◊ fall-verify
prime 401 ┃ R5 mirror · ◐ verify ┃ v1.0 · sovereign

did Claude hallucinate that·3-vote adversarial verification·sovereign

Catch the plausible nonsense
before it propagates.

52% of LLM-generated claims fail adversarial verification
(n=25, deep-research run, May 2026)

Convene the panel

paste a claim · pick presets · run the 3-vote panel

3
2
demo mode · no key · click a preset

Panel deliberating

3 panelists · each given a distinct lens

VERDICT
confidence · —

SDK · drop in anywhere

import { verify, batchVerify } from 'https://sjgant80-hub.github.io/fall-verify/fall-verify-sdk.js'; const v = await verify({ claim: "GPT-4 was trained on 13 trillion tokens", panelSize: 3, // 2-5 · default 3 killThreshold: 2, // refute votes needed · default 2 sources: [ // optional · gives panel context "https://arxiv.org/abs/2303.08774", "https://openai.com/research/gpt-4" ], context: "Discussing model scaling trends", // optional byoKey: 'sk-ant-...', onPanelist: (v) => render(v), // optional · per-panelist stream }); // → { // verdict: 'confirmed' | 'refuted' | 'inconclusive', // confidence: 'high' | 'medium' | 'low', // votes: [{ panelist, role, vote, reasoning, citations, elapsed }, ...], // summary: 'Claim refuted 2-1 · GPT-4 token count never disclosed publicly', // suggestedFix: 'GPT-4 training token count has not been publicly disclosed by OpenAI' // } // batch interface · parallel const results = await batchVerify(claims, options);

Use cases · where verification earns its keep

RAG fact-check

Verify every claim your retrieval-augmented chain produces against its own sources. Catch the moment the model wanders off the corpus.

Content moderation

Adversarially vet user-generated claims before they ship. False-statistic posts caught upstream of publication.

Code review

Hallucinated API signatures, fabricated library names, wrong version numbers — the panel catches what the LLM smuggled past you.

Audit chain

Pro tier signs every verdict with Konomi keys. Your audit trail proves you ran the gate before propagating the claim.

Deep research

The workflow that birthed this primitive. Kill 50%+ of plausible nonsense before it lands in the final report.

LLM eval

Use the panel as a judge over your own model's outputs. Cheaper than human review, sharper than single-model self-critique.

Tiers · Konomi shim · prime 401

Free

trial · branded
  • 10 verifications / day
  • panel size ≤ 2
  • fall-verify badge
  • demo presets unlimited

Sovereign

MIT · BYO key
  • unlimited verifications
  • panel size 2-5
  • no branding
  • IndexedDB history
  • works from file://

Pro

batch · webhook · audit
  • batchVerify in parallel
  • webhook on verdict
  • Konomi-signed audit chain
  • every verdict cryptographically signed

Why adversarial

Every LLM application generates plausible nonsense. A single-vote check inherits the model's blind spots — it doesn't know what it doesn't know. The 3-vote panel works because each panelist is prompted with a distinct adversarial lens: skeptic, technical detail, nuance. Disagreement is the signal. When the panel splits 2-1, the claim deserves the human's attention. When the panel agrees 3-0 to refute, the model can keep walking. That's the gate. No public competitor packages this as a primitive.