AI reliability / Recorded prototype

ClaimGuard.

Check the claim. Preserve the rest.

Explore the repository

TARGETED REPAIR / ILLUSTRATION

Fix the claim. Keep the context.

Orion publishes a product FAQ. The warranty lasts two years. Customers can read it before buying.

The middle sentence conflicts with the source.

Source / README fixtureThree-year warranty in the FAQ.

Fictional sample. Read the recorded report for the real run and its framing limitation.

WHAT IT DOES

A verification pipeline that checks generated articles against sources, repairs only failing sentences, and leaves an inspectable audit trail.

WORK IN THIS REPOSITORY

Deterministic gates, quoted verification, bounded repair, and offline guarantee tests.

THE INTERESTING DECISION

Evidence quotes must literally occur in the named source. A verdict without a valid quote is downgraded to UNSUPPORTED in code.

START WITH THE EVIDENCE2 attempts

Maximum per failing sentence, not an unbounded agent loop.

Inspect the source ↗

BUILT WITH

PythonGeminiLexical retrievalpytest

PROJECT TYPE

Recorded prototype

START HEREREADME & setup

01 / THE IDEA

Check the claim. Preserve the rest.

The problem

Generated content can contain plausible but unsupported numbers, invented experience, and contradictions. Another model's confident verdict is not enough to establish evidence.

The approach

ClaimGuard segments an existing article, extracts atomic claims, applies deterministic gates, verifies quoted evidence, and selectively repairs failing sentences. Unchanged sentences must remain byte-identical; unresolved claims are routed to a person.

02 / THE SYSTEM AT A GLANCE

Follow the flow.

01Stable sentence IDs
02Atomic claims & gates
03Quoted verification
04Bounded repair
05Audit report

A simplified architecture overview. The repository documents the implementation details.

03 / THE ENGINEERING DECISIONS

Why it's built this way.

Verify the verifier

Evidence quotes must literally occur in the named source. A verdict without a valid quote is downgraded to UNSUPPORTED in code.

Bound the repair

Each failing sentence gets at most two repair attempts. Rewrites are re-verified; unaffected sentences are protected by a byte-identity assertion.

Keep simple work deterministic

Phrase-pattern and orphan-figure gates avoid model calls. IDF-weighted lexical retrieval is deliberately sufficient for the four-document fixture.

04 / WHAT TO LOOK AT

Evidence over adjectives.

Repair budget2 attempts

Maximum per failing sentence, not an unbounded agent loop.

Recorded evidenceFull audit trail

Committed report, repaired article, and claim-level JSON from a real Gemini run.

Test fixture6 planted failures

Authored sample includes a deliberately unfixable manufacturing claim; this is not a production benchmark.

Capabilities and result snapshots follow the project's current README, reviewed October 10, 2026. Read the source

Recorded diff and report Claim-level audit Offline pipeline tests

05 / THE HONEST PART

What I'd improve next.

Current boundaries

The first-person blocklist can miss unsupported experiential framing; a recorded rewrite exposes this limitation.
All recorded end-to-end runs used Gemini. The configured Anthropic path has not been exercised end to end.
The small authored fixture does not establish general factual accuracy or production-scale retrieval quality.

The next iteration

Require explicit source attribution for experiential claims and evaluate a broader independently collected corpus.

Open my Software & AI CV

KEEP EXPLORING

ClearStep

Less paperwork. More clarity.