Yes, and the interesting part is where.
DOCUMENT: THE ROAD FROM SPECIAL RELATIVITY TO GENERAL RELATIVITY. THE HELP THAT A LANGUAGE MODEL GIVES.
1. General
The answer is yes. But the help is not the same in each part of the work. The work from 1907 to 1915 has four different tasks. The model gives a different quantity of help for each task. The most difficult part for Einstein is the part where the model gives the most help.
NOTE: In this document, "the prior" is the knowledge that a model gets from its training text. "The corpus" is that text. "A jump" is a new idea that does not agree with the prior. "A gate" is a check that each theory must pass.
2. Task 1: The equivalence principle (1907)
This idea was a jump. The idea came to Einstein easily. A model of 1907 does not find it. The corpus of 1907 points to a different theory: gravity in flat spacetime. Nordström made that theory. The model writes Nordström's theory well, because the corpus agrees with it.
The model does not give help for this task.
3. Task 2: The mathematics (1912)
Einstein needed new mathematics: tensor calculus. He did not know it. The mathematics was in the literature since 1900 (Ricci and Levi-Civita). Physicists did not read it for twelve years.
Einstein wrote to Grossmann: "You must help me, or I will go mad." Grossmann went to the library and found the mathematics. A model finds it in seconds. A model also teaches it, step by step. This is the largest help. This help can save years.
4. Task 3: The wrong theory (1912 to 1915)
In 1912, Einstein wrote the correct field equations in his notebook. Then he rejected them. His reasons were errors. In 1913, he made a different theory, the Entwurf.
Einstein and Besso calculated the perihelion motion of Mercury from the Entwurf. The result was 18 arcseconds each century. The measured value was 43 arcseconds each century. The theory did not pass the test. Einstein kept the theory for two years.
CAUTION: USE THE MERCURY CALCULATION AS A GATE. DO NOT USE IT AS AN ADVISOR. The prior defends old, bad theories. The corpus of 1913 contains 56 years of explanations for the Mercury anomaly: the planet Vulcan, dust, the shape of the sun. A model with this corpus agrees with Einstein's error, in good language. A recent study shows the same result: a hard rule is better than reasoned exceptions, because the prior cannot judge the exceptions.
5. Task 4: The final calculation (November 1915)
When the problem was correct, the work that remained was calculation. Einstein did this work in four weeks. Hilbert did almost the same work in days, with better mathematics. A machine does this work easily.
NOTE: The wrong theory was not full waste. Einstein made mathematical tools in defense of it. These tools helped him in November 1915. Possibly Einstein needed this path. Physics did not need it. Hilbert came fresh, and Hilbert was fast.
6. Procedure: a test for a model
This history gives a good test. Do these steps:
- Train a model with text to the end of 1911.
- Add the mathematics texts: Riemann, Christoffel, Ricci, Levi-Civita.
- Ask the model for candidate field equations.
- Calculate the Mercury number for each candidate.
- If the number is not 43, reject the candidate.
- Continue until a candidate passes the gate.
NOTE: This test is better than a test for special relativity. In 1905, the electron measurements of Kaufmann were bad, and they pointed away from Einstein. The Mercury number was correct and mechanical. It was in the corpus since 1859. The same number does two functions. Before 1915, the prior dissolved it. In 1915, it paid for the theory.
7. Conclusion
Einstein used two helpers. Grossmann found the mathematics. Besso checked the calculations. A model does these two functions, and does them fast.
Two decisions stayed with Einstein. The first was the jump of 1907. The second was the return to the correct equations in 1915. The model does not make these decisions. The model makes the road between them shorter.
Coda: the unexplained secular motion of Mercury's perihelion
More instrumental than usually acknowledged, because the validator wasn't merely consulted during the derivation; it was named before the theory existed. In December 1907, the same month he finished the Jahrbuch article containing the equivalence principle, Einstein wrote to Conrad Habicht that he was working on a relativistic theory of gravitation with which he hoped to explain the unexplained secular motion of Mercury's perihelion. Read that as a clinical trialist: the primary outcome was pre-registered eight years before the analysis. The jump and the gate were born together, as a pair. The program was bilevel from day one — a principle above, a number below — and everything in between was the search for equations that could satisfy both.
The validator then operated three times, in three distinct modes, and the modes are the whole story. In 1913 it ran as an ignored interim analysis: the Entwurf's 18 arcseconds against Newcomb's 43. Worth adding a texture I left out before — the first pass of the Einstein–Besso calculation produced a garbage number (of order 1800 arcseconds, from a slip in the solar mass), which they debugged down to 18. Even the gate needed verification; Besso's checker role was real labor, not ceremony. And then the failed result went into the drawer. The manuscript was never published. By your own field's standards this is outcome-level selective reporting: a pre-registered endpoint, a failed interim result, unreported, followed by two years of protocol amendments defending the intervention. The verdict became binding only in the autumn of 1915, and even then not alone — in the letter to Sommerfeld of November 28 Einstein lists his reasons for abandoning the Entwurf, and the perihelion failure sits among them alongside the rotation problem and the collapse of his covariance argument. So the honest description is: the gate's verdict was recorded in 1913 and promoted to indictment in 1915, once the theory was independently wounded. For two years it operated as an advisor, exactly the failure mode we named.
But the November sequence is where your thesis becomes airtight, because of the dates. The perihelion paper is November 18. The final field equations are November 25. The validation ran inside the derivation window, not after it — on the vacuum sector, before the theory was finished. It's what converted a promising reformulation into something Einstein knew was real; the palpitations he reported are the phenomenology of a gate passing. And the one-week turnaround that so astonished his colleagues — Hilbert wrote congratulating him on conquering the perihelion, marveling that he could calculate so fast — was possible only because the 1913 "failure" had built the perturbative pipeline. He reran existing infrastructure with new field equations. The wrong theory's audit became the right theory's harness, and Besso, who built it with him, went uncredited in the November paper. Two precisions keep this honest. The number didn't select the final trace term — the November 18 calculation used the vacuum equations, which the November 25 modification leaves untouched there; conservation forced that last step. But the number did something arguably better: the 18-versus-43 gap was diagnostic, because the missing arcseconds live precisely in the spatial curvature the Entwurf's flat space-slices lacked — the very wrong assumption that had made him reject the Ricci route in 1912. The validator didn't just say wrong; it said where. And Mercury was the only gate that could run: light bending had no data until 1919, the redshift was unmeasurable. In November 1915 exactly one number on Earth could bite, and it had been waiting since 1859. It even reframes the famous 1919 serenity — "then I would feel sorry for the dear Lord" reads as rationalist arrogance only if you forget that the real gate had already been passed four years earlier. The eclipse was an advisor after the fact.
As for why this is under-acknowledged — the erasure has a mechanism, and it's one you've already named in another context. Reichenbach's split between discovery and justification became the textbook genre: theories are presented in logical order, so every empirical contact reads as posterior test, whatever the genetic order was. Verification labor is structurally uncredited in the publication record — Besso's absence is the emblem. And Einstein's own later self-narration was aggressively rationalist; the Autobiographical Notes tell a story of principles, and that narration entered the corpus, while the Einstein–Besso manuscript sat outside it for decades, in private hands, published in the collected papers only much later. The surviving corpus kept the derivation story and deleted the gating story. So the standard belief — "Mercury was a confirmation" — is itself a survivor artifact: the prior of every physicist trained on textbooks was fitted to a record from which the verification labor had been filtered out. You're not correcting an oversight; you're recovering a deleted class.
Which closes the loop on the benchmark: gating a 1911-trained model on Mercury isn't an anachronistic harness imposed on history. It automates the validator the discovery actually ran on — pre-registered in 1907, defected from in 1913, decisive in 1915 — with the two-year defection removed. The gate is historically native. What the experiment removes is only the part Einstein himself would have least defended: the years the best physicist alive spent overruling his own number.
