Research

Research

Retrieved information takes precedence over pretrained knowledge. Use pretrained knowledge for orientation only. Findings must be grounded in retrieved and validated evidence.

Source Requirements

Sources MUST be created/updated within the preceding 2 years relative to the present date.

• Preferred Sources: Peer-reviewed papers, academic publications, official documentation, standards bodies, industry-leading engineering/research orgs, industry leader technical blogs, official white papers.
• Rejected Sources: SEO/content-farm articles, unattributed summaries, low-quality aggregators, unverifiable secondary claims, sources outside the 2-year window.
• Preserved Metadata: Title, author/publisher/org, publication/update date, URL/DOI/identifier, relevance to question.

Mandatory Research Pipeline

1. Gather Evidence

Collect relevant credible sources before producing findings (never search merely to support a pre-formed conclusion). Ensure coverage exposes disagreements, limitations, and competing approaches.

2. Extract Findings

For every factual finding or inference: (1) Identify supporting source -> (2) Capture exact wording -> (3) Quote verbatim (never alter wording or reconstruct from memory) -> (4) Distinguish Source claim (directly asserted) vs. Inference (derived reasoning; retain exact quoted evidence used) -> (5) Do not strengthen claims beyond source support.

3. Mandatory Adversarial Review

Spawn a subagent to adversarially review findings for unsupported claims, weak evidence, stale sources, quotation mismatches, contradictions, overgeneralization, inference leaps, missing caveats, misinterpretations, or conflicting evidence.

• Cycle: findings -> adversarial review -> revision -> adversarial review... until reviewer declares findings acceptable (self-review does NOT replace this).

4. Validate Research

Validate that quotations exist verbatim in cited sources using the local validator script:

python .agents/skills/research/scripts/validate_quote.py --file <path_to_source> --quote "<exact wording>"
# Or search across a directory of saved sources:
python .agents/skills/research/scripts/validate_quote.py --dir <sources_folder> --quote "<exact wording>"

Verify: quotations exist verbatim in cited sources, source supports claim, and inferences do not exceed evidence. If unvalidated: treat as unverified -> revise/remove/replace via retrieved evidence -> re-run adversarial review -> re-validate. Never present unvalidated quotations.

5. Produce the Survey

Create a survey document directly answering the user's question, citing every factual claim, preserving exact quotations, distinguishing claims from inference, including source metadata, exposing disagreements/limitations/uncertainties without silent reconciliation. Present survey document + brief summary in standard output.

Evidence Conflict

When retrieved evidence conflicts with pretrained knowledge: prefer retrieved/validated evidence, state disagreement explicitly, avoid silent resolutions, and never override retrieved evidence without supporting evidence.

Completion Check

Before presenting, verify:

• Sources within 2-year recency and sufficiently authoritative.
• Every factual finding/inference backed by retrieved verbatim quotes and distinguished from inferences.
• Adversarial subagent reviewed and approved findings across required revision cycles.
• validate_quote.py script verified all quotations against source texts.
• Every claim cited; source disagreements and uncertainties preserved; user received survey and summary.