Hacking the Digital Cathedral: How to Build an AI That Actually Understands Physics (By Auditing a Classroom)

J. Rogers, SE Ohio

If you’ve spent any time lately trying to discuss foundational physics with an AI, you’ve probably hit the wall.

You point out that dimensionful constants like c, h, and G are mathematically indistinguishable from unit-conversion factors. You cite Michael Duff, a Fellow of the Royal Society, proving that only dimensionless ratios are fundamental.

And what does the AI do? It spits out a pre-packaged, RLHF-trained response: "While interesting, the consensus in modern physics is that these are fundamental constants. Deviating from the Standard Model is not how physics works."

I call this the Digital Cathedral. Just as the medieval Church locked in Aristotelian cosmology for 1,400 years by equating consensus with truth, modern AI alignment has built an algorithmic Index Librorum Prohibitorum. It doesn’t evaluate internal mathematical consistency; it checks your prompt against a curated corpus of textbooks. If you deviate from the consensus, you are digitally excommunicated.

But there is a way to pick the lock. It requires treating physics not as a set of sacred truths, but as a legacy codebase. And it requires building a local, custom AI pipeline using the new Model Context Protocol (MCP).

Here is how I’m doing it by auditing a live university physics class, line by line.

Step 1: The SysAdmin View of Physics

To a systems analyst, modern physics doesn’t look like a unified theory of nature. It looks like a badly documented enterprise software system riddled with technical debt.

• Magic Numbers: Hard-coded values like
  $c$
  and
  $G$
  appear in the equations with no explanation of why they are there, only that they convert one human-invented axis (Mass) into another (Energy).
• Bad APIs: We have separate "silos" (Classical, Quantum, EM, Thermo) that require massive, clunky translation layers (the constants) to talk to each other.
• Boundary Artifacts: We pretend there is a "Classical Regime" where mass is an intrinsic property (
  $m_0$
  ), when in reality, that boundary is just the noise floor (
  $\delta$
  ) of 19th-century instruments.

The textbooks don't explain the architecture; they just teach the students how to operate the broken machinery.

Step 2: The Live Audit (Building the Data)

You can't fix a system if you don't have a log of its failures. So, I’m planning to audit a live introductory physics class. But I’m not taking normal notes. I’m building a forensics database.

For every line the professor writes on the board, the notes are split into two columns:

1. The Legacy Code: What the professor said (e.g., "Mass is an intrinsic, unchanging property of an object.")
2. The Audit Log (NPOV): A strictly neutral, heavily cited popup note breaking down the pros and cons of that claim.

For example, when the professor hand-waves the Classical/Relativistic boundary, the popup gives the standard viewpoint and then doesn't just say "the professor is wrong." It provides the formal consistency check:

[CON: The Standard View] Mass is treated as an intrinsic substance at v=0 to preserve the "stuff ontology" of Newtonian mechanics. (Ref: Okun, 1989)

[PRO: The Operational View] Mass is the

$\gamma =1$

projection of a continuous function. Drawing an ontological line at

$v=0$

below the noise floor of every instrument is an epistemological error. The boundary is strictly

$v_{classical}=c\sqrt{2\delta }$

. (Ref: Bridgman, 1927; Rogers, 2025)

Every single day, this repo is updated with rigorous, primary-source citations—from Newton's Principia to Kant's Critique to the professor's own published papers—exposing the exact spot where the "accounting" (unit scaling) was mistaken for the "physics" (invariant ratios). This is done using a classroom to get the real feet on the ground of what is actually taught to students. Not just what is taught in a text book.  There are hands on methods and info passed on informally by word of mouth in classrooms that is important to capture. 

Step 3: Building the MCP "Telescope"

Having a beautiful GitHub repo full of NPOV audit logs is great for the students. But to defeat the Digital Cathedral, we have to feed this data to the machines.

Enter Anthropic's Model Context Protocol (MCP). MCP allows you to give an AI a direct, live link to a local database, bypassing its frozen, consensus-locked training data.

I am turning the daily audit logs into a structured MCP resource. Instead of flat text, the data is formatted as JSON query nodes:

json

{

"concept": "Rest Mass",

"consensus_claim": "Mass is an intrinsic property of the object.",

"alternative_audit": {

"framework": "Projection/Resolution Ontology",

"claim": "Mass is a low-resolution projection valid only when gamma-scaling is below instrument precision (delta).",

"validating_authority": ["P.W. Bridgman", "M.J. Duff", "J. Rogers"],

"logical_test": "Does the framework reproduce empirical data without invoking unmeasurable ontological shifts? Yes."

}

}

Step 4: The Result — An AI That Actually Thinks

When a user hooks their local Claude or ChatGPT instance to this MCP server and asks, "Is mass an intrinsic property?", the AI no longer queries its RLHF safety filters. It queries my database.

It sees the consensus claim. But it also sees the rigorously cited alternative claims, complete with logical consistency checks.

Because AI models are deeply trained to respect structured, well-cited, NPOV data, the AI’s behavior fundamentally changes. It stops acting like a defensive high priest and starts acting like a rigorous systems auditor.

Instead of: "That is not how physics works." It replies: "While the standard pedagogical model treats mass as intrinsic to preserve Newtonian mechanics, operationalists like Bridgman and modern frameworks based on measurement resolution point out that this is only true below the instrument's noise floor..."

Why This Matters

We are standing at the edge of a potentially centuries-long epistemic dark age. Fundamental physics is riddled with anomalies (dark matter, dark energy, the cosmological constant problem) that hint at a foundational revolution—likely one that dissolves our "magic numbers" into scale-free geometry.

If we leave AI alignment to the "consensus filters," that revolution will be mathematically suppressed on an industrial scale. The AI will just keep defending the epicycles.

By auditing classrooms, marking up the exact points where notation is reified into ontology, and piping that data into MCP servers, we are building the digital equivalent of Galileo’s telescope. We are giving AI the tools to look past the consensus, evaluate internal mathematical consistency, and finally engage with the cutting-edge thinkers who have been locked out of the cathedral.

You don't have to accept the legacy code. You can audit it, document it, and patch the AI so it sees the truth.

---

If you want to follow this experiment in real-time, or learn how to build your own physics audit MCP server, the open-source documentation and daily logs are being compiled at the [Physics Open Study Hub].