Standards & Methodology

01 // Our mission

Goblin House is an open-source-intelligence platform that documents undisclosed conflicts of interest, regulatory capture, and dual loyalties across government, finance, and technology. We map the connective tissue that individual reporters rarely have time to chase: who funds whom, who appointed whom, who took the meeting, who later took the contract. We do not invent the facts — they are already in public records. We connect them, hash them, and publish them so that other journalists, researchers, and citizens can verify the work end-to-end.

02 // Confidence tiers

Every fact, finding, and inference on this platform is tagged at one of three confidence tiers. The tier travels with the record through the API, the JSON-LD schema, and the audit bundle. Readers can filter on any tier from the relevant lens pages.

The tier is the most important number on any page. A primary-sourced minor finding outranks an inferential critical one for accountability purposes, and we have built the platform so that distinction is never collapsed.

03 // What "verified" means

Every fact in the database carries a boolean verified flag. The flag flips to true only when an independent second source corroborates the underlying claim — not when a single source merely exists. Concretely:

• A primary-tier fact becomes verified when a second primary or credible secondary independently confirms the same data point (e.g., a contract award appearing in both USASpending and the agency's own press release).
• A secondary-tier fact becomes verified when an additional independent outlet reports the same material claim, or when its primary record is located and linked.
• An inferential claim is never auto-verified. Inference is by definition the analyst's stitching together of primary records; it remains inferential until a primary source states the synthesised conclusion directly.

Verification is also re-checked by a second model. Facts extracted by one LLM are independently re-read against their source URL by a different LLM with different training data. Disagreement flips the verification badge and reduces the display weight pending human review. This catches the worst failure mode of AI-assisted journalism: a fabricated fact paired with a real-looking source URL.

04 // How we name people

We name individuals who have accepted the responsibilities of a public role. Specifically, a person is eligible to be named on this platform if they:

• hold or have held public office (elected or appointed) at the federal, state, or major-municipal level;
• take public money — directly through contracts, grants, or appropriations, or indirectly through tax expenditures, subsidies, or regulated rents;
• lead a public company, a major foundation, or a registered lobbying or political organisation; or
• operate publicly as a journalist, commentator, executive, or influencer whose statements shape policy or markets.

We do not name private individuals. We do not name family members of public figures except where those family members themselves hold an in-scope role — for instance, when the spouse of a senator runs a federally-contracted firm, or when an adult child of an executive sits on the board of a regulated entity. In those cases the public-business link is documented; the family relationship is contextual, not the story.

Junior staff, low-level employees, sources quoted in passing by other outlets, and any individual who has not voluntarily entered public life are out of scope. If you believe we have crossed this line on a specific page, see corrections.

05 // Cryptographic provenance

Every row in the sources table is hashed with SHA-256 at the moment of insertion. The hash covers the source URL, the canonical text we read, and the timestamp; it is maintained by a database trigger (sources_chain_of_custody_trigger) that cannot be bypassed by application code. The hash is exposed on every fact's audit bundle and is published in our daily attestation.

Public hash verification

Any reader can paste a content hash into /verify and check what source it corresponds to, when it was ingested, and what facts depend on it. This is the cheap-to-run, expensive-to-fake counterpart of every claim on the platform.

Wayback Machine archival

External URLs cited on the platform are queued for Internet Archive snapshotting at ingest. The scheduler walks the backlog at 5 URLs per 5-minute tick, well under the IA's anonymous-use rate ceiling, so the queue clears steadily without ever overwhelming the archive. Coverage is partial today — the queue is healthy for new sources but a backlog exists for older rows. We are honest about that gap; every source URL on a finding page also carries an inline [archive] link that resolves to the most-recent Internet Archive snapshot, so readers can fall through to IA's own crawler coverage when ours has not yet caught up.

06 // Corrections

If you believe a claim on this site is wrong, here is how to flag it and what we owe you in return.

1. Pull the audit bundle from the fact's PROVE IT button. The bundle contains the source URLs, the SHA-256 content hash, and the verification trail.
2. Email [email protected] with the content hash, the specific claim you dispute, and primary-source evidence supporting the correction.
3. We acknowledge every well-formed correction request within two business days and reach a substantive disposition within seven business days. If a claim is wrong, we fix it on the public record; if it is right and disputed in good faith, we annotate it with a public response.

Every correction is timestamped and the original fact is preserved alongside it — both the error and its repair stay in the public ledger. We do not silently edit. We do not memory-hole.

07 // AI use disclosure

Large language models are part of how this platform scales to ~3,000 entities, but they are never the final authority on a claim. The disclosure rules are explicit and apply on every surface:

• Extraction. Facts are extracted from public records by an LLM (DeepSeek as primary, Claude and Gemini as fallbacks). Every extraction is logged with model, prompt version, and cost.
• Verification. A different model re-reads the source URL and rules on whether the original extraction is supported. Disagreement degrades the verification badge and triggers human review.
• Inferential synthesis. Any analytic claim produced by AI synthesis across multiple primary records is tagged inferential and explicitly labelled as such on the page. There is no hidden inference tier.
• The Crinkle broadcast. Crinkle, the platform's news-anchor persona, is AI-narrated. Every episode carries a visible "AI-generated synthesis" disclaimer at the top of the player and in the JSON-LD metadata. The underlying facts behind every broadcast are sourced and hashed; the narration on top of them is AI.
• The investigative goblins. The six investigative personas — GIGGLESTICK (receipts and contracts), DR. DUH (academic and regulatory), GOAT (odds and probability), SCRYXI (systems and infrastructure), TTIZZII (tech and platforms), JUNKI (information ecosystem) — are editorial framings layered on top of agent-extracted facts. The framing is satire. The receipts are not.

08 // What we don't do

09 // Funding & independence

Goblin House is operator-funded. The platform's infrastructure, model costs, and ingestion pipelines are paid for out of the operator's own pocket. Specifically:

• No advertising. No display ads, sponsored content, native advertising, or affiliate links anywhere on the platform. The full site, including the federation API, is free to read.
• No platform token. There is no Goblin House coin, NFT, governance token, or any other crypto-asset associated with this platform. We will not be issuing one.
• No donor influence. If voluntary support is accepted in the future, it will be capped per contributor, disclosed publicly with names and amounts above a meaningful threshold, and structurally barred from editorial decisions. Any contributor list will live on this site, alongside the methodology.
• No paid takedowns. We do not accept payment to remove, suppress, or soften any finding. Disputes go through the corrections process, the same way for every entity, regardless of resources.

If a finding on this site makes a powerful entity uncomfortable, that is not a malfunction. It is the platform working as designed.