Skip to content

Repository files navigation

🛡️ Aether by SF2X

Don't trust. Verify.

Aether is an AI trust verification layer that catches LLM hallucinations in real time using a 3-model tribunal architecture. It turns unverified AI outputs into audit-ready, cryptographically warranted answers.

📊 Benchmark Results

All numbers are Aether's own internal suites, run through the live pipeline and published — regardless of outcome — on the in-app methodology page. They are vendor-run measurements on small, versioned suites, not independent audits.

Suite n Measures Latest published result
Correlation audit 24 (12 true / 12 hallucinated) Does the trust score separate true from hallucinated claims? AUC 1.0 · Pearson r=0.98 on this suite — a strong internal signal at this n, not proof of perfect separation in general
Negative controls 30 (+5 thin-coverage probes) Fabricated-claim catch rate, true-claim pass rate, honest abstention Gates every release; runs published live
Tribunal lift 6 hard adversarial questions Single model vs. full tribunal Per-item table published

Withdrawn claims. Earlier versions of this README compared Aether (91/100) against "Plain LLM 14/100, Vanilla RAG 28/100, RAG + Output Validation 51/100" with derived "hallucination rates," and cited a "91 with red-team / 58 without" ablation. Those baselines were never evaluated by any code in this repository, so they are withdrawn until they can be re-measured and published with n and methodology attached. Live, always-current results: aether.sf2x.com/benchmark.

🏗️ How It Works

Input Text → [Proposer] → [Critic] → [Verifier] → Cryptographic Warrant

The red-team loop is architectural — the tribunal is adversarial by design. (The previously cited 91-vs-58 ablation is among the withdrawn claims above.)

Two paths run that shape, and they are not the same depth — worth knowing which one you are calling:

  • Fast path (/verifyResponse, the widget, the extension, the Action) — one model plays proposer, critic and verifier in a single pass, ~2-4s. Cheap and quick; not three models.
  • Full tribunal (the playground and escalation path) — mid-tier proposers from three independent labs each reconcile with their own critic, then a cross-firm top-tier model verifies and tries to falsify the result. Asymmetric routing: the answer is written by a cheaper model, every catch is made by the best one.

🔐 Verify it yourself

Warrants are published to an append-only transparency log you can audit without trusting us: signed tree heads, RFC 6962 inclusion and consistency proofs, and a published signing-key document (warrantRegistry?op=keys). The registry publishes integrity metadata and hashes — never answer content. See docs/API_REFERENCE.md.

🔌 API Quick Start

/verifyResponse answers without a key, capped at 5 verifications/day per IP. Every other endpoint requires an x-api-key header and is metered against your plan's monthly credits.

curl -X POST https://aether.sf2x.com/api/functions/verifyResponse \
  -H "Content-Type: application/json" \
  -d '{"text": "According to Section 4.1, all employees get 15 vacation days."}'

🔧 GitHub Action

Advisory by default. v2 reports findings and exits 0; it gates on per-claim dispositions, not a single score. Set mode: enforcing to fail the build. (threshold is deprecated — consulted only against older servers that return no dispositions.) Source and full input reference: github-action/.

- uses: sf2x/aether-hallucination-guard@v2
  with:
    api-key: ${{ secrets.AETHER_API_KEY }}
    text: prompts/output.txt
    # mode defaults to 'advisory' — reports, never fails the build

📁 Structure

  • app/ — the Aether web app + Base44 backend functions (the product itself)
  • github-action/ — CI/CD hallucination guard
  • sdk/ — Python + JavaScript SDKs
  • mcp-worker/ — MCP server, so agents can call Aether as a tool
  • claude-skill/ — Claude skill package
  • chrome-extension/ — Browser extension
  • desktop-extension/ — Desktop extension bridge
  • docs/ — API reference + integration guides
  • docs-site/ — Static docs landing page
  • marketing/ — Launch materials

💰 Pricing

Metered in credits, not raw calls: a verification costs 2 credits, a warrant 5, a full inquiry 10, and gate checks are free. Monthly allowances (enforced by PLAN_QUOTAS in app/base44/shared/apiAuth.js):

Plan Price/mo Credits/mo
Free $0 5 verifications/day without sign-in · 100 credits on a free API key
Starter $20 250
Pro $100 1,000
Enterprise $1,999 15,000
Enterprise BYOK $999 ~200,000 fair-use (you bring your own provider key)
Scale $9,999 150,000
API Starter $49 10,000 (~5,000 verifications)
API Pro $199 50,000 (~25,000 verifications)

No tier is unlimited. Running out returns 429 rather than billing an overage.

🔗 Links

👤 Author

Cameron Piper@Deepsea1


Don't trust. Verify.

About

Aether by SF2X — AI trust verification layer. 3-model tribunal that catches LLM hallucinations. 91/100, AUC 1.0. Chrome extension, API, GitHub Action, and public playground.

Topics

Resources

Contributing

Security policy

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages