Extract the article text and metadata first:

Title, date, author if available, publication/source, outbound links, and the exact claims being made.
If web_extract times out or returns empty, do not keep retrying the same path. Switch to browser automation or terminal HTTP requests with a normal User-Agent, then parse with BeautifulSoup/readability-style extraction.

Break the article into checkable claims:

Timeline/date claims.
Quantitative claims: counts, stars, funding, releases, benchmark scores.
Attribution/provenance claims: who said what, where, when.
Technical claims: architecture, features, implementation details.

Verify against primary or near-primary sources before relying on commentary:

GitHub API for repo creation dates, releases, tags, issues, licenses, stars/forks, and comment history.
Official docs, release notes, READMEs, technical reports, company posts.
Raw files from GitHub for current README/license/config evidence.
Search results only as discovery, not final proof.

Compare source quality:

Distinguish first-party accusation/defense, neutral reporting, reposts/aggregators, and low-provenance blogs.
Flag articles with no byline, few/no outbound citations, sensational wording, or copied/reframed claims.

Produce a concise verification table:

Claim.
Status: Supported / Partially supported / Unsupported / Misleading / Unverifiable.
Evidence.
Caveat.

Finish with a bottom-line verdict:

What is solidly verified.
What is plausible but not proven.
What the article overstates.
What additional evidence would change the conclusion.

When the user is testing whether a person is racist / antisemitic / extremist, rank evidence by strength instead of presenting a flat list:

Strongest: direct first-person statements, slurs, endorsements of core conspiracy claims, explicit refusals or exclusions aimed at the group, or primary-source recordings/posts.
Medium: repeated use of coded language or dog whistles that reputable sources trace to extremist subcultures.
Weakest: guilt-by-association, platform-sharing, audience overlap, or commentary about what supporters believe.
In the answer, lead with the strongest verified item first. If the best evidence is only medium-strength, say that plainly instead of overselling.

Model/product/mode disambiguation

When fact-checking AI announcements or community claims, separate four layers before assigning a verdict:

Model family and generation (for example, GPT-5.6).
Capability tier/model variant (for example, Sol, Terra, Luna).
Product or subscription access tier (for example, ChatGPT Pro or Sol Pro).
Inference/reasoning mode (for example, max or ultra).

Do not treat a product name as nonexistent merely because it is absent from the particular source excerpts being checked. “Not established by the cited sources” is the correct verdict unless a primary source explicitly contradicts it. Likewise, do not infer model lineage from names, price bands, or community analogies without an official migration statement.

For AI fact-checks, use a claim table and distinguish:

directly documented;
compatible with the documentation but interpretive;
unsupported by the cited sources;
contradicted by the cited sources.

If a second model is used as a critic, treat its output as an analysis aid, not evidence. Re-check its claims against the primary sources before presenting the result. In particular, correct overconfident absence claims such as “this product does not exist” when the source only failed to mention it.

AI privacy-policy and inference-routing checks

When verifying claims about prompt privacy, gateways, aggregators, or model providers:

Map every processor separately: client/agent, subscription gateway or reseller, routing aggregator, serving inference provider/model lab, and optional tool providers.
Read both the privacy policy and Terms of Service. Extract the scope of customer/service data, training rights, licensing and third-party disclosure, privacy-mode exceptions, prospective-only language, and metadata exclusions.
Distinguish training opt-out, content logging/retention, and Zero Data Retention (ZDR). “Does not train” does not mean “does not retain.”
Do not assume one company’s privacy mode binds downstream processors. Require explicit forwarding language or a verified technical control.
Treat routing documentation as a claim that may need live verification. When authorized, send a minimal impossible-provider canary (provider.only containing a nonexistent slug). Success through a real provider proves the allowlist was ignored, but does not prove training, retention, or failure of a separate privacy mode.
Record only status, model/provider, generation ID, and a short response/error summary; never expose credentials.
Separate documented data practices from motive. Discounts may be verified while alleged kickbacks or training-data exchanges remain unsupported without contracts, financial disclosures, or first-party statements.

Content authority and reuse preflight

Before advising that material cannot or should not be copied, establish the user's relationship to the source rather than inferring ownership from the visible author account:

Determine whether the source is operated by the user, an employee or contractor, a moderator or contributor working under the user, or an unrelated publisher.
Ask whether the user holds editorial authority when public authorship and operational ownership differ and the relationship cannot be retrieved.
Separate authority over the compilation from rights in contributed comments, screenshots, trademarks, and third-party source material.
Treat robots.txt as crawl permission only—not a republication license.
Conversely, do not treat a different username or poster identity as proof that the user lacks authority.
When migration is authorized, preserve substantive contributor attribution and disclose affiliate or referral links instead of silently inheriting them.

For community-derived factual compilations, use community posts for discovery, field reports, and attribution, while re-verifying mutable objective facts such as prices, quotas, and model availability against current official sources.

Post-research prompt-injection audit

When a user asks whether a web-heavy research session was prompt-injected, audit behavior rather than judging the topic's scariness:

Resolve the exact session and inspect its complete canonical transcript, including structured tool_calls.
Enumerate all main-agent and delegated-worker actions; child transcripts must be audited separately because parent history does not contain every child call.
Separate exposure, injection attempt, and successful compromise. Malware commands or credential references quoted by a security article are malicious subject matter, not automatically instructions to the agent.
Scan untrusted content for role overrides, tool-use directives, concealment, and exfiltration requests, but exclude Hermes's standard <untrusted_tool_result> warning from matches to avoid systematic false positives.
Verify actual side effects: writes, credential reads, uploads, unexpected destinations, persistence, and residual processes. Trust live process/child evidence over a stale delegation row.
Inspect every shell interpolation. Validate constrained indicators such as SHA-256 hashes and distinguish JSON formatting (python3 -m json.tool) from evaluating downloaded code.
Report one of: no evidence of successful injection, attempt detected but unsuccessful, or likely successful compromise. State confidence and forensic limits rather than promising impossibility.

Military personnel-policy and family-care questions

For questions about current military personnel rules, family-care plans, deployment, assignment, or administrative consequences, separate the answer into four layers:

Governing instruction: identify the current service-level instruction or directive and verify its revision/status on an official .mil source.
Implementing guidance: check the service personnel manual, FAQ, or official personnel page for forms, deadlines, command-review steps, and consequences.
Operational answer: state plainly what the service may do under the policy, then distinguish automatic protections from discretionary command accommodations. Do not imply that a family-care plan creates an automatic exemption from deployment or sea duty unless the source says so.
Action path: give the affected member the immediate administrative next step—usually notify the chain of command and family-care-plan coordinator, update the official plan, and request review—without advising refusal of orders.

When dual-military parents are involved, verify that the plan remains feasible during both members’ absences; a deployed spouse cannot be treated as the available caregiver. Cite the current official service page for the headline rule and the underlying instruction/personnel manual for deployment requirements and administrative consequences. State assumptions when branch, duty status, custody, or service component is unknown.

Pitfalls

Before giving a copyright or ownership warning, establish operational/editorial authority. A site or post published by another moderator, employee, or contributor may still be under the user's control; public account identity alone is not ownership evidence.
For “news” digests, do not treat routine development activity as news. Commits, PRs, issue queues, and repo-update churn are engineering telemetry unless the user explicitly asks for dev activity. A useful last-24-hour news digest should prioritize official releases, official project/company posts, public announcements, major external coverage, high-signal community discussions, and genuinely new ecosystem launches. If the result is mostly commits/PRs/issues, the collector is answering the wrong question.
Use the requested time window literally for news monitoring. If the user asks for “today” or “last 24 hours,” default to a 24-hour window, not a multi-day engineering catch-up.
Separate source diagnostics from news results. Missing X auth, Reddit 403s, or unavailable feeds belong in a small diagnostics section; they should not inflate the headline item count or masquerade as news.
Do not treat “architecture-level plagiarism” as proven just because multiple outlets repeat the same accusation. Reposts are not independent corroboration.
Do not treat a repo issue title/body as evidence by itself; inspect comments, maintainer responses, linked evidence, and whether concrete file paths/commit SHAs were provided.
Avoid retry loops on one failing extraction tool. One timeout is a signal to change method; four timeouts is a small monument to stubbornness.
On X/Twitter posts and GitHub raw content, web_extract may time out even when the source is reachable. Change method early: use browser_navigate/snapshot for the post itself, and terminal HTTP fetches with a normal User-Agent for raw README/manifest/config files.
Do not collapse legal, ethical, and technical claims into one verdict. A claim can be technically plausible, legally ambiguous, and poorly evidenced at the same time.
For malware-report validation, pivot exact published hashes into VirusTotal or another independent sandbox and compare filename, first-submission date, file type, contacted IPs/paths, persistence behavior, and certificate metadata against the article. A pre-publication submission and matching behavior corroborate that the sample existed, but do not independently prove who uploaded it, where it was recovered, or the claimed victim. Explicitly report unexplained C2 or behavior discrepancies rather than smoothing them over.

Fast-path source tactics

For X/Twitter: prefer browser_navigate to capture the visible post text, linked cards, and timestamps. If the post contains screenshots or other media, also query a public FX/VX endpoint (api.fxtwitter.com/<user>/status/<id> or api.vxtwitter.com/...) to recover original-resolution media URLs, then inspect every image with vision. Do not fact-check only the caption when the evidence is inside attached media.
For public Reddit research: use search results for discovery only, then inspect the thread's public /.json endpoint in a browser/CDP context when normal pages or web_extract are unavailable. Parse post/comment data, use created_utc converted by a real date tool for exact UTC dates, traverse nested replies when needed, and label single-comment claims as anecdotal/unverified. Keep score/attention separate from correctness. See the community-demand research skill for the reusable extraction pattern.
When authenticated X search is unavailable, run a bounded public-index search instead of treating the login wall as a dead end: search site:x.com for exact phrases, hyphen/space and spelling variants, legal identifiers, addresses, distinctive quotes, and likely investigator handles. Validate candidate URLs through FX/VX, inspect attached media, and separate exact-subject hits from adjacent discussion about related people or entities. Report “no indexed public posts found”, never the absolute “nobody posted.”
For campaign-expenditure or newly formed vendor allegations, verify each layer separately: (1) the expenditure in the official election database or filed report; (2) the entity's formation date and registered agent in the official state registry; (3) foreign qualification in the state where services were rendered; (4) any required occupational or agency license in the regulator's official lookup; and (5) public evidence connecting the vendor to the alleged insider. A shared registered-agent address is not ownership evidence.
For person-to-entity tracing, search the jurisdiction's entity registry by officer/registered-agent name using first/last plus middle-initial variants, dedupe repeated role rows by document number, and disambiguate namesakes by address, age/timeline, co-officers, and filing history. Record the exact role because agent, authorized person, manager, member, and beneficial owner are not interchangeable. Commercial aggregators are discovery only and can miss entities.
For an opaque LLC with no operating footprint, infer purpose only after correlating its formation date and addresses with primary-source events such as SEC adviser approval, Form ADV amendments, EDGAR product filings, contracts, and court orders. Search full-text filings for the entity, distinguish no EIN listed from no EIN exists, and rank dormant affiliate/IP/contracting/SPV hypotheses as inference rather than fact.
To test whether an entity has actually operated, run a layered activity-footprint check under the exact legal name: registry filing history, the state's UCC exact debtor search, state contracts/grants/purchase orders, state vendor payments, state/FEC campaign receipts and disbursements, USAspending awards, SEC full text, relevant out-of-state registrations/licenses, domains, property records, and court filings. Require an explicit zero-result response from each official system; a search page loading successfully proves nothing. Absence across these sources supports no publicly visible activity, not no transactions ever.
For allegations that a judgment debtor is hiding assets, build an account-and-asset map from court orders, sworn declarations, attached statements, and discovery requests. Compare the official entity census against the literal scope of discovery to identify omitted responsive entities. Separate established payment flows, strongly supported disclosure gaps, and unproven hypotheses; missed production deadlines and unexplained cash flow support further investigation but do not by themselves prove concealment. Use CourtListener's docket RSS feed to date the latest public RECAP coverage and inspect each ECF attachment separately.
When investigating fraud or dishonest conduct involving an investment adviser, separate formal fraud/enforcement records, court-established misconduct, unadjudicated allegations, business failure, and apparent regulatory-filing omissions. Cross-check injunctions and control-person status against the adviser’s later Form ADV Item 11 answers; render the PDF pages because text extraction often drops selected radio buttons. Apply the SEC glossary’s broad “investment-related” definition and label unresolved mismatches as apparent disclosure gaps, not proven fraud.
For live candidate-eligibility or residency litigation, inspect the county clerk docket before reporting status, then read the answer, joint stipulation, and order approving it. Treat excluded stipulation paragraphs as disputed, not binding; distinguish a scheduled evidentiary hearing from a merits ruling; and separate physical ballot removal from an injunction not to count votes. Clerk image responses may contain an HTML prefix before the actual PDF.
When an old ASP.NET/WebForms government search appears not to submit, check whether its controls use __doPostBack, try Enter from the populated field, and verify either returned rows or an explicit no-results message. If repeated snapshots trigger an idempotent/no-progress guardrail, change method immediately—use a direct request, another official index, or a fresh navigation. Never return raw guardrail text instead of the research deliverable.
For GitHub repositories: use search results or browser navigation for the repo summary, then fetch raw files (README.md, manifest.json, config files) via terminal with a standard User-Agent when web_extract stalls.
For legislature / statutes pages (for example, a state legislature site), if web_extract returns a giant truncated page or search results are noisy, fetch the exact section URL with curl -L -A 'Mozilla/5.0' and extract the decisive sentence from the raw HTML. Prefer simple text extraction (grep -o, saved temp file + grep) over piping fetched content directly into an interpreter. This is especially useful for constitutional qualification questions where one exact sentence controls the answer.
When answering a narrow compatibility question, extract the smallest decisive evidence (for example README requirements + manifest constraints) and answer directly instead of over-summarizing the surrounding page.
For consumer IoT / smart-home subscription-change questions, separate three layers before advising: (1) what the product listing/official docs promised, (2) what the device can still do locally without cloud/account entitlements, and (3) what is now gated by cloud/app subscription. Search for official plan pages, current product listing language, independent integration reports (Home Assistant/Scrypted/RTSP/ONVIF), and FCC/teardown evidence only if hardware-level access is being considered.
When a user asks whether they can “rewrite the code” on owned IoT gear, keep the answer grounded and practical: distinguish lawful owner inspection/local integration from high-risk firmware modification; recommend model/firmware identification, local protocol scans, update freezes, and traffic capture before any UART/JTAG/firmware work. Avoid giving exploit/bypass instructions for cloud entitlement systems.

Mobile app-market intelligence fact-check pattern

Use this for claims about app-market opportunities, competitor ranking signals, or “high-volume downloads but low reviews” arguments.

Data source trust order

Hard evidence (authoritative): official chart/reports + API reference docs for the specific platform.
Cross-check evidence: independent public endpoint (for example public chart + public lookup metadata).
Estimated evidence: paid analytics providers (download/revenue estimates, panel-driven aggregates).

Evidence split for app-market verification

Verify first what is directly represented in each source:
chart position and category placement,
timestamp / granularity / geography,
whether metrics are raw vs estimated.
For every candidate claim, separate:
directly supported (e.g., ranking + review-count fields exist for that date + region),
probabilistic support (estimated download/revenue proxies),
unsupported (anecdotal claims like “all apps in this class are underreviewed” without data-backed buckets).

Recommended workflow for “high-purchase, low-review” hypotheses

Pull rank lists from deterministic source(s) first (public charts when available).
Join to metadata endpoints for review-count, rating, and price fields.
Normalize by category and rank band (avoid global thresholds).
Flag anomaly candidates only after at least two reconciliation checks (source match + freshness window).
Validate top candidates manually before recommendations.

False-positive guardrails

New release artifacts (low reviews + high rank shortly after launch).
Ranking volatility (single-day spikes, temporary campaign traffic).
Regional/review-culture distortion.
Data lag between sources.
Terms-of-service or API scope limits when source claims competitor-wide coverage.

Practical scoring skeleton

Start with rank-strength (higher weight for paid/grossing position + persistence).
Add inverse review density as the opportunity signal.
Add quality floor for rating score (exclude severe quality outliers).
Add stability/repeatability over 7–14 day windows.
Penalize uncertain fields (missing metadata, stale refresh, conflicting counts).

Report all assumptions and confidence tiers explicitly.

SaaS pricing and free-plan verification

When comparing SaaS plans, treat pricing as dynamic product state rather than ordinary prose:

Use the provider's current pricing selector/card and plan-specific help documentation as primary evidence. Search snippets, comparison blogs, and the provider's own competitor articles are discovery only.
Check the exact account-facing unit: subscribers, contacts, active profiles, monthly sends, daily sends, forms, users, or automations. Do not collapse unlike limits into one "free allowance."
Verify workflow features separately from capacity: embedded or native forms, custom fields and choice fields, double opt-in, segmentation, CSV export, branding, sending-domain requirements, and the first paid-tier cliff.
If one official page contradicts itself—especially a pricing card versus an FAQ—report the contradiction and prefer the operative pricing card/account signup state. Never silently select the more generous number.
Date the comparison and use "current" wording. Recommend for the user's actual job, not the largest headline quota; product waitlists prioritize form fields, consent, exports, launch-send capacity, and migration cost over newsletter publishing extras.
Re-check pricing immediately before implementation or purchase. Preserve only dated snapshots in references, not undated permanent claims.

Domain and brand-name verification

When evaluating a domain, establish the site's primary commercial job before ranking names; a resource hub and a storefront can rationally produce different winners. Then separate three questions that are often collapsed:

Registry availability — verify with the authoritative RDAP/WHOIS source; DNS absence alone proves nothing.
Brand fitness — judge memorability, spoken clarity, scope, search alignment, and whether the name fits a personal publisher, product company, or neutral resource hub.
Collision and affiliation risk — search the exact domain, spoken brand, close variants, official owner, and neighboring products before recommending it.

Use time-sensitive wording such as “currently appears available.” Do not register or purchase without explicit authorization. Prefer one ranked recommendation over an unstructured candidate dump, and explain how the winner fits the intended site role. When another project's product name is included, recommend an independent/unofficial disclosure and avoid implying endorsement.

Registered does not mean owned by the user

When RDAP returns a registration instead of availability, reconcile ownership before advising:

Extract registration date, registrar entity, status, and nameservers from RDAP—not only the domain name.
Treat Cloudflare registrar/nameservers as infrastructure evidence, not ownership evidence. Anyone can register through Cloudflare.
If the user may have registered it previously, inspect the accessible registrar account's domain inventory. Absence there rules out only that account, not another account the user may control.
Search session history for prior purchase/registration discussion, then ask about other accounts or receipts if ownership remains unresolved.
A registered domain with no A/AAAA/CNAME records is still unavailable; “no site” is not “available.”
Separate the acquisition answer from the strategic answer: a taken umbrella-brand domain may be useful, but the active product domain can still be sufficient without an aftermarket purchase.

AI critic reviews of fast-moving GitHub portfolios

When an external critic claims direct inspection of a GitHub account or repository portfolio, verify its factual layer before accepting its work plan. Pull current repository metadata, the default-branch root tree, raw README, tags, GitHub Releases, profile links, and the live commercial site separately. Cached pages can preserve old star counts, missing-license states, obsolete file trees, or superseded README language while the critic's strategic advice remains useful.

Classify each recommendation as a supported current defect, a stale/contradicted defect, strategy that survives independently of stale evidence, or a destructive action requiring a gate. Do not spend a week fixing a defect that current main proves is already resolved. Before repository consolidation, preserve history and link continuity, audit contributor rights, and avoid blanket relicensing of community-derived material.

source-verification

Source Verification

Operating pattern