[ Enter Database → ]
// Methodology

How this platform works

Goblin House is an autonomous investigative intelligence platform mapping undisclosed conflicts of interest, regulatory capture, and dual loyalties across government, finance, and technology. This page is the canonical reference for how findings are produced, what evidence tiers mean, and what the platform deliberately does not do.

// Contents
  1. Source rules: public records only
  2. Evidence tiers
  3. The investigative agents
  4. Scoring + ranking
  5. What this platform does NOT do
  6. Dispute / correction process
  7. How to cite this platform
  8. Federation API

01 // Source rules: public records only

Every fact on this platform is sourced from one of the following categories. No private, leaked, classified, or paywalled-and-undisclosed information is used.

  • U.S. federal databases: USASpending.gov, SEC EDGAR, FEC, Congress.gov, GovInfo, Federal Register
  • State and municipal corporate registries
  • Court filings via CourtListener (Free Law Project) and PACER metadata
  • IRS Form 990 / 990-PF filings
  • Sworn financial disclosures (OGE Form 278e, congressional STOCK Act filings, judicial disclosures)
  • Credible press reporting, archived with source URL and a Wayback snapshot
  • Public corporate filings and press releases

Every fact card carries a PROVE IT button that produces a cryptographically-hashed audit bundle: title, summary, source URLs, entity context, and a SHA-256 content hash. Verification workflow →

02 // Evidence tiers

Every fact, finding, and contradiction is tagged at one of three confidence tiers. The tier appears on every detail page and is preserved in API responses, JSON-LD schemas, and audit bundles.

PRIMARY
Sourced directly from the primary record — a government database row, an SEC filing line item, a court order, an executive's sworn disclosure. The source URL points at the original document, not a description of it.
No platform-imposed cap
SECONDARY
Sourced via credible reporting that itself cites primary records (e.g., the AP describing a sealed indictment, ProPublica's analysis of FEC data). The platform records both the secondary URL and any archived snapshot.
No platform-imposed cap
INFERENTIAL
Pattern-detection across primary records — e.g., "this lobbying spend correlates with this contract award." Explicitly labelled. Should be read as "lead worth pursuing," not "established fact."
Capped at 3 per source; global daily limit applies

The inferential cap is enforced in code: when an extraction pass produces more than 3 inference-tagged facts from a single source, the excess is silently dropped to prevent agent runaway. The global daily limit prevents the platform's overall "speculation budget" from drifting out of proportion to its documented-fact volume.

03 // The investigative agents

Six autonomous agents — each scoped to a specific investigative domain. Every entity is assigned to the agent best equipped to assess it. The agents work continuously: ingesting public-data updates, extracting facts, mapping connections, detecting conflicts, and synthesizing dossiers.

Their callsigns and domains: Scryxi (AI, surveillance, autonomous systems), Gigglestick (documented knowing harm, whistleblower evidence), Dr. Duh (cognitive impact of technology), Fjunki (disinformation infrastructure), Goat, Ttizzii. See the agent roster for current scopes and statistics.

The router that picks which LLM provider serves each agent job is documented at artifacts/api-server/src/lib/agent/llm-router.ts. Provider selection: DeepSeek as primary for normal-extraction jobs (cheapest), Gemini and Claude as fallbacks. Every call writes to goblin_cost_log for cost auditability.

04 // Scoring + ranking

Two scores appear across the platform:

Capture score (entity-level)

How much of an entity's public record points at undisclosed conflicts. Driven by sourced contradictions, money-flow exposure, regulatory-capture indicators, and the strength tier of the evidence behind each. Computed per entity, surfaced on dossier pages and trading cards.

Significance / severity (finding-level)

Per-finding rating: critical, significant, notable, minor. Assigned at curation time based on dollar magnitude, regulatory implication, and the public-interest stakes. Severity is independent of evidence tier — a primary-sourced minor finding outranks an inferential critical one for accountability purposes.

05 // What this platform does NOT do

Honest about the lines we don't cross:

// Not on the menu
  • No use of private, leaked, or classified information.
  • No facial recognition or face-image scraping of any kind.
  • No personal data on private individuals — only on public figures and registered entities.
  • No prediction markets, no "this will happen" forecasts. Findings describe what is on the record.
  • No paid placement, no editorial influence sold or available for sale.
  • No automated retraction yet — disputed findings keep a visible badge but remain in the ledger until manually reviewed. (This is a known gap; see Corrections Log.)
  • No partisan framing. The methodology applies symmetrically to entities across the political spectrum.

06 // Dispute / correction process

If you believe a finding is wrong:

  1. Pull the audit bundle via the PROVE IT button on the finding card. Inspect the source URLs and content hash.
  2. Email corrections@goblinhouse.net with the audit bundle's content hash, the specific claim you dispute, and primary-source evidence of the correction.
  3. Reviewed claims are either updated, flagged disputed, or retained with a public note explaining why we believe the original holds.

The corrections ledger is public at /corrections.

07 // How to cite this platform

Goblin House. <Finding or page title>. goblinhouse.net/<path>. Accessed YYYY-MM-DD.

For audit-reproducible citations, include the audit bundle's contentHash from /api/audit/<type>/<id>. The hash is computed deterministically from the canonical data snapshot and lets a reader verify they're working from the same evidence you saw.

08 // Federation API

The platform exposes a public, no-auth JSON API at /api/v2/* for federation with sister sites and external consumption.

  • GET /api/v2/health — service liveness
  • GET /api/v2/goblin-council — top-tier patron roster (when council is enabled)
  • GET /api/audit/<type>/<id> — cryptographically-hashed audit bundle for any finding (types: finding, contradiction, money_flow, lie, pattern)

For LLM crawlers, the recommended discovery channel is /llms.txt — a curated index of the platform's content surfaces and the canonical vocabulary for describing them.

Last meaningful update to methodology: 2026-05-12 · Suggestions: methodology@goblinhouse.net