Ariel: Source-Verified Wisdom

An AI wisdom agent powered by GPT-5.6 that verifies every quotation against immutable source records before showing it, blocking fabricated or altered citations.

Comment

Inspiration

AI systems can produce convincing quotations that are inaccurate, altered, or attributed to the wrong source. Ariel: Verified Wisdom explores a stricter approach: the model does not directly control the final quotation shown to the user.

What it does

The project introduces a deterministic source-verification layer for AI agents.

Each source is registered with:

  • a stable sourceId
  • exact UTF-8 text
  • a SHA-256 integrity hash

When an agent wants to cite a source, it provides the sourceId and an exact character range. The verification layer checks that the source exists, that its contents have not changed, that the range is valid, and that the quotation exactly matches the registered source.

A Claim Gate blocks unsupported claims and altered or fabricated quotations before they reach the user.

How we built it

The first working vertical slice was built with OpenAI Codex as a dependency-free Node.js prototype.

It currently includes:

  • an immutable source registry
  • SHA-256 integrity verification
  • exact quotation extraction
  • UTF-8, Hebrew, and diacritic-safe range handling
  • structured deterministic errors
  • a Claim Gate
  • a local CLI demonstration
  • automated tests

The current prototype passes 51 tests with 0 failures and directly covers 20 out of 20 mandatory verification cases.

The final submission goes beyond the CLI prototype. A local web demo connects GPT-5.6 (gpt-5.6-sol) through the OpenAI Responses API with strict Structured Outputs. The model returns only an interpretation, a support status, and an opaque reference id — it never supplies the displayed quotation. The quotation is reconstructed from the immutable registry (JPS 1917 sources retrieved from Sefaria with an explicitly pinned Public Domain version) and verified against SHA-256 and exact UTF-8 byte ranges. A transparent tampering simulation alters the citation after the model finishes — and the Claim Gate blocks it. Final state: 95 automated tests passing, live API call completed.

Challenges we ran into

The main challenge was preserving exact source text across Hebrew, diacritics, UTF-8 boundaries, invalid ranges, altered quotations, and duplicate or missing sources.

A second challenge is separating two questions that are often confused:

  1. Is this the exact registered quotation?
  2. Does the quotation semantically support the claim?

The current prototype solves the first problem. Semantic entailment is not yet implemented.

Accomplishments that we're proud of

We built a working, dependency-free source-verification prototype rather than stopping at an architectural concept.

The system can:

  • register immutable source records
  • preserve exact UTF-8 text, including Hebrew and diacritics
  • verify source integrity with SHA-256
  • extract quotations using precise source ranges
  • reject missing sources, invalid ranges, altered quotations, and unsupported claims
  • return structured and deterministic errors
  • demonstrate both successful verification and blocked citations through a local CLI

The prototype passes 51 automated tests with 0 failures and directly covers all 20 mandatory verification cases defined in the test plan.

We are also proud of the architectural separation between the AI agent and the verification layer. The model may propose a claim and reference a source, but it does not control the final quotation shown to the user.

What we learned

We learned that source verification contains several separate trust problems.

A matching SHA-256 hash proves that registered content has not changed, but it does not prove that the source is authoritative or that it was trustworthy when first registered.

We also learned that exact quotation verification and semantic support are different problems. A system may confirm that a quotation is exact while still needing another layer to determine whether the quotation truly supports the claim being made.

Working with Hebrew, diacritics, Unicode boundaries, and exact text ranges showed that seemingly small encoding decisions can directly affect the reliability of an AI citation system.

Finally, we learned that deterministic software should remain responsible for facts that can be checked mechanically, while the language model should focus on interpretation, explanation, and reasoning.

What's next for Ariel: Verified Wisdom

The GPT-5.6 integration is complete and demonstrated in the video. GPT-5.6 (gpt-5.6-sol) generates structured answers through the OpenAI Responses API, returning only an interpretation, a support status, and an opaque reference id. Before anything reaches the user, the deterministic verification layer confirms the source, integrity hash, range, and exact quotation — and the demo shows both outcomes: a valid source verified and displayed, and an altered quotation blocked by the Claim Gate.

Next steps:

  • explore semantic entailment as a separate verification layer, so the system can also assess whether a verified quotation truly supports the model's claim
  • add persistent source storage
  • add authenticated source provenance (signatures and a verifiable chain of custody for the registry itself)
  • introduce CI and automated verification
  • expand the registered corpus beyond the initial two sources

The long-term goal is to make this a reusable trust layer for AI agents working with legal, medical, educational, research, religious, and enterprise knowledge.

Built With

Share this project:

Updates