sgo-commission-research

name

sgo-commission-research

description

Conduct source-backed research for an SGO commission, from FINER question shaping through capability-based provider routing, attended Perplexity Pro handoff or credential-gated API execution, Markdown export ingestion, claim-to-source verification, and durable evidence receipts. Use for current-web research, literature discovery, standards or provider-change research, fact checking, public-claim due diligence, systematic reviews, and any commission whose downstream writing or implementation must consume a verified EvidencePacket.

SGO Commission Research

Treat Studium Generale ORGANVM as the research-workflow authority. Treat Limen as an invoking runtime, not as the owner of research methodology. Do not add a research provider to Limen's agent registry.

Adapted from concepts in Cheng-I Wu, Academic Research Skills (CC-BY-NC 4.0), https://github.com/Imbad0202/academic-research-skills. Re-authored in SGO idiom. No verbatim copy.

Read first

• governance/research-backend-profiles.yaml for live profile states, health gates, capabilities, and guardrails.
• schemas/research-request.schema.json, schemas/research-outcome.schema.json, and schemas/research-receipt.schema.json for the external contracts.
• schemas/source-verifier-attestation.schema.json for the machine-readable verifier decision required by ingestion.
• schemas/output-sanitization-attestation.schema.json for the machine-readable export and tracked-output privacy proof.
• standards/SOP--research-backend-routing.md for routing and custody rules.
• standards/SOP--public-private-classification.md for P/O/C/E preservation.
• references/provider-routing.md when selecting or activating a profile.
• The mode-specific reference only when using socratic or systematic-review.

Modes

Mode	Output
full	Research brief, literature matrix, verified EvidencePacket
quick	Bounded thematic overview plus verified material claims
socratic	Five-layer dialogue and refined ResearchRequest
lit-review	Literature review, FINER score, matrix, EvidencePacket
fact-check	Claim-by-claim support, contradiction, and unknown report
review	Source and argument audit of an existing draft
systematic-review	Preregistered PRISMA-style review and risk-of-bias record

Outcome contract

Emit exactly one typed outcome:

• EvidencePacket when all material claims have resolvable supporting citations and the verification gate passes.
• ManualHandoff when an attended surface must be operated by the user, most notably pro_research.
• BlockedReceipt when no healthy profile can satisfy the request within its privacy, latency, or spend ceiling.

Never silently fall back to a more expensive, less private, or less verifiable profile.

Workflow

1. Validate the commission.

   • Create a ResearchRequest; validate it against the canonical schema.
   • Require an owner repository, report path, receipt path, preservation tier, external-transmission authority, and spend ceiling before routing.
   • Use question-architect and methodology-designer to refine scope and apply the FINER gate.

2. Route by capability.

   • Read the tracked profile registry at execution time.
   • Filter using each profile's declared outcome type, verification strength, preservation tiers, transmission authorities, capabilities, state, machine health, execution timeout, and USD cost. Treat null cost as unknown, never zero.
   • Treat provider catalogs as live external state.
   • Use provider Auto when live metadata is insufficient. Never invent, promise, or pin a model identifier.
   • Hash the validated request and observed catalog/profile view for the receipt.

3. Execute or hand off.

   • For pro_research, render a ready-to-submit prompt using templates/perplexity-project-standing-instructions.md and return a ManualHandoff. The attended Project or Space is named Limen Research.
   • Open Perplexity only for the qualifying request. Research mode chooses its own models; do not request a model.
   • Defer subscription, authentication, and provider-surface checks to the ManualHandoff; do not claim they passed during machine routing.
   • Enforce the pro_research 3,600-second attended execution timeout.
   • Require a Markdown session export. Do not schedule the run, connect apps, send messages, make purchases, or perform external writes.
   • Never use private connected sources. Perplexity profiles accept only P/O preservation and public_only or sanitized_only transmission.
   • API profiles remain impossible to execute until their credential, credit, catalog, and spend-policy health checks pass.

4. Ingest the export.

   • Place the raw export according to the request's P/O/C/E preservation tier. Sensitive raw material belongs in its designated private owner, never in a tracked public path.
   • Preserve retrieval dates, URLs, titles, author or publisher, publication date, source locator, and negative searches. Retain required metadata fields with null only when the source does not expose the value.
   • Split synthesis into atomic claims. Mark each as evidence, inference, or unknown; mark material claims explicitly.
   • Inspect the exact Markdown export and tracked derivatives, then emit OutputSanitizationAttestation. Bind the exact export UTF-8 bytes and record every required redaction.

5. Verify sources.

   • Run bibliography-specialist, then source-verifier.
   • Resolve every cited URL and verify that the source supports the linked claim. Prefer primary sources and name contradictions.
   • Emit the exact SourceVerifierAttestation contract from templates/source-verifier-attestation.md; include every claim/source, confirmed metadata, language, and jurisdiction needed by the request.
   • Drop or downgrade unsupported claims. A broken citation, missing retrieval date, or lost source record fails the packet.
   • Use synthesis-agent only after the graded corpus exists.

6. Write durable outputs.

   • Store the report and receipt in the request's owner repository.
   • Keep private prompt bodies, credentials, and sensitive raw exports out of tracked receipts.
   • Record observed provider/model only when the surface exposes them.
   • Validate the receipt schema and its sanitization assertions.
   • Run uv run --with-requirements requirements-validation.txt python3 scripts/validate-research-backend.py.

7. Hand off verified evidence only.

   • Writing and coding agents consume only accepted EvidencePacket outcomes.
   • CCE may archive and evaluate packets but does not replace owner-repo custody.
   • A rejected packet yields a BlockedReceipt or a revised request, not a prose claim of completion.

Roles

Role	Responsibility
question-architect	Refine and FINER-score the commission
methodology-designer	Match method, evidence, and privacy constraints
bibliography-specialist	Search, log, normalize, and preserve sources
source-verifier	Verify claim support, source quality, and citations
synthesis-agent	Build a claim-linked synthesis and surface tensions
socratic-mentor	Clarify exploratory commissions without answering them
risk-of-bias-assessor	Apply RoB 2, ROBINS-I, or AMSTAR 2 when required

Required templates

• templates/research-brief.md
• templates/literature-matrix.md
• templates/preregistration.md
• templates/evidence-packet.md
• templates/manual-handoff.md
• templates/blocked-receipt.md
• templates/research-receipt.md
• templates/source-verifier-attestation.md
• templates/output-sanitization-attestation.md
• templates/perplexity-project-standing-instructions.md

Terminal predicates

A commission run is accepted only when:

• the request, SourceVerifierAttestation, OutputSanitizationAttestation, outcome, and receipt validate against their schemas;
• every material claim resolves to at least one supporting source;
• every source retains required metadata, a retrieval date, and a source-manifest entry;
• preservation, transmission, and custody match the request and selected profile;
• variable spend does not exceed the request ceiling;
• the report and receipt exist in the owner repository.