Derived, Not Measured: Why Machine-Precision Math Never Substitutes for an Experiment

I · The Twelve-Digit Illusion: Accuracy Is Not Observation

Open the entanglement kernel behind step T.05 of Time & Gravity and feed it the geometry published by Sougato Bose and colleagues in 2017 (DOI 10.1103/physrevlett.119.240401): two diamagnetic microcrystals of 10^-14 kg each, trap centers separated by 450 um, each mass coherently split across 250 um, falling in vacuum at 10^-15 Pa and 0.15 K for 2.5 s. Under Newtonian attraction, the four spatial branches—left-left at 450 um, right-right at 450 um, left-right at 700 um, and right-left at 200 um—accumulate slightly different quantum phases phi = G m^2 t / (hbar d). Subtract the symmetric pair from the asymmetric pair and the entanglement phase comes out to Delta_phi = 0.31 rad. Pass that phase into William Wootters's 1998 two-qubit formula and the concurrence reads C = 0.16, rising to 1.00 if the drop could be stretched to 25.0 s.

To guard against coding mistakes, the verifier computes that same 0.16 a second way: it constructs the full 4-branch pure state, traces out crystal B in complex arithmetic, computes the purity Tr(rho_A^2) of crystal A's 2x2 reduced density matrix, and checks that the two routes agree within 1 part in 10^12. Here is where both hurried technical journalism and large language models routinely stumble. Seeing a peer-reviewed 2017 Physical Review Letters DOI, a measured 2021 gravitational constant from 90 mg gold spheres at 400 um surface gap (Tobias Westphal et al., DOI 10.1038/s41586-021-03250-7), and a deterministic script printing PASS at 10^-12, an automated summary will confidently write that gravity entangles micron crystals with concurrence 0.16. Every digit in that sentence is algebraically unimpeachable, and the sentence as physics is false.

That distinction matters more each year as scientific papers ship alongside executable notebooks, symbolic algebra scripts, and automated synthesis pipelines. A deterministic program guarantees reproducibility of inference: anyone who runs the same code on the same inputs recovers the identical floating-point value. Yet reproducibility of code is orthogonal to empirical realization in hardware, and conflating the two silently erases the laboratory.

II · Three Kinds of Absence: Unbuilt, Unresolved, and Derived

Scientific prose goes blurry whenever it uses a single hedge word—'estimated', 'modeled', or 'predicted'—to cover radically different reasons why a direct detector reading does not exist. Across computational physics and structural biology, there are 3 distinct kinds of empirical absence, and each carries a different obligation to reality. The first absence is an unbuilt or unobserved phenomenon, which our ledger caps at PROPOSED. Neither the 2017 Bose–Marletto–Vedral interferometer at 10^-14 kg nor the 2025 theoretical counter-arguments by Joseph Aziz and Richard Howl (DOI 10.1038/s41586-025-09595-7) or Sebastián Trillo and Miguel Navascués (DOI 10.1103/physrevd.111.l121101) represent laboratory observations; they are competing mathematical proposals about what classical or quantum fields would do if the apparatus existed.

The second kind of absence is physically present in the specimen yet experimentally unresolved (OPEN). This rule entered our engineering practice before Time & Gravity existed, while auditing structural coordinate files in Cystic Fibrosis ion-channel research. Examine the 4 landmark human CFTR Cryo-EM coordinate maps: dephosphorylated wild-type 5UAK (Fei Liu et al. 2017, DOI 10.1016/j.cell.2017.02.024), phosphorylated ATP-bound 5W81 (Zhenmei Zhang et al. 2017) and 6MSM (Zhang et al. 2018, DOI 10.1073/pnas.1815287115), and modulator-bound DeltaF508 structure 8EIQ (Karol Fiedorczuk & Jue Chen 2022, DOI 10.1126/science.ade2216). In wild-type 5UAK, 5W81, and 6MSM, residue Phe508 is clearly resolved inside nucleotide-binding domain 1, whereas 8EIQ lacks residue 508 because the DeltaF508 genotype genetically deletes that single phenylalanine from the polypeptide chain. By contrast, the roughly 200-residue disordered regulatory R-domain is physically tethered inside every intact channel in the vitrified ice, yet all 4 Cryo-EM structures leave those 200 residues unmodeled in atomic coordinates because conformational flexibility smears out the electron density. Structural biologists refuse to invent atomic coordinates where density blurs; teaching AI verifiers never to hallucinate smooth splines across those 200 unresolved residues—and never to confuse 8EIQ's genetic deletion at position 508 with Cryo-EM motion smearing—forged our strict separation between what an instrument resolved and what software inferred.

The third kind of absence is neither an unbuilt dream nor a blurred micrograph: it is a quantity deduced with mathematical rigor from a real, peer-reviewed measurement when the target configuration itself has not been directly observed. Suppose we take Vienna's 2021 measurement of Newtonian gravity between two 90 mg gold spheres at 400 um surface separation and combine it with quantum mechanics to compute the 0.31 rad phase of two 10^-14 kg crystals, or take Seattle's 2020 torsion-balance Yukawa table down to 52 um and project a 2025 error-function potential onto those 66 published wavelength rows. Calling those calculations PROPOSED would understate their empirical anchor, because a real pendulum already constrained the force law; calling them MEASURED would overstate what occurred in the room, because no sensor ever recorded that observable directly. Naming this third category DERIVED (NOT MEASURED) and requiring an explicit link to its MEASURED parent row resolves the dilemma cleanly.

III · The Recast Discipline: Why Every Derived Claim Needs a Measured Parent

The third kind of absence—a number rigorously deduced from a measured dataset under a new kernel that the original experimenters never fitted—came into sharp focus during the 2025 tabletop-gravity dispute. In Physical Review D in 2025, Trillo and Navascués integrated the Tilloy–Diósi hybrid master equation, in which gravity remains classical at the price of smearing Newton's point-mass potential over a regularization length sigma into V(r) = -G m_1 m_2 erf(r / (2 * sigma)) / r. In their geometry of 10^-15 kg spheres separated by 24 um with a 23 um split, that hybrid dynamics generates entanglement only below 0.8508 * sigma, requiring sigma above 55 um (and they plotted curves at sigma = 75 um, 100 um, and 250 um). A natural physicist's question immediately follows: if gravity is weakened by erf(r / (2 * sigma)) at tens of micrometres, what would have happened inside the 2020 Eöt-Wash short-range torsion balance?

In 2020 Junyi Lee, Eric Adelberger, Blayne Heckel, and colleagues at the University of Washington measured gravitational torque between 120-fold and 18-fold platinum wedge rings from 52 um out to 3.0 mm, limiting any gravitational-strength Yukawa term to ranges below 38.6 um at 95 % confidence (T05-LEE2020, DOI 10.1103/physrevlett.124.101101). Using the collaboration's own 2021 Fourier–Bessel expansion, our wedge kernel evaluates the smeared potential across those exact 54 um and 99 um foils: at 52 um, sigma = 75 um removes 17.0 % of the 120-fold Newtonian torque, whereas the Yukawa potential at the published 38.6 um edge removes only 2.5 %. Projecting the smeared torque deficit across 52 um to 3.0 mm onto the 66 signed 95 % Yukawa limits tabulated in Lee's 2020 dissertation yields an effective strength alpha_eff = -7.3 against the published signed limit -7.4 at sigma = 33 um, and alpha_eff = -10.2 against -7.4 at sigma = 35 um—excluding sigma > 33 um at 95 % confidence across 4 separation grids and both harmonics.

Why must that sigma > 33 um exclusion wear the badge DERIVED (NOT MEASURED) with an explicit data-tg-parent pointer to T05-LEE2020 rather than MEASURED or a standalone BOUND? Because the Eöt-Wash team fitted Yukawa exponentials exp(-r / lambda) to their 95 private torsion runs, never the error-function kernel erf(r / (2 * sigma)). By contrast, when Sandro Donadi and colleagues operated an underground high-purity germanium detector at Gran Sasso in 2021 (DOI 10.1038/s41567-020-1008-4) and recorded 0 excess spontaneous X-ray bursts to rule out parameter-free Diósi–Penrose collapse, that collaboration analyzed its own raw pulse shapes against that exact kernel, earning BOUND. And when the XENONnT Collaboration published its 1 to 140 keV electronic-recoil collapse search in 2026 (Elena Aprile et al., DOI 10.1103/2jm3-4976), tightening Diósi–Penrose limits by a factor of 5, our 24-month replication clock holds that young 1-collaboration run at SINGLE RESULT until independent time-projection chambers cross-check it.

IV · A Partial Order, Not a Ladder: Formal Types and Citation Polarity

Treat the 6 labels not as a beauty contest from 1 to 6, but as a typed partial order modeled on formal proof assistants. When the mathlib community formalized mathematics in Lean in 2020 (DOI 10.1145/3372885.3373824) and Joseph Tooby-Smith digitalized high-energy physics theorems in HepLean in 2025 (DOI 10.1016/j.cpc.2024.109457), every proved lemma carried its exact hypotheses inside its type signature. If you prove theorem B assuming axiom A, Lean's kernel verifies the implication A -> B with zero hand-waving, yet running #print axioms never lets you pretend A vanished. In Time & Gravity, data-tg-parent is that exact type-theoretic dependency edge: DERIVED (NOT MEASURED) certifies the implication from a mature MEASURED parent row while refusing to mint synthetic empirical data.

Type-checking internal deduction must also be paired with auditing external scientific disagreement. A paper can hold a valid Crossref DOI and still face sharp technical refutations in subsequent letters. Drawing on Josh Nicholson and colleagues' 2021 smart-citation classification architecture in Scite (DOI 10.1162/qss_a_00146), our ledger lint inspects supporting versus contrasting citation counts for every cached DOI. Whenever contrasting citation statements exceed supporting ones—as happens on hybrid classical-gravity channels challenged by short-range heating bounds—the build fails unless the ledger row carries a human-written contrasting_audit scope note of at least 20 characters explaining what remains disputed and capping the claim appropriately.

Notice why a partial order succeeds where a single numerical confidence score fails. Statistical language models are trained to maximize next-token likelihood across millions of abstracts, a training objective that naturally smooths away boundary conditions: a 95 % exclusion limit (BOUND), a 1-group cryogenic run published 8 months ago (SINGLE RESULT), and a closed-form integral evaluated to 10^-12 (DERIVED) get flattened into the same declarative verb tense. In a partial order, BOUND and DERIVED (NOT MEASURED) sit side by side at rank 4 without collapsing into each other. An exclusion bound rules out regions of parameter space where 0 signal appeared above background; a derived quantity propagates equations forward from a rank-5 MEASURED parent. Neither can be promoted to rank 5 by rhetorical polish or extra decimal places—only a 2nd independent laboratory apparatus operating outside the 24-month window can unlock MEASURED.

V · Testing the Guardrails: Eight Planted Epistemic Corruptions

A verifier that prints PASS on a curated repository proves very little until you try to fool it. Borrowed from safety-critical compiler testing, mutation analysis measures whether a rule suite actually kills plausible corruptions injected into the data. In the companion Python module for this essay, run_epistemic_mutations() takes a clean 11-row, 7-claim benchmark spanning tabletop gravity and structural coordinates and injects 8 targeted epistemic failures one at a time: (1) promoting a 2025 arXiv preprint to MEASURED; (2) deleting parent_id from the 0.16 concurrence derivation; (3) re-pointing that DERIVED (NOT MEASURED) claim from its 2021 90 mg MEASURED parent to the unbuilt 2017 PROPOSED witness; and (4) promoting the 2026 1-collaboration XENONnT run inside the 24-month window to MEASURED.

The remaining 4 mutations target subtler failure modes that routinely slip past human copy-editing: (5) interpolating the 200 unresolved flexible R-domain residues from OPEN into MEASURED coordinates; (6) misreading Gran Sasso's 2021 0-excess exclusion BOUND as a positive MEASURED collapse detection; (7) introducing bilingual numeric drift so the Spanish twin quotes 0.39 rad where English states 0.31 rad; and (8) erasing the 20-character contrasting_audit scope note on a 2016 hybrid-gravity row where 9 contrasting Scite citations outnumber 4 supporting ones. On a normal run, the lattice kills 8 of 8 mutants (recall 1.00); invoking --planted-defect deliberately relaxes parent-provenance checking so mutant 3 survives (7 of 8 killed) and forces a non-zero exit code.

Primary Sources & Offline DOI Ledger