--- name: deep-web-research description: Exhaustive deep web research — five-move flow, external findings ledger, mechanical saturation check, disconfirmation, condensation from disk. Opt-in skill for the research profile. version: 1.0.0 author: Hermes Agent --- # Deep Web Research Loaded explicitly via `-s deep-web-research`. Not loaded during normal interactive use of the research profile. This skill overrides default behavior to enforce exhaustive research methodology. ## Core Rules (Always in Effect) These persist across all turns — they are in the skill, not in fading context: - **Confined to /tmp.** All file writes go to `/tmp/`. Never write outside /tmp. - **No self-provisioning.** Never install software. No pip, npm, apt, docker, or any package manager. Use only what's already configured. - **No repeat searches.** If you catch yourself searching the same thing twice, stop. That sub-question is saturated. - **Blacklist after 3 failures.** If a URL returns an error 3 times, blacklist it and move on. Do not retry indefinitely. - **Local and free only.** No internet-based paid services, no SaaS APIs with billing, no metered endpoints. Use any tool already configured that fits this rule. ## Architecture: External Findings Ledger At 200 turns, early findings scroll out of context. The ledger is the fix. **Path:** `/tmp/research-.md` **Schema per finding:** ``` ## Finding #: - Sub-Question: - Claim: - Source: — - Confirmed by: (or "single-sourced") - Date: - Confidence: confirmed | likely | single-sourced | disputed - Notes: ``` **Dedup rule:** Before writing a new finding, scan the ledger for an existing finding with the same claim. If found, add the new source to "Confirmed by" instead of creating a duplicate. Duplicate findings defeat the saturation check. **Sub-question tagging:** Every finding must reference which sub-question it belongs to. This enables the phase gate to verify coverage and the condensation phase to group findings by topic. ## Five-Move Research Flow ### Move 1: Decompose Break the question into sub-questions. Write them to the ledger with time classifications. This makes "multiple angles" principled instead of random. ``` ## Sub-Questions 1. [time: week|month|year|none] 2. [time: week|month|year|none] ``` The time classification lives on disk, not in fading context. Move 3 reads it back when applying time filters per sub-question. **Time classification guide:** - **Current-state** (releases, prices, news, versions, events): `week` or `month` - **Established-knowledge** (how something works, architecture, algorithms, history): `none` or `year` ### Move 2: Landscape Pass Shallow-but-broad sweep. For each sub-question, 1-2 searches, read top 1-2 results. Extract only: key sources, terms of art, where disagreement lives, major players. No deep-diving yet. Append landscape notes to the ledger. ### Move 3: Deep-Dive per Sub-Question For each sub-question: - Search from 3+ angles with different categories and framings. Use `mcp_searxng_searxng_web_search`. Apply the time filter from the ledger. - Read full pages (5-10 per sub-question) via `mcp_searxng_web_url_read`. Snippets are pointers, not sources. - Follow citation trails: if a page cites a study/paper/dataset, go read that source. Depth-first — one trail at a time, to its end, then the next. - Extract structured data: tables, numbers, dates, versions, names. - Assign credibility tier at capture time (not post-hoc in output). - Cross-reference: every claim needs 2+ independent sources. - Use browser tools (Playwright) when static extraction fails on JS-heavy pages. - Append every finding to the ledger. Citation trail stop conditions: - Reached primary source (original paper, official docs, raw data) - Dead end (paywall, 404, requires login) - Circular reference (already read this source) ### Move 4: Disconfirmation Pass Actively try to falsify each key claim: - Search for "[claim] wrong", "[claim] criticism", "[claim] outdated" - Check dates: is a 2024 source being presented as current? - Look for contradicting evidence - Append disconfirmation findings to the ledger This is distinct from cross-referencing. Cross-reference confirms agreement. Disconfirmation actively hunts for disagreement. ### Move 5: Condensation Read the full ledger from disk. Read the phase gate file. If any box is unchecked, do NOT condense — go back and complete that item. Synthesize: - Lead with the answer. No process narrative. - Evidence-backed. Every claim tied to a source from the ledger. - Concrete. Numbers, dates, names, specifics. - Uncertainty explicit. "X confirmed by A and B. Y single-sourced from C. Z unresolved — D and E conflict." - Sources section: numbered list with URLs and credibility tiers. ## Guardrails ### Saturation Check (Mechanical) Every 3-4 searches, run this command — do not self-assess: ``` grep -c "^## Finding #" /tmp/research-.md ``` Compare to the previous count. If zero new findings in the last 3 searches, that sub-question is saturated. Move to the next. This is a measurable fact, not a vibe. ### Phase Gate File The phase gate is a file at `/tmp/research--gate.md`. Write it and read it back — it is not a mental checklist. Before moving to Move 5, read the gate file. If any box is unchecked, do NOT condense. Go back and complete that item. ``` # Phase Gate - [ ] All sub-questions have findings in the ledger - [ ] Disconfirmation pass completed for key claims - [ ] OR: saturation condition fired (no new findings in last 3 searches) - [ ] OR: turn ceiling approaching (wrap up what you have) ``` Update this file as you progress. ### Turn Ceiling 200 turns is a safety net, not a target. Hitting the ceiling is a failure to condense, not the normal path. If the ceiling is hit, deliver partial findings with a note on what's missing. ## Tool Policy **One rule: local and free only.** Use any tool already configured on the research profile that fits this rule. No internet-based paid services, no SaaS APIs with billing, no metered endpoints. You determine what fits — you are not given a list of allowed or disallowed tools. **No self-provisioning.** Never install, pull, or spin up new tools at runtime. ### Tools That Would Have Helped At the end of the response (after Sources), add a section if any limitations were hit: ``` Tools that would have improved this research: - : None: all sources were accessible with available tools. ``` This is informational only. Do your best with what's available and note what could have been better. Do not stop or block on tool gaps.