I built a pipeline that takes a single editorial brief and produces fact-checked, structured content that search engines and language models can cite. When a claim turns out to have no source behind it, the pipeline stops the piece before anyone can publish it.
Hi, I'm genuinely a writer and genuinely a builder.
I spent the last few years writing and curating large volumes of training data. Alongside that I've worked on grading rubrics, production and system prompts, and LLM-as-judge systems that score conversations at scale. Lately I've been working on agentic pipelines that draft and check content, though I still make the final call on what ships.
Before that, I spent fifteen years as a writer and editor, much of it in science and medical publishing for specialist audiences, where proper fact-checking is extremely important.
I can read Python, run and modify scripts and debug my way through them. I work in Claude Code and Cursor, and I'm reasonably comfortable with the Anthropic and OpenAI APIs.
The brief holds the questions the piece has to answer, the rules for its voice, and a numbered list of sources, each with a line I wrote saying what that source is allowed to support. Nothing outside that list is admissible: the drafting step can't use anything else, and the checking step is told plainly that whatever the model happens to know does not count as evidence.
If a question can't be answered from those sources, it comes back empty along with a note explaining what evidence would be needed to answer it. Deciding what counts as a source is a judgement I make before anything is generated, and it stays with me rather than with the model.
What that looks like — draft answer to Q3, before checking
Solidgate states that orchestration can lift approval rates by 10–30% using the same providers you already have. It does not publish the methodology, sample size or measurement period behind that range. Test it against your own transaction data before using it in a business case.
C10 · advisory · unsupported · no source in the pack gives this advice
The model added that last sentence on its own, and none of the four sources in the brief say anything of the kind. It’s harmless here, since a vendor would probably compare their own figures against whatever is found in the wild anyway. However, the same move in a piece on a more demanding subject would hand a reader something they could act on and be wrong about, and the pipeline already takes that possibility into account.
The first run was blocked because four of its claims had nothing behind them, and three of those four turned out to be advice to the reader rather than statements of fact. I had expected to catch an invented statistic, but what the model produced instead was guidance, written into answers that were otherwise accurate and properly sourced, which is considerably harder to notice while reading. A third run followed, after I edited the copy for readability and put it back through the same check.
| Claim | Type | Verdict | Source |
|---|---|---|---|
| A payment gateway handles payment card transactions on a merchant's behalf. | definitional | supported | S1 |
| A gateway is not a system for tailoring or optimising payment flows. | definitional | supported | S2 |
| Solidgate states orchestration can lift approval rates by 10–30%. | attributive | supported | S3 |
| Approval rate is always equal to or higher than first-attempt authorisation rate. | comparative | supported | S4 |
| Test it against your own transaction data before using it in a business case. | advisory | unsupported | — |
An extract from the first run, which checked sixteen claims in allEach claim is checked in a separate call of its own
The prompts are kept as numbered files rather than edited in place, so the drafting prompt has four versions and the reason I replaced each one is written down beside it in the repository.
The pipeline produces a draft that has passed its fact-check, which is not the same as copy that reads well. Below is what the pipeline wrote, what I changed, and why. My edited version went back through the check afterwards, because copy that skips the gate defeats the point of having one.
What does a payment gateway do?
A payment gateway handles payment card transactions on a merchant's behalf. In standard industry definitions, it's essentially another name for a payment processor or payment service provider, not a separate category of function. It's distinct from an acquirer, which is the financial institution that processes card transactions for merchants and is defined by a payment brand as an acquirer.
A payment gateway handles your card transactions for you. It's essentially another name for a payment processor or payment service provider. It's distinct from an acquirer, which is the financial institution that processes card transactions for merchants and is formally defined as such by a payment brand.
Editor’s noteThe first sentence answered the question, just not in words anyone would use, so I simplified it. "In standard industry definitions" — does the reader care? "Not a separate category of function" is machine speech. And an acquirer defined as an acquirer reads awful.
What does a payment orchestration layer add on top of a gateway?
Question: what does an orchestration layer add on top of one or more gateways?
An orchestration layer integrates multiple payment service providers, gateways and payment methods into a single platform. A gateway provides a means of access to payment processing; it is not a system for tailoring or optimising payment flows. Solidgate describes orchestration as routing, retrying and reconciling transactions centrally across multiple providers.
Question: what does an orchestration layer add on top of a gateway?
An orchestration layer integrates multiple payment service providers, gateways and payment methods into a single platform. A gateway provides a means of access to payment processing; it is not a system for tailoring or optimising payment flows. Solidgate, a payments vendor, describes orchestration as routing, retrying and reconciling transactions centrally across multiple providers.
Editor’s note"One or more gateways" is needlessly specific for a question someone would actually type. Solidgate also turns up out of nowhere, so I clarified who they actually are and added a link to the source.
How much do approval rates improve after adding orchestration?
Solidgate states that orchestration can lift approval rates by 10–30% using the same providers you already have. It does not publish the methodology, sample size or measurement period behind that range, so the figure is a vendor claim rather than a measured benchmark.
Solidgate states that orchestration can lift approval rates by 10–30% using the same providers you already have. They don't share the methodology, sample size or measurement period behind that range, so keep in mind that it's a vendor's estimate rather than a measured benchmark.
Editor’s note"It does not publish" is clumsy, and "they don't share" is plainer — it's clear from the context that we're still talking about Solidgate. The ending was too dry, so I clarified it in simpler words.
Authorisation rate is the percentage of card authorisation requests approved by the issuing bank, calculated as approved authorisations divided by total authorisation attempts. Approval rate is broader: it also counts transactions recovered through retries, failover or alternative payment methods after an initial decline. Because of that, approval rate is always equal to or higher than the first-attempt authorisation rate.
No answerI simply couldn’t answer this from the evidence pack. The only related figure is Solidgate’s $400,000 monthly threshold, which says when orchestration pays off, not when it doesn’t. To answer it properly I’d need a source about smaller or single-market merchants, and until I have one, I’d rather skip than lie.
Everything is public: the scripts, the prompts with their version history, all three sample runs with every claim and its verdict, and the notes recording what I changed between them and why. Most of the prompt work I do is covered by NDA or belongs to my employers, so I built this one from scratch on a neutral topic in order to have something I could actually show.