Cross-Validation Methodology
Methodology · v1.0 · 12 May 2026 · Grant Bishop
There are publications that ask the reader to trust the byline. We're not one of them. The byline matters, but the byline alone is not a substrate. What matters is the work behind the byline — the data the work rests on, the claims the work makes, the sources those claims trace back to, and the auditable record that lets a reader see all of it.
This page describes the verification substrate the publication runs on every piece. It is technical only where the technicality is the point. Most of what follows is a description of editorial discipline, expressed in the language of the actual mechanisms.
Three layers of verification
Every piece we publish runs through three independent verification layers before it lands on the reader's screen. The layers do different work. Each is necessary; none is sufficient.
Layer one — the data
Every numeric claim in every piece — market caps, backlog disclosures, earnings figures, the delayed prices tracked on the Bishop's Board — is cross-validated against multiple independent data sources before it enters the publication.
When we cite a quarterly revenue figure, the figure is sourced from the company's filing and validated against at least one second source. When we read a Bishop's Board price, that reading is pulled from multiple independent providers and the values reconciled. When the values disagree beyond a threshold the desk has set for material variance, the data is held until the disagreement is resolved. When the data sources agree, the published value carries the consensus. When they don't, the publication says so.
This is not the same as "we double-checked the number." It is a mechanical pass that runs without exception, logged for every published claim, with a record the reader can inspect. The discipline matters more than any individual data source. Sources fail. Cross-validation catches the failure.
Layer two — the claims
The desk does not score its own work. We never have. Every editorial piece is evaluated by multiple independent judges before publication — and independent here means structurally independent, with different training inputs, different reasoning approaches, and different failure modes. They evaluate the same piece against the same rubric. They produce separate scores. The scores are reconciled.
When the judges agree, the piece publishes. When they disagree within a tolerance the desk has calibrated against the rubric, the piece publishes with the variance recorded. When they disagree beyond that tolerance, the piece is held for desk review. The desk reads the dissent. The desk decides whether the dissenter caught something the others missed, or whether the dissent reflects a known failure mode in the evaluator itself. Sometimes the piece publishes with a note. Sometimes it goes back to the writer. Occasionally it gets pulled.
The rubric is not secret. The dimensions the judges evaluate against are: whether the voice on the page matches the standing voice of the desk, whether every factual claim traces to a source the desk can verify, whether the piece holds together as a piece of writing in the structural conventions the publication carries, and whether the piece addresses the full set of moves a piece of its type is expected to make. Failure on any dimension can block publication. The thresholds are calibrated to catch real failures without false-flagging normal editorial variance.
This is the layer that catches what a single set of eyes — or a single judge — would miss. The desk has watched it work in practice. It catches things.
Layer three — the citations
Every primary-source citation in every piece is mechanically validated before publication. When the desk cites a 10-Q filing, the publication pulls the actual filing from primary sources and verifies the cited number appears. When the desk cites a quote from an earnings call, the publication pulls the transcript and verifies the quote substring. When the desk cites a press release or a regulatory docket, the publication fetches the source document and confirms the cited claim is present.
The result is recorded per-citation, per-piece, in an audit log the reader can inspect at the bottom of every piece. A piece that cites fourteen sources carries an audit-log entry showing fourteen citation checks: which sources were fetched, when they were fetched, whether the cited claim was found in the fetched content. When a citation fails to validate — which happens — the piece carries a flag at that citation, naming it as unverified. The publication does not pretend the failure didn't happen. The publication tells the reader the citation couldn't be confirmed and lets the reader decide what to do with that.
This layer catches the failure mode where a writer cites a source incorrectly, paraphrases a quote that no longer matches the original, or pulls a number from a press summary that doesn't appear in the primary filing. We have seen it catch all three.
The audit log
Every published piece carries a link to its audit log. The audit log is not a marketing feature. It is the record of the work the publication did to verify itself before asking the reader to read it.
The audit log shows: which data sources were consulted for each numeric claim, which evaluators scored the piece and what scores they produced, the variance across evaluator scores, every citation that appeared in the piece and whether each was mechanically validated against its source, timestamps for each fetch and each score, and the version of the methodology under which the piece was processed.
A reader who clicks into the audit log is looking at the same data the desk looks at. There is no separate internal version. The audit log is what the desk used to decide the piece was ready.
We publish the audit log because we think the reader has the right to see it. Most publications don't. Most publications make a series of editorial judgments and ask the reader to trust the result. We make the same judgments. We just make them on the page and on the audit log, in writing.
What we commit to
A few things follow from the substrate, and we name them explicitly:
No quiet corrections. When a piece's citation fails to validate, we surface the failure. When a piece's data source revises and we have to revise downstream, we publish the revision and link it from the original piece. The original piece is not edited silently to look like it never made the original claim. The chain of revisions is part of the work.
No claims of certainty we don't have. The audit log shows uncertainty when uncertainty exists. When a citation could not be verified, the citation reads as unverified. When evaluator variance was wide, the variance shows. When the desk made a judgment call to publish despite some dissent, the judgment call is visible.
No reliance on any single source. The substrate is designed around the assumption that any single source will fail. Data providers go down. Evaluators have blind spots. Citation fetches occasionally return stale content. The verification works because multiple independent surfaces have to fail simultaneously for an error to make it through — and even then, the audit log makes the failure inspectable after the fact.
Methodology versioning. This page is versioned. When we change how the verification substrate works, we publish the change with a version bump, dated and signed. The methodology is part of the editorial product. We don't get to change it quietly any more than we get to delete a Final Call quietly.
The line this draws
We don't publish to be impressive. We publish to be useful. Useful means the reader can rely on the work without having to do it themselves — and can verify the work when they want to.
The verification substrate is what lets the publication say things it would otherwise have to qualify. When the desk writes that a Watchpoint has cleared a threshold, the desk has already mechanically verified the underlying data, has already had the claim scored by multiple evaluators, and has already validated every citation in the piece. The claim is not a guess. It is the output of work the reader can see.
That is what this page describes.
— Grant Bishop