Avoiding the Pitfall of Unchecked AI

July 2026

A glowing checklist-like sequence of small verification nodes in a row, one node flagged with a subtle warning-colored glow just before being corrected back to a clean matching glow.

Status: scaffolding — planned, not yet the finished article. This page lays out the structure Matthew wants this piece built around, and pre-populates the one part that’s already available: a real inventory of the specific errors AI-assisted work introduced into this site’s own papers, later caught and corrected. The deeper layer — the actual root-cause analysis of how those errors got generated in the first place — is waiting on a Grok chat transcript export Matthew is retrieving from xAI. Once that’s available, it gets fed through for genuine root-cause analysis rather than guessed at here. The eventual goal, in Matthew’s own words: turn this into both a publishable article and a formal paper.

Why this is worth writing at all

Most of this site’s papers were built with heavy AI assistance (Grok, and later Claude for the audit pass), and that assistance introduced real, sometimes significant numerical errors — a rotation-curve claim off by 32x, a proton-decay comparison on the wrong side of its own experimental bound, a decay-rate figure with the wrong sign of exponent, borrowed physics results presented ambiguously as if independently derived. All of them were eventually caught, but only because a dedicated audit process went looking for them — they didn’t announce themselves. That’s the actual subject of this piece: not “AI makes mistakes” (uninteresting, everyone already knows this), but the specific, checkable mechanics of how confident-sounding, plausible-looking technical output can be wrong in ways that pass a casual read, and what a real process for catching it looks like in practice.

The framework

Matthew’s own manufacturing-engineering background supplies the structure, and it’s the same vocabulary already governing how this vault steers its own work (see item 11’s “Root Cause, Corrective Action, Look-Across, Poka-Yoke” in the CLAUDE.md guardrails behind this whole site) — this article is that same discipline turned outward into a public case study rather than kept as internal process:

Case study inventory: errors already found and fixed on this site

This is real, not hypothetical — every item below is a documented correction already live on solvetheuniverse.com, most surfaced by a systematic 2026-07-21/22 audit pass. Grouped by what kind of failure it actually was:

Numerical errors that contradicted the paper’s own stated formula or inputs:

A claim that was actually on the wrong side of the real experimental bound it cited as support:

Real, independently-confirmed numbers stated incorrectly:

Real results borrowed from established physics, presented ambiguously as if independently derived:

Unverified statistics stated as if computed:

Corrections that were made in one place but never propagated to where the same claim was repeated:

Structural/technical bugs, not content errors, but caught the same way — by someone actually checking:

What’s already functioning as poka-yoke on this site

Some of what came out of catching these errors wasn’t just a fix — it changed the site’s own structure so the same class of mistake is harder to make silently again:

What’s still pending

The genuine root-cause layer — why these specific errors got generated by the AI assistance in the first place, what patterns in the original prompting or verification (or lack of it) let them through, and what Matthew’s own actual reaction was in the moment he first suspected something was wrong — needs the real Grok transcript, not a reconstruction. That’s the part this page is deliberately leaving blank rather than inventing plausible-sounding root causes to fill the gap. Once the transcript is available, the “What Did I Do” and “Resolution” case-study sections above get filled in with Matthew’s real, specific account rather than this session’s own inference from the final corrected text.