysis.evaluation, and analysis.key_findings preserve the mechanism and experimental facts from which the field model is built. field_thesis records the direct answer and explicit belief updates that follow from those propositions. Do not skip from abstracts directly to strategy prose.

Non-negotiable contract

1. Treat this workflow as arXiv-centered, not arXiv-only. Use the installed literature-search-arxiv skill for discovery and full-text retrieval. Read its complete SKILL.md first and obey its license notification, rate limit, URL-listing, download, and safe-extraction rules.
2. Never issue arXiv requests in parallel. Use the supplied helper and leave at least three seconds between separate invocations.
3. Preserve every raw search response. Normalize it with scripts/normalize_arxiv_stream.py; never hand-rewrite source metadata.
4. Analyze into the exact v3 contract in references/report.schema.json. HTML is a deterministic view of the JSON, never a second analysis pass.
5. Record exact discovery queries, query layer, sort, limit, and date boundary. A list of already-known arXiv IDs is not a reproducible discovery query.
6. Build an atomic evidence ledger. Every proposition and model judgment must resolve through IDs to appropriate source evidence and locators.
7. Distinguish metadata, abstract, full_text, and external evidence. Apply an honest evidence_ceiling; missing depth cannot be repaired by prose.
8. Distinguish shared narrative from independent validation. Several papers repeating the same premise show framing convergence, not that the premise has been independently tested or reproduced.
9. Do not claim peer review, publication, citation impact, influence, replication, adoption, or authority without external academic verification. Record source, URL, access date, and the exact verified fact.
10. Preserve original English titles. title_zh is editorial analysis, not source metadata. Separate author statements, measured results, and analyst inference.
11. List the arXiv URL of every analyzed paper in the report and final response.
12. Frame one primary decision maker, choice, and time horizon. Secondary contexts may receive subordinate consequences but may not become co-equal report goals.

Evidence ceilings and interpretive boundaries

• Latest is not frontier. Frontier requires a demonstrated boundary and a bounded delta from it.
• Shared language is not validation. Independent groups must directly test a compatible proposition under interpretable conditions.
• One paper is not a transition. A single work is a signal regardless of novelty or headline result.
• Submission volume is not prevalence or adoption. Search counts describe this retrieval only.
• A benchmark win is not deployment readiness. Preserve the demonstration context and generalization ceiling.
• Prestige is not evidence. Author, lab, venue, affiliation, citations, or recency never replace direct evidence.
• Current-year-only abstract scans produce signal hypotheses only. They may describe recent submissions, but cannot establish historical change, established propositions, structural transitions, validated maturity, or real-world capability. Mark the missing anchor/full-text/external layers.
• Coverage is not a scientific result. “This search found no deployment evidence” belongs in coverage gaps and the epistemic ceiling, not in a field proposition claiming the technology has no deployments.
• Modules are not automatically alternatives. Memory, gating, feedback, and execution may be complementary stages. Rank only options that address the same choice under genuinely comparable conditions.
• Analyst design is not source mechanism. Label a reference architecture and every design-completion edge visibly; never let a useful synthesis inherit the authority of its source papers.

Read references/analysis-rubric.md completely before screening or synthesis.

Defaults

• Decision context: exactly one primary research agenda, technology strategy, architecture, product, or due-diligence decision; optional secondary contexts remain subordinate.
• Corpus layers: historical anchor, counterevidence, recent frontier from roughly 12–24 months, and emerging signal from roughly 3–6 months.
• Search breadth: 4–8 real query families, normally 5–20 results each; paginate.
• Report corpus: normally 10–25 papers; prefer mechanism and negative-evidence coverage over a target count.
• Evidence: abstracts for discovery; full text for decisive mechanisms, comparisons, results, and limitations; external sources for scholarly status, replication, adoption, and real-world claims.
• Language/output: zh-CN, exact English titles, plus -field-report.{json,html} unless otherwise requested.

Workflow

1. Frame the decision before the search

Populate scope with the operational field boundary and one structured decision_context:

• primary decision type;
• one decision_maker, one choice_at_stake, and one time_horizon;
• optional secondary contexts that remain subordinate to the primary choice;
• current or prior belief that may be updated;
• operational definition and unit of analysis;
• preferred, adjacent, and historical terminology;
• inclusion, exclusion, date, and category boundaries;
• what evidence would change the decision;
• report as_of date.

The schema-valid decision types are research_agenda, technology_strategy, architecture, product, and due_diligence. Listing several secondary uses does not make them co-equal; only the primary choice controls scope and the top answer. Record scope.coverage_gaps rather than burying known omissions in final caveats.

The research question must contain a decision consequence. “有哪些论文” is a search request, not a field-intelligence question.

2. Build a four-layer corpus

Create scope.corpus_layers entries with an id, role, status, dates, purpose, and an honest coverage_note. Use searched, partial, or not_searched for status:

1. Historical anchor — actual earlier papers, surveys, paradigms, benchmarks, or capability records needed to define the prior state.
2. Counterevidence — failures, critiques, negative results, robustness gaps, competing mechanisms, and boundary conditions.
3. Recent frontier — work that may move a capability, mechanism, evaluation, engineering, deployment, or governance boundary.
4. Emerging signal — very recent or isolated work that suggests a hypothesis but is not yet a transition.

Missing layers remain explicit. Never derive a historical baseline solely from recent papers describing their own novelty. A failure paper found incidentally is not a dedicated counterevidence search; if none was run, keep that layer partial or not_searched in the Decision Brief and Evidence Boundary.

Build complementary title/abstract, mechanism, benchmark, application, failure, and replication queries. In scope.queries, preserve the exact query string and reference its layer_id. Distinguish genuine discovery queries from later ID-based metadata retrieval. Enforce date windows both in the query where supported and again against normalized API dates during screening.

3. Search arXiv compliantly

Resolve literature-search-arxiv from the active skill list and run its bundled search helper from that skill directory. Redirect every response to a dedicated run directory.

The dependency requests uv. If no uv skill is exposed, check uv --version. Continue when the executable exists and document the fallback. If it is absent, stop and report the missing prerequisite; do not replace the helper.

uv run "[ARXIV_SKILL_DIR]/scripts/search_arxiv.py" \
  --query "[QUERY]" --max_results 10 \
  --sort_by relevance --sort_order descending \
  > "[RUN_DIR]/raw/query-01.jsonstream"

Normalize each stream:

python3 "[THIS_SKILL_DIR]/scripts/normalize_arxiv_stream.py" \
  --input "[RUN_DIR]/raw/query-01.jsonstream" \
  --output "[RUN_DIR]/normalized/query-01.json" \
  --query-label "Historical anchor — canonical mechanism"

Deduplicate by base arXiv ID, retain the newest retrieved version and all query provenance, and use supplied dates rather than inferring dates from IDs.

4. Screen for field value, not topical resemblance

Screen in two passes:

1. Topical fit — the title and abstract satisfy the operational definition.
2. Model value — the paper supplies an anchor, mechanism, experiment, capability demonstration, counterexample, comparison, or signal needed by the field model.

Assign one primary corpus role: anchor, frontier, signal, or counterevidence. Preserve exclusion reasons for borderline and high-salience papers. Record retrieved, deduplicated, screened, and included counts in scope.screening; record full-text and external coverage in evidence_coverage. Counts, query logs, and layer coverage must reconcile.

5. Enrich papers that control the answer

Retrieve full text whenever a decisive proposition depends on architecture, mechanism, experimental conditions, exact comparisons, figures, tables, author limitations, or claimed novelty. Record section/page/table/figure locators.

Use official venue/publisher records, DOI/Crossref, OpenAlex, Semantic Scholar, official project or replication records, or another primary academic source for external verification. Search snippets are discovery aids, not evidence.

6. Build atomic evidence and paper facts

Create one evidence record per indivisible source fact. Separate author framing, measured result, negative result, replication, derived comparison, analyst inference, and external validation. External evidence requires source_url. Do not combine a mechanism claim and several metrics into one evidence record. Each paper metric must have its own unique evidence_id; never reuse one compound evidence record for two values.

For each paper, populate analysis.research_object, analysis.mechanism, analysis.evaluation, analysis.key_findings, analysis.author_stated_limitations, and analysis.analyst_inferred_limitations. Use model_links to connect the paper to relevant stage_ids, option_ids, capability_ids, and proposition_ids. Preserve abstract_original; keep it out of the main decision narrative. Assign an independence_cluster_id so overlapping labs, author groups, and companion work cannot be miscounted as independent validation.

Reconcile evidence_coverage with the included papers and evidence records.

7. Form testable propositions and the field thesis

Write propositions before conclusions. Every item in propositions must contain:

• statement, proposition_type, status, and scope_conditions;
• supporting and counter evidence IDs plus alternative_explanations;
• evidence_ceiling, confidence, uncertainty, and what_would_change;
• decision_relevance;
• an evidence_profile covering directness, consistency, external_validity, reproducibility, and rationale.

Apply the exact status rules in the rubric. A repeated author claim can be multi_source_aligned while remaining an unvalidated shared narrative. Corpus coverage statements such as “this scan found no deployment paper” are not scientific propositions. Keep them in scope.coverage_gaps, evidence_coverage, report limitations, and the thesis ceiling.

Derive field_thesis from the completed propositions. Populate direct_answer, field_stage, stage_rationale, bottom_line_proposition_ids, belief_updates, decision_relevance, and epistemic_ceiling. Every belief update must state prior_belief, updated_belief, and the proposition IDs that justify the change. Allowed field stages are pre_paradigm, emerging, consolidating, maturing, and deployed. Do not choose a stage from paper volume or recency.

Set epistemic_ceiling.authority, .trend, .capability, and .rationale before writing direct_answer. When the authority ceiling is none, the valid top answer is “当前没有可成立的 authoritative field view”; do not manufacture an authoritative-view section. With a current-year abstract corpus and no searched or partial anchor layer, set trend to signals_only and keep every transition at signal or insufficient_evidence.

8. Establish the capability frontier before explaining mechanisms

Populate capability_frontier.overall_statement and each capability's current_boundary, demonstration_context, strongest_evidence_ids, frontier_level, observed_failures, anticipated_risks, generalization_ceiling, unresolved_requirements, next_boundary, and proposition_ids.

Every observed failure contains statement and nonempty, unique evidence_ids; it must be directly source-bound. anticipated_risks are explicitly analyst inference. Never relabel an unmeasured concern or coverage gap as an observed failure.

Put unachieved desired capabilities in readiness_gaps, not in the demonstrated frontier. Each gap records its identity and labels, current_state, blocking_evidence, what_would_close, and proposition_ids. State when the strongest evidence is benchmark-only and show its exact result, baseline, task, and simultaneous failure in the main narrative.

9. Build a labeled mechanism model and technical options

Only after the capability boundary is explicit, build mechanism_model:

• choose model_type as source_synthesized or analyst_reference and set its epistemic_status;
• system_statement defines the functional boundary;
• stages state purpose, inputs, outputs, current methods, failure modes, and linked propositions;
• every edge records its relation and support_type: source_supported, cross_paper_inference, or analyst_design_completion;
• critical_path_stage_ids and critical_path_note explain which stages are necessary and why;
• structured bottlenecks bind a stage, statement, propositions, and confidence;
• model_limitations state omissions and analyst completion.

An analyst reference architecture must be visibly labeled in JSON and HTML. Do not present analyst-added authorization, rollback, memory, or control relations as source-supported mechanisms.

Populate technical_options, not a generic list of competing approaches. Each option states its technical_bet, relationship, comparison_status, comparison_note, mechanism stages, hypothesis, representative papers, advantages, costs, failures, validity conditions, supporting propositions, and falsifiers. Use alternative only for substitutable choices that address the same stage and objective. Use complementary, baseline, or hybrid honestly. Rank options only when comparison_status is comparable; partially comparable or non-comparable options may be contrasted but never ranked. When the primary decision compares architectures, include material static-rule, end-to-end LLM, human-confirmation, or mixed-initiative baselines as applicable. Mark unevaluated baselines baseline / not_comparable; never invent a direct comparison, result, or paper support.

10. Derive transitions, maturity, decisions, and indicators

A transition_theses item must specify from_state, to_state, the old_constraint, drivers, time window, status, supporting and counter propositions through supporting_proposition_ids and counter_proposition_ids, plural alternative_explanations, falsifiers, linked leading indicators, and confidence. Submission clustering alone yields signal, not a structural transition.

Create exactly one maturity_assessment entry for each dimension:

• scientific_mechanism
• benchmark
• engineering
• deployment
• governance

Each entry requires a level, bounded statement, proposition links, blockers, and upgrade conditions. Never average the five dimensions into a decorative score.

Every decision_consequences item must identify decision_type, audience, the actual decision, a concrete action, because_proposition_ids, conditions, risks, reversibility, and confidence. Generic “值得关注” prose is invalid. The first and strongest consequence must answer the primary decision context; secondary consequences must be visibly subordinate and may not redefine the choice or time horizon. It explicitly states: invest now or not; fund validation or productization; the ordered priorities; and what must not be done in this horizon. A reversible validation investment is not production authorization. When scope.decision_context includes technology_strategy, express its concrete consequences through the schema's research_agenda, architecture, product, or due_diligence decision types rather than inventing a fifth value.

Every leading_indicators item must define a label, observable event or measure, current_baseline, threshold_or_trigger, threshold_basis, threshold_rationale, data_source, observation_window, affected propositions, interpretation if met and missed, and the resulting decision_trigger. Mark a threshold invented for an internal decision rule as analyst_policy; never imply that “two teams” or “four weeks” came from literature unless source_derived is actually supported. “更多研究出现” is not an indicator unless the unit, source, threshold, and belief update are explicit.

11. Validate the v3 research record

python3 "[THIS_SKILL_DIR]/scripts/validate_report_data.py" \
  "[OUTPUT_DIR]/-field-report.json"

Fix every error. Never weaken validation to accommodate incomplete analysis.

Required negative-test classes include:

• duplicate IDs, unknown fields, dangling evidence/model references, duplicate base arXiv IDs, impossible dates, and inconsistent coverage counts;
• missing/ambiguous primary decision context or co-equal secondary contexts;
• a thesis that outruns its propositions or epistemic_ceiling, including a manufactured authoritative view when authority is none;
• a coverage gap encoded as a scientific proposition;
• non-atomic evidence, two metrics sharing one evidence ID, full-text evidence without a locator, or external facts without URL/access date;
• an established proposition supported only by metadata/abstracts or shared narrative;
• false independent validation from overlapping authors/labs or repeated claims;
• a capability level beyond its context, an unmet capability modeled as demonstrated, legacy common_failures, an unreferenced observed failure, or an analyst risk mislabeled as observed;
• an analyst reference model presented as source-synthesized, an unlabeled design-completion edge, or an unsupported bottleneck/critical path;
• complementary modules labeled as alternatives, a non-comparable option ranked, or a technical option without conditions, costs, failures, and falsifiers;
• a transition without a historical state, independent sequence, counterpropositions, alternatives, falsifiers, or linked indicators;
• maturity above its evidence ceiling or without blockers/upgrade conditions;
• a decision without proposition support, conditions, risks, or reversibility;
• an indicator without threshold basis/rationale or an internally invented threshold not marked analyst_policy;
• a current-year abstract fixture without an anchor that is not capped at signals_only, or that claims established/structural/validated/mature, real-world, longitudinal, or independent reproduction.

Run representative mutations from these classes and confirm rejection for the expected reason, not merely a generic parse failure.

12. Render an outcome-first HTML report

python3 "[THIS_SKILL_DIR]/scripts/render_report.py" \
  --input "[OUTPUT_DIR]/-field-report.json" \
  --output "[OUTPUT_DIR]/-field-report.html"

Do not hand-edit generated HTML. The fixed information architecture is:

1. Decision Brief — direct answer, visible authority/trend/capability ceiling, changed beliefs, and the primary decision.
2. Current Field Beliefs — proposition status, scope, uncertainty, and what would change each judgment.
3. Capability Frontier — demonstrated boundary, context, source-bound observed failures, analyst-inferred risks, ceiling, and next boundary.
4. Mechanism Model & Technical Options — source-supported versus analyst-added relations, option relationships, comparability, and actual technical bets.
5. Transition Theses — bounded from/to shifts, alternatives, and falsifiers.
6. Five-Dimensional Maturity — separate scientific, benchmark, engineering, deployment, and governance readiness.
7. Decision Consequences — act, conditions, risks, and reversibility.
8. Leading Indicators — measurable belief and decision update triggers.
9. Evidence Boundary — coverage, missing layers, and material uncertainty.
10. Evidence Appendix — paper dossiers, evidence ledger, exact search method, exclusions, external verification, and all source URLs.

Evidence chips and paper cards must not interrupt the main reasoning. Main-view citations use short stable ordinals such as [E01]; raw evidence IDs belong only in the appendix and link target. Keep key qualifications visible, while detailed dossiers, abstracts, records, and search logs remain in the appendix.

13. Verify artifact and research acceptance

python3 "[THIS_SKILL_DIR]/scripts/verify_report.py" \
  --data "[OUTPUT_DIR]/-field-report.json" \
  --html "[OUTPUT_DIR]/-field-report.html"

Acceptance requires all of the following:

1. JSON validation and semantic cross-reference checks pass.
2. Required negative mutations fail for the intended reasons.
3. Two renders of unchanged JSON have identical hashes.
4. HTML has doctype, UTF-8 charset, viewport, methodology, evidence ceiling, exact queries, external links, and every analyzed arXiv URL.
5. Desktop and narrow layouts are visually inspected; English titles do not overflow, Chinese text does not create awkward single-character wraps, and print remains legible.
6. A cold reader can answer the six value questions, including invest-now, validation-versus-productization, priority, and prohibition, without a dossier.
7. Every decision consequence names the propositions that justify it, and every leading indicator states how it changes a belief and decision.
8. The top answer remains valid when it says no authoritative view or no trend is currently supportable; no empty section is filled to imply otherwise.

Editorial and visual voice

• Calm, restrained, professional, and outcome-first: an Academic Decision Dossier, not a dashboard, newsletter, or promotional page.
• Lead with bounded judgments and decisions, never “本文介绍了” or paper counts.
• Prefer mechanism diagrams, frontier matrices, technical-option tables, and indicator registers only when the structured relationships exist.
• Never manufacture numeric scores, maturity averages, or charts for decoration.
• Use off-white paper, near-black ink, cool-gray rules, and muted steel-blue; no gradients, glossy cards, decorative icons, or SaaS dashboard blue.
• Keep conclusions, scope conditions, uncertainty, and falsifiers visible. Put evidence details and original abstracts in the appendix.

Included resources

• references/report.schema.json and analysis-rubric.md — v3 contract and rules.
• scripts/ — normalization, validation, rendering, and artifact verification.
• assets/report.css and examples/sample-report.json — visual system and v3 example.

Final response

State both skills used and link JSON/HTML; report date, primary decision, coverage, investment/validation priority and prohibition, limitations, and every arXiv URL.