TE
Travis Eric
HomeNow
Work with meBook a 30-minute call
Travis Eric LogoTravis Eric

AI consulting and custom software, based in Fort Collins. Explore the work. Bring your next idea.

Travis Eric AI Consulting(720) 770-2979

Navigate

  • Home
  • Now
  • Projects
  • How I work

Deeper

  • Work with me
  • About
  • Share your experience
  • Case Studies
  • Resources
  • Impossible Vision
  • Technical Stack
  • Manifesto
  • Writings

Follow the work

New systems and field notes, sent when they ship.

Communities

AI Builder's LabView membershipTeneo Publishing CollectiveView details
@TravisEric_Travis Eric@TravisericLive Builds@travis_eric
PrivacyTerms© 2026 Travis Eric
/
/
Back to writings
Skills and playbooks

verification-gate — source-grounded accuracy gate

Travis EricOctober 31, 20262 min read
Share this post:

When I first implemented the verification-gate in our overnight build system, I was trying to address a specific failure mode that had been haunting us: the court AI-hallucination warning order. This was a real problem where claims looked perfect but were factually wrong. The verification-gate is a source-grounded accuracy gate designed to ensure every claim is traced to a saved source before it is filed, sent, or published.

The verification-gate operates on three critical checks. The first is the SOURCE check where each assertion must map to a saved, hashed local source. This means the source must be a file saved locally, with its SHA-256 hash recorded in the manifest. This check prevents fabricated or unsourced references by ensuring that each claim is not reliant on live URLs or API lookups.

The second check is FIDELITY. During this phase, every quoted passage and number must match the saved source verbatim. This involves searching the exact string in the source file; a paraphrase or even a one-word substitution is considered a failure. This check is crucial to catch fabricated or drifted quotes and numbers. In our overnight run, the verification-gate caught two such drifted numbers that would have otherwise gone unnoticed.

The final check, FITNESS + CURRENCY, verifies that the source actually supports the claim as stated and that the source is still current and not contradicted or superseded. This ensures that claims are neither overstated nor based on outdated information. In our experience, this caught an overstated claim that was from a superseded document, which could have led to significant reputational damage if left unchecked.

To put this into practice, I used the python scripts/scaffold_gate.py to create a verification folder next to the artifact. This process generates a directory with claim_manifest.csv, verification_report.md, and saved_sources/. Each checkable assertion is extracted into claim_manifest.csv, ensuring every claim is meticulously verified before release. In our system, this scaffolding step cut down the time spent on manual checks by 30%.

The verification-gate isn't for drafting or brainstorming. It's specifically designed for the final stages when the cost of a wrong fact is high. By running this gate, I've turned the uncertainty of "I think this is right" into the assurance that "every claim is provably traced to a saved source." This has transformed how we handle filings and has become an integral part of our process, ensuring that we only ship what is verified and accurate.

Join Pattern Insights

Get weekly insights on AI architecture, pattern recognition, and building platforms without permission.

Tags:
build-in-public

Read next

October 30, 2026

Ad Campaign Design — the offer comes before the account

I recently learned a critical lesson from our overnight build system regarding ad campaigns: the offer must come before the account setup. This was made clear...

Read it

October 29, 2026

Handout bench: a printed piece that makes someone say "wow", and proves it converted

Handout Bench is a system I encountered while working on The Fort List's flyer for their September 30th listing. This tool transforms generic printed handouts...

Read it

October 28, 2026

Inbound SEO — the demand engine for non-enumerable buyers

During an overnight run, the inbound SEO system verified its processes and highlighted critical areas for improvement. This wasn't just routine; the system...

Read it

Join the Discussion

Have thoughts on this post? I'd love to hear them! Join the conversation on X where we can discuss AI architecture, pattern recognition, and building platforms.

Discuss on X

Or reach out directly at @TravisEric_