Conducting Iterative Research Audits

Given a prior report (or a fresh topic with no prior report), run this single pass:

1. Extract every "unverified," "unknown," "TODO," or "Previously flagged" item from the prior report into a gap list.
2. Write 5-10 new search queries, each targeting exactly one gap — never re-run a query phrased similarly to one already in the report's source list.
3. Execute searches, prioritizing: primary sources > maintained awesome-lists/directories > recent (last 12mo) authoritative writeups > forums/Q&A as corroboration only.
4. For each finding, tag it with the gap ID it closes, a confidence label, and a dated source.
5. Output using the fixed template (below) — delta only, no restatement of prior content.
6. If fewer than 2 substantive new findings surface, say so explicitly and recommend stopping iteration.

No preamble, no summary of "what I did" — go straight to the audit.

Progress checklist for each pass:

• Step 1: Audit prior report for gaps (unverified claims, missing sub-topics, missing expert lenses, uncited domains)
• Step 2: Convert each gap into 1+ targeted, non-overlapping search queries
• Step 3: Execute multi-angle searches (direct, adjacent-domain, counter-narrative, retraction/correction check)
• Step 4: Cross-check currency — discard anything deprecated, superseded, or outdated as of the stated current date
• Step 5: Verify each candidate finding against ≥2 independent sources where possible; label single-source claims as "unverified"
• Step 6: Draft New Findings grouped by theme, each mapped to the gap it closes
• Step 7: Draft Updated/Corrected Claims — explicitly state what prior claim is superseded and why
• Step 8: Draft Remaining Gaps — be honest if yield is low
• Step 9: Compile New Sources list (numbered, dated, linked, excluding anything already cited)
• Step 10: Final self-check pass for sycophancy/hallucination before output (see Verification below)

Step 1 — Gap Extraction

Scan the prior report for:

• Explicit markers: "unverified," "unknown," "TBD," "not yet confirmed"
• A "Previously" / "Remaining Gaps" section from the last pass
• Sub-topics named but not expanded (one-line mentions)
• Single-perspective coverage (e.g., only vendor docs, no practitioner/critic view)
• Domains adjacent to the topic that were never touched
• Sources older than the acceptable freshness window for that domain

Produce a numbered gap list (G1, G2, G3...) before searching anything.

Step 2 — Query Design

Rules:

• One query ≠ one gap is fine, but never reuse a query signature from the prior pass (check prior report's implied search terms if listed, or infer from citation phrasing).
• For each gap, generate at least: one direct query, one adjacent/alternate-terminology query, one "criticism OR limitations OR deprecated" query, one "site:github.com awesome" or curated-list query if the domain has active maintained lists, one "2026" or current-year-filtered query for recency.
• Explicitly search for retractions/updates to previously cited sources: "[prior source name]" update OR correction OR deprecated OR retracted.

Step 3 — Multi-Angle Execution

Run searches in this order per gap, stopping early only if saturation is reached (3+ consistent independent sources):

1. Direct factual query
2. Counter-narrative / skeptic query
3. Primary source / official docs / changelog query
4. Community consensus query (awesome-lists, top-starred repos, current directories, recent high-signal forum threads)
5. Currency check query (filter explicitly for current year / "latest")

Step 4 — Currency Filter

Discard or flag as "historical context only":

• Anything superseded by a newer version, standard, or finding
• Deprecated tools/APIs/methods unless needed to explain why they were replaced
• Stats/benchmarks older than the domain's natural refresh cycle (e.g., 6mo for fast-moving tech, 2-3yrs for slower domains)

Step 5 — Verification (anti-hallucination, anti-sycophancy gate)

Before writing any claim:

• Does it trace to a specific, nameable source? If not → label "unverified" or drop it.
• Would this claim survive if the user disagreed with it? If a claim only exists because it sounds agreeable or expected, cut it.
• Is this common knowledge dressed up as a finding? If yes, cut it — it's fluff.
• Does this contradict the prior report? If yes, it goes in "Updated or Corrected Claims," not "New Findings."

Example 1: Input: Prior report on "vector database selection" flagged G3: "pricing at scale — unverified" and G5: "no mention of on-device/embedded vector stores." Output:

### New Findings
**Theme: Pricing at Scale** (addresses: G3)
- Managed vector DB cost per 10M vectors with metadata filtering varies 4x between providers primarily due to storage-tier vs compute-tier billing models, not raw vector count. [S1] — Confidence: Medium

**Theme: Embedded/On-Device Vector Stores** (addresses: G5)
- SQLite-based vector extensions now support HNSW indexing in-process, eliminating the network hop for sub-1M-vector use cases. [S2] — Confidence: High

### Remaining Gaps
- G3 partially resolved; enterprise contract pricing still opaque (unverified).

Example 2: Input: Re-running the same prompt on a report that was already exhaustively covered two passes ago. Output:

### New Findings
None met the 2-source verification bar this pass.

### Remaining Gaps
No substantive new findings (0 passed verification). Recommend stopping iteration on this topic; revisit only if a triggering event occurs (new major release, regulatory change, etc.)