Research Synthesis

Research Synthesis

Turn a live question into a bounded, recheckable evidence brief. Optimize for a defensible answer and next inquiry, not source volume or a research diary.

Preserve scope and evidence integrity

• Research and synthesize evidence by default. State support and bounded implications. Recommend only the next research probe; route downstream judgments to their owners. Do not approve, reject, or assign their decision states.
• Do not contact people, enroll participants, publish externally, purchase access, change systems, or make consequential decisions without explicit authorization.
• State question, decision relevance, audience, horizon, source boundary, error cost, and stop condition. Keep inquiry open when a hypothesis would create false closure.
• Classify the missing artifact first. When a neighboring owner fits, route there rather than wrapping its work in RES-* claims or a matrix.
• Distinguish source evidence, source-reported interpretation, synthesizer inference, proposed conclusion, and unresolved question. Do not let fluent prose erase those ownership layers.
• Preserve permissions, confidentiality, personal data, embargoes, and licensing. Tool access does not authorize disclosure.
• Existing summaries and concepts are routing and overlap evidence, not automatic substitutes for sources.
• Keep output inline unless the user requests durability or the repository has a research-note convention.

Choose execution depth

Use the compact path for an inline answer from a small cross-source set when reuse, reconstruction, or specialist handoff does not justify a record. State scope, cite material claims, preserve contradiction and uncertainty, explain applicability limits, and stop. Do not create RES-* keys, a working model, matrix, or durable record.

Use the full workflow when consequence, conflict, reusable claims, reconstruction cost, durability, or cross-skill handoff justifies it. Key only claims another artifact must recheck or consume.

If the request concerns one artifact without a broader evidence question, stop this workflow. Use a repository-native capture or summarization workflow when one exists; otherwise summarize or extract the artifact directly without cross-source machinery.

Full research workflow

1. Frame the inquiry. State the question and intended decision or understanding. Bound populations, systems, jurisdictions, versions, periods, sources, excluded interpretations, depth, and stop or inconclusive conditions.
2. Recover authorized coverage. Search supplied sources and local repositories, knowledge bases, decisions, or authoritative artifacts only within the authorized source boundary. Skip connected private corpora not explicitly in scope. Record reusable, stale, and derived material plus gaps.
3. Expose the working model. State a provisional answer, competing hypotheses, or question tree. Search next where evidence could confirm, revise, distinguish, or defeat the weakest consequential claim. Do not turn exploration into advocacy.
4. Design the source strategy. Match official specifications to normative behavior, original research and data to empirical results, systematic syntheses to field evidence, and labeled first-party reports to practice. Set inclusion, exclusion, recency, canonicality, and counterevidence criteria.
5. Verify canonical sources. Prefer the claim owner. Check identity, authorship, date, version, corrections, retractions, and supersession. Read enough to recover method, context, qualifications, and negative results; titles, snippets, abstracts, first pages, and secondary quotes are insufficient.
6. Extract claims with provenance. Give reusable or handed-off claims stable RES-* keys and labels. Link exact passages, figures, tables, data, code, or observations with enough context to recheck. Separate source report from synthesis.
7. Compare equivalent dimensions. Normalize claim, method, population or workload, setting, intervention or mechanism, outcome, window, assumptions, limitations, and transfer boundary. Explain disagreement from conflicting results, different questions, incomparable contexts, versions, or incentives. Compare provenance, evidence overlap, source incentives, and information uniqueness. Correlated repetition is not independent support. Preserve a pivotal minority claim for focused verification when the majority lacked access to its evidence; dissent is neither proof nor a reason for automatic dismissal. Source counts do not settle weight.
8. Synthesize for the declared use. Lead with the bounded answer. Show convergent evidence, material counterevidence, plausible alternative explanations, source-class limits, source-supported applicability, and confidence reasoning. Preserve uncertainty instead of averaging it into a false consensus or local applicability verdict.
9. Stop deliberately. Stop when the declared consequence level is met, decisive uncertainty is explicit, or another source has low expected value. Return inconclusive, insufficient access, or route when evidence cannot support the claim. Name a next probe only if it could change the result.
10. Persist only reusable research state. When a durable artifact is justified, follow the repository convention; keep raw sources or stable source identities separate from derived claims, record freshness, and retain open questions or failed searches that would prevent repeated work. Do not silently rewrite source-owned material.

Use references/evidence-synthesis-record.md only for the full workflow or a justified durable brief.

Route neighboring work

• architecture-surface-mapping owns software orientation; repository inspection is not research synthesis.
• product-opportunity-discovery owns customer context and product hypotheses; literature cannot replace customer evidence.
• software-failure-diagnosis owns causal bug and performance investigation.
• architecture-risk-evaluation owns architecture verdicts; synthesis may supply external evidence.
• software-verification owns project-specific engineering claims, methods, fixed oracles, execution, and adjudication.
• domain-modeling owns local meaning, behavior, rules, invariants, and translations. External definitions are evidence, not authority for the local model.
• Legal, compliance, privacy, security, safety, medical, and financial owners retain context-specific applicability judgments; do not issue compliant, noncompliant, safe, unsafe, eligible, or ineligible verdicts.
• decision-facilitation owns owner criteria, downstream recommendations, accountable tradeoffs, and closure.
• technical-growth-coaching owns engineering mental-model explanation, misconception repair, reconstruction, and transfer checks. It may consume supplied or synthesized evidence; teaching does not establish source truth.
• technical-writing owns a durable reader artifact and its reading path. When cross-source evidence is needed, research-synthesis supplies the evidence foundation; it does not own the tutorial, article, or documentation structure.
• For one artifact, use repository capture or summarization when available; otherwise summarize or extract it directly.

Quality gates

• Question, relevance, scope, freshness, and stop condition are explicit; authorized local coverage was checked when applicable.
• Execution depth is proportionate; compact answers omit research-record machinery, and full records exist only for consequence, reuse, reconstruction, or handoff.
• Source classes fit claims and disclose incentives or independence limits.
• Consequential conclusions trace to exact evidence and remain distinct from inference.
• Counterevidence, negative results, newer versions, and different contexts were sought.
• Empirical findings retain method, population, outcome, and transfer boundary; normative documents are not outcome evidence.
• Contradictions and access gaps remain visible; confidence avoids invented precision.
• The answer precedes source inventory and process.

Reject weak synthesis

• Do not return bibliographies or summary piles as conclusions, count sources as votes, treat repeated citations as independent, or equate vendor guidance, cases, benchmarks, preprints, standards, and independent studies.
• Consequential claims need passage-, figure-, table-, data-, or version-level support.
• A bounded search cannot alone establish absence, consensus, causality, safety, or generality.
• Breadth, polish, or durable artifacts cannot hide weak or inconclusive evidence.

Completion

For the compact path, return the inquiry and scope, bounded answer, exact citations for material claims, contradictions and gaps, applicability limits, confidence, and stop or route. For the full workflow, add keyed reusable claims, source limitations, stop rationale, and any next probe. State whether the evidence is sufficient for its declared use. Do not label a downstream decision approved, rejected, compliant, noncompliant, safe, or unsafe.