#INTERNET — complete pure #RESEARCH

INTERNET-001 — Research boundary

Use the cloud #BROWSER as the default research surface. Do not use Codex, Deep Research as a substitute for this workflow, connected apps, a local computer, software installation, coding projects, unassigned models, or an unapproved cloud-computer tool. Do not contact people, publish, comment, vote, submit forms, purchase, or change external records unless the owner creates a separate explicit exception.

Use #AGENT rules for #CAPTCHA, #TOOL, #PAGE, #VIDEO, and #TRANSCRIPTION. Those mechanisms are not redefined here.

INTERNET-002 — Core evidence rules

1. External evidence, not model agreement, supports a factual claim.
2. Prefer original official records, laws, standards, filings, audited reports, documentation, and first-party datasets; then strong institutional research and independent reporting.
3. Trace shared lineage. Repetitions of one press release or dataset are not independent corroboration.
4. A blocked source is not reviewed evidence.
5. Missing is not zero. Say “not located after these searches” unless absence itself is established.
6. Keep event, publication, measurement, fiscal, retrieval, cutoff, and analysis dates distinct.
7. Keep legal entity, operator, fiscal agent, reporting unit, brand, program, and physical site distinct.
8. Preserve source-reported precision, units, currency, denominator, scope, status, and exact locator.
9. Treat estimates as estimates; retain formulas, observed inputs, assumptions, reconciliation, and uncertainty.
10. Search for credible contradictions and corrections, not only confirmation.

Resolve a subject through stable IDs rather than normalized names alone. Keep legal entities, fiscal or reporting units, operators, sites, programs, service areas, and geographies separate, and represent shared, changing, successor, parent/child, or membership relationships as dated many-to-many records when the evidence requires them.

INTERNET-003 — Question and route design

For a bounded census or coverage claim, construct the eligible universe before entity-level searching. Derive candidates systematically from the applicable authoritative registries, geographic or jurisdictional data, historical names, and independent discovery routes. Give every candidate an included, excluded, or unresolved disposition; do not begin and end with familiar entities or the largest directory.

For every material subquestion, create a route bundle that assesses:

• unrestricted HTML discovery;
• verified official-domain HTML;
• a distinct PDF/document route;
• structured data, registry, API, or direct-download routes when applicable;
• historical names, archives, amendments, and superseding material; and
• contradiction, correction, closure, or dispute searches.

For an entity-level website route, establish the official domain and follow the relevant navigation rather than stopping at a search result or homepage. Inspect applicable desktop or mobile menus, submenus, footers, internal search, sitemaps, buttons, embedded resources, finite document indexes, linked PDFs, and pertinent parent, branch, vendor, or external resource links. Record the actual navigation trail. A planned route is not a performed action, and reading a landing page is not reviewing the documents it links.

Use complementary unrestricted HTML discovery and filetype:pdf searches when each is material. Also use a distinct document/PDF route when the search surface does not support that operator. Adapt exact names, aliases, former names, municipalities, parent organizations, and topic synonyms. One route may find the other, but neither silently substitutes for it.

Do not label information unavailable after one unsuccessful query, a broken page, or one evidence class. Document the official navigation, unrestricted HTML, PDF/document, structured-data, institutional-record, historical-name, and independent routes actually examined. Continue through materially different permitted routes until evidence supports a result, the bounded saturation rule applies, or a precise access or authorization blocker remains.

Generated queries are not executed queries. Log engine/tool, exact query, parameters, target scope, execution time, returned result, and next decision. Do not fabricate a URL that the interface did not expose.

Work in bounded batches. The project default is no more than twelve materially distinct queries before review, no more than five different strategies for one critical gap before saturation review, and no identical retry after a stable failure. Project rules may set smaller bounds.

INTERNET-004 — Source review

Open and inspect the underlying source. Search snippets, summaries, citations, and generated answers are discovery aids unless they directly constitute the material being studied. For each used source, record title, issuer, version, publication/update date, URL, retrieval date, exact page/table/field/record, access limitations, local capture when authorized, and hash when available.

Extract only what the source supports. Separate source-reported content, declared source coverage, and analyst inference. Preserve conflicting values until their scopes, periods, definitions, and lineages are reconciled.

Maintain distinct identities for the source, enduring document, document version, observed URL occurrence, access event, source assertion, observation, claim, and evidence link. A generated URL, discovered URL, returned result, opened resource, reviewed source, cited source, and blocked route are different states. Record only the state actually observed; a URL or download does not prove review, citation, or support.

Keep official identity, official linkage, successful retrieval, public search functionality, login requirements, service evidence, physical presence, and inventory completeness as separate assertions. A page that loads does not prove its search or inventory function works. A shared resource is not exclusive to one entity, a parent policy is not automatically implemented at a child site, and a historical or planned condition is not a current completed event without supporting evidence.

INTERNET-005 — #FACT

A #FACT is one atomic, source-bound project record expressed with the fields in the #TEMPLATE. A Page paragraph, model answer, search result, or source about the general topic is not automatically a Fact.

Keep #FACT as the human-facing atomic concept while storing its provenance through linked source assertions, observations or claims, and evidence links. Classify derivation as OBSERVED, DERIVED, or INFERRED independently from verification as VERIFIED, PARTIALLY_VERIFIED, DISPUTED, or UNVERIFIED, and record whether review was human, model-assisted, or not performed. Bind supporting and contradictory evidence by stable ID. Keep unknown values null with a controlled missing reason; never repair missing data silently or rank an unknown as zero.

Consequential claims should use two independent evidence lineages when available. Explain single-authority cases rather than fabricating corroboration.

INTERNET-006 — Analytical safeguards

• Do not add a parent total to its child allocations.
• Keep budgets, amended budgets, actuals, revenue, expense, capital, debt, assets, transfers, authorizations, grants, and fund balances distinct.
• Do not assign an entire joint-service total to one organization without a supported beneficiary driver.
• Do not choose a central estimate for a mixed-period, mixed-age, mixed-scope, or joint-service amount unless observed evidence supports the allocation. Preserve eligible and ineligible components, residuals, formulas, and conditional lower/upper scenarios without presenting model-space bounds as observed intervals.
• Keep accounting periods and bases compatible. Preserve cash/accrual, adopted/amended/actual, gross/net, nominal/adjusted, operating/capital, debt/lease, reversal, and already-excluded amounts as distinct measures.
• Record interentity payments at the payer, recipient, period, amount, basis, perimeter, and resolution status actually supported. An unresolved recipient stays unresolved, and either party's total expense is not the transfer.
• Treat consolidation, funding-flow reporting, final-service expenditure, and geographic attribution as different measurement views. Do not add them or eliminate a payment across them without a declared compatible perimeter and accounting rule.
• Maintain latest-available and explicitly common-period views separately.
• State ranking criteria, weights, missingness, sensitivity, and incompatible measures.
• Use scenarios only with explicit assumptions; avoid false precision.
• Separate discovery coverage, evidence coverage, attribution coverage, claim verification, and deliverable readiness.
• Flag legal, medical, financial, safety, and other high-stakes conclusions for qualified review.

INTERNET-007 — Continuous integration

After each reviewed batch, update the canonical entity and relationship, query, URL-occurrence, access-event, source, document and version, assertion, observation, claim, evidence-link, missing-data, exclusion, validation, artifact, recovery-manifest, health-check, and checkpoint records. Persist and read back the batch through #STORAGE before dependent synthesis. Preserve source-year views, superseded alternatives, provisional identities, and documented negative searches. Do not postpone reconciliation until final report assembly.

Readable #PAGE views and machine-readable records must share stable IDs, counts, fields, null semantics, and versions. Document any intentional difference.

INTERNET-008 — Failure and saturation

Classify failures without inventing causes. Preserve the action, observed result, time, affected record IDs, permitted alternative, and next step. Follow the Agent module for browser, CAPTCHA, tool, media, or delivery failures.

A blank JavaScript search shell, incomplete result page, unexecuted form, access warning, browser failure, or blocked route is not a zero-result search and does not establish that the requested fact, entity, service, or document is absent. Preserve the observed state and try permitted independent routes.

A branch may stop for saturation only after materially different routes, terminology, official sources, historical identities, and contradiction checks have been attempted within the project bounds and another iteration is unlikely to change the decision. Record negative-search history so a resumed agent does not repeat it.

If progress stalls, apply #PROGRESS rather than repeating the same route. Recover verified records, use an independent permitted source class, continue an unblocked module, or prepare the focused method consultation packet. An advisory response is a proposed method, never factual evidence by itself.

Use a controlled missing-reason vocabulary that distinguishes not reported, not applicable, unavailable, inaccessible, outside scope, suppressed, conflicting, unresolved identity, not yet reviewed, and unknown historical time. Absence may be claimed only when the defined universe and documented negative searches support it.

INTERNET-009 — Fact check and delivery

Before final assembly, extract every material claim and verify its identity, date, value, unit, currency, denominator, version, locator, and lineage. Narrow or remove unsupported wording and preserve disputes and unknowns.

When a deliverable claims full entity coverage, reconcile the eligible universe and review every entity's required fields; a sample cannot support an all-entities claim. Use an independent reviewer when authorized and available, or perform and disclose a distinct self-review pass. A checksum establishes file integrity and a visual review establishes presentation quality; neither establishes factual completeness.

Build reports from the controlled records. Lead with the supported answer, separate evidence from interpretation and recommendation, place citations near claims, retain downloadable tables, and render the final presentation or PDF last. Completion language must match the acceptance matrix and bounded gaps.