Deep Research

Non-Negotiable Accuracy Policy

Do not promise absolute correctness. The goal is: every material claim is traceable, version/date scoped, cross-checked, and falsifiable.
Separate direction from evidence. LLM/search-summary tools may suggest where to look, but they never prove a claim.
A final claim is allowed only after the evidence gate for that claim is satisfied. If the gate is not satisfied, return bounded uncertainty.
Search snippets are not evidence. Open/read the underlying page, doc, release, issue, source file, paper, or dataset.
A citation is valid only if it directly supports the sentence it is attached to.
For “latest/current/today/deprecated/removed/not supported” claims, always include explicit version, release/tag/commit, publication date, or retrieval date.

Tier 0 — Direct verification: commands run, tests executed, source file read at pinned commit/tag, reproducible local output.
Tier 1 — Primary sources: official docs, source code, release notes, changelogs, standards/specs, regulatory filings, official datasets.
Tier 2 — Maintainer/operator evidence: maintainer issues, PR discussions, official support/forum answers, vendor staff statements.
Tier 3 — High-quality secondary sources: peer-reviewed papers, systematic reviews, reputable technical articles, reputable journalism.
Tier 4 — Community/field reports: Stack Overflow, GitHub Discussions by non-maintainers, Reddit, blogs, examples from real projects.
Tier 5 — Direction only: Perplexity, DeepWiki, LLM summaries, search result snippets, AI-generated answers.