Truth Onion

What's built

The engine runs today. This page states what exists, what is designed but not yet built, and the numbers behind both.

Built and running

The engine

  • The rules layerDecides where every claim sits. Enforced in application code, with database constraints and triggers as an independent second layer, so no request can set a tier directly
  • Source library with rippleKeeps every claim answerable to its sources. When a reviewer marks a source discredited, every claim resting on it is re-evaluated in the same transaction
  • Challenge and promotionHow claims move. Promotion runs a claim through a battery of challenges it must survive; demotion follows once the support fails, with a correct-and-resubmit path. Withdrawals and routing decisions are two-phase — filed by one party, adjudicated by another, with both outcomes kept permanently
  • Import and exportLets topics travel between instances without trusting the file. Every imported item re-runs through the rules layer, and a tampered export is refused with the same plain reasons as a live violation
  • 2D and 3D viewsTwo ways of working the same map — flat concentric rings, or nested spherical shells you move through — sharing one dataset and one dial. Neither view can change a placement itself; every change goes through the rules layer, whichever view you work in
  • The depth dialControls how much uncertainty you see. Core-only by default, deeper layers by choice, and the count of what sits deeper is always shown
  • Private working spaceWhere half-formed notes live before they become claims. Outside the epistemics entirely — unlinkable, invisible to every view — until a note is ready to earn a placement
  • Kernel links and lineagesA debunked claim links back to the documented ground it overreaches from: what the kernel establishes, what was asserted beyond it, and the evidence path inward. Kernel links carry no evidentiary weight in either direction, by structure
  • The time machineReplays any topic or claim as it stood at any past moment, from the recorded event log. Reconstructions name what cannot be honestly rebuilt instead of guessing, and no historical view can reach a write path
  • Claim pagesStable, script-free, server-rendered permalinks for every claim, with share cards that carry status inseparably. Each page has its own on-page time machine and a feedback form with no identity fields
  • Record permanenceRecord entities are never hard-deleted. Sources leave a case only through a proposed-and-adjudicated withdrawal with a stated reason; what left, and why, stays visible
  • The sandboxAct as Curator, Contributor, and Reviewer against the real rules layer; a proposer never upholds their own entry, enforced in code. Sandbox saves export as files — the same save files the demo’s contribute / feedback panel accepts
  • Walkthrough and tourA setup walkthrough and a guided tour ship with the engine

Verification

Checked by machine, not asserted by model

The engine retrieves sources itself and verifies quoted text server-side, against the page actually fetched. The verified state is computed, not authored — no model can declare a source checked.

Retrieval runs through a guarded proxy, with a headless browser only where a page will not render otherwise. Where a page cannot be read, that is recorded as unreadable rather than silently treated as a failed quote.

The companion

Built to argue with you

An optional AI companion, connected to a provider of your choosing with your own key. Nothing is proxied through this project and no keys are held server-side: OpenRouter, Anthropic, or Google Gemini, billed directly by your provider. There is no inference markup.

OpenAI direct is deliberately unsupported — api.openai.com does not accept direct browser connections (no CORS), so the settings entry is disabled with that reason and points at OpenRouter instead of failing silently.

Its win condition is your claims surviving scrutiny, not your claims being promoted. It argues against you, surfaces adverse sources unprompted, and never proposes a tier. Its tools are read-only, and the executor refuses any tool name it doesn't recognize.

  • Analysis passRuns with the character card structurally absent from the request — not instructed against, not stripped afterward, not included
  • Rendering passThe persona receives the findings and delivers them in voice, having taken no part in producing them
  • Fidelity checkConfirms the substance survived the rendering; it is what stands between a persona and an unproven claim that sounds likely
  • Character cardsWritten by you, any persona you like — the guarantee comes from where the card sits in the pipeline, not from limits on what it may say
  • Write accessNone. It advises; the rules decide

The reason for the split is the same reason the rules live in the database. A promise that a model will ignore its own persona is enforced by good behavior. Removing the persona from the request is enforced by construction.

Tests

289 automated tests across twenty-one suites

A permanent pressure-test suite attempts laundering, circular support, self-published stacking, cross-topic smuggling, and edit-after-placement on every run — and every later stage pins its own guarantees the same way, from kernel-link grammar to sandbox isolation.

  • Definition of done14
  • Pressure tests10
  • Stages 2–2.5 (multi-topic, library, dial)19
  • Stages 2.9–2.9e (kernel links, legibility, tokens, providers, progress)71
  • Stages 2.95–2.98b (time machine, tour, parking, claim pages, permanence)59
  • Sandbox & personas (2.99a)29
  • Kind adjudication & recast (2.99b)8
  • Release checklist & history fixture20
  • Demo enforcement5
  • Companion40
  • Fetch proxy14

Content

Five complete topics

MKUltra

12 claims about the CIA's mid-century behavioral research program, each placed by the documentary record behind it.

COINTELPRO

9 claims about the FBI's domestic counterintelligence program, each placed by the documentary record behind it.

The Replication Crisis

10 claims about replication and reproducibility in psychology research, each placed by the evidence behind it.

UAP Disclosure

11 claims about government UAP releases, hearings, and reports — sworn testimony split from its content, releases split from what footage depicts, documented disinformation beside the phenomena it imitated.

AI Evaluation

19 claims about AI capability and safety evaluation — benchmark scores split from what they measure, self-published system cards from their contents, and expert forecasts recorded as claims about surveys, never about the future.

Designed, not built

What comes next

Stated as designs rather than features, because that is what they are.

  • Taxonomy revisionThe category system was built against the first three topics; the fourth and fifth strained it in ways the strain journal records. Its vocabulary gets one deliberate revision, drawn from those recorded cases where the current terms misfit — rather than drifting patch by patch. Sequenced next, before multiplayer, so the revision sees every mechanic it has to name
  • A heightened bar for claims about living peopleAccusations against living, identifiable people will carry the strictest evidence requirements in the system — a named gate in the rules layer that lands before multiplayer opens
  • Multiplayer and adversarial reviewOpening the map to other people's challenges, under persistent pseudonymous identity. Demo save files import through the real rules layer entry by entry, carrying no standing in from the file; the tier challenge splits in two, filed by one party and adjudicated by another, as withdrawals already are. Reputation is earned as much by careful demotion as by confirmation, and dissent is preserved rather than overwritten
  • Topic pagesPublic pages for a topic, as claim pages already are for a claim — surfacing what has been reviewed lately and what moved, so the rulings most worth checking are findable rather than buried
  • A shared worldA 3D space where user-built sections surround a common bedrock of established claims — the world's geometry is the epistemology