Andrew Edmond

Independent buy-side technical due diligence

linkedin.com/in/andrewedmond ↗
Southern Spain · Remote & onsite

← All posts

Get Camino build log · 4 of 10

The truth pipeline

For an app that tells people what the law requires of them, being right can’t be a value on a poster. It has to be a system that refuses to ship the alternative.

Right is a processHow a fact earns its way into Get Camino
  1. 01

    Source at scale

    15 hours of expert webinar transcripts, mined by parallel agents. The catalog grows to 60 obligations.

    Fast and broad — but not yet trustworthy.

  2. 02

    Ground in primary sources

    28 uncited obligations re-verified against the law itself: the BOE, the tax agency, the DGT.

    Caught: a penalty the courts had struck down. A form that no longer exists.

  3. 03

    Gate the build

    No page may state a number its cited source doesn't carry. The lint fails the deploy.

    First run: caught a violation before any user saw it.

My app tells people what the Spanish government requires of them. “Mostly right” isn’t a quality bar here — a wrong fact is someone missing a residency deadline.

So the most important system in Get Camino isn’t a feature. It’s the process that decides what’s true.

Step 1 — Source at scale. Claude mined 15 hours of expert webinar transcripts with parallel agents, growing the catalog to 60 obligations. Fast, broad — and not yet trustworthy.

Step 2 — Ground everything. 28 of those obligations arrived without primary citations. Every single one got re-verified against the actual source: the BOE, the tax agency, the DGT. Not blogs quoting blogs. The law itself.

That pass caught two things no prompt would have: a penalty that a court had struck down (still cited all over the expat internet), and a form that had been abolished. The web remembers rules that no longer exist. An AI trained on the web remembers them too.

Step 3 — Make truth a build gate. The honesty rule — no page may state a number its cited source doesn’t carry — stopped being a value and became a lint that fails the deploy. On its first run it caught a violation before any user ever saw it.

Here’s the lesson I’d offer every leader shipping LLM products: hallucination is not a prompt-engineering problem. It’s a systems-design problem. You don’t ask the model to be honest. You build a pipeline where dishonesty can’t ship.

AI collapsed the cost of drafting 60 obligations. It did nothing to collapse the cost of one of them being wrong.

This is one post from a longer build log. If you're weighing an AI-built codebase, that's the work I do — or find me on LinkedIn.