Synthesis, scale and testing extraordinary claims

A useful claim risks being wrong

Specify the proposed structure, operating conditions, expected response and observations that would contradict it. Distinguish absence of a signal from an inadequate detection limit. Preserve raw data, calibration, sample identity and unsuccessful trials where they affect interpretation.

A slab conducts well after release from its formation conditions, and a shallow structural signal resembles the proposed phase. Does that establish a bulk material that retained the phase? Both observations can be real while sampling a thin surface region. The question is not whether the signals look impressive, but whether a specified competing material arrangement would produce them too.

All specimens, phase Q, geometries, response coefficients and observations below are invented teaching records. No new phase was synthesized or measured here. Predictions, targets and levels of evidence separates the earlier evidence stages; here we develop an entire bounded test, including which observation can contradict its claim and when a negative result lacks the required sensitivity.

Choose the claim, endpoint and controls before seeing the result

Define claim C precisely: a rectangular gauge contains a spatially uniform Q volume fraction q ≥ 0.90 through its thickness and has axial bulk conductivity σ ≥ 1.00 × 10⁶ S/m, at 300 K and 0.100 MPa, at the specifically observed elapsed times 10 and 60 minutes after release. The gauge length is L = 10.0 mm, width w = 5.00 mm and thickness b = 1.00 mm. This is a claim at two observation times, not continuous monitoring between them or a service-life guarantee.

Uniform here refers to a resolved local phase fraction: let f(z) be the Q volume fraction in a thin region at depth z from one broad face. C requires f(z) = q ≥ 0.90 in this depth description. It does not demand that every atomic position is Q. Direction, specimen state and spatial resolution belong to the claim. A calculation for an ideal Q lattice alone supplies none of these specimen observations.

Formation history can leave gradients. Local temperature, chemical environment, interfaces and constraints affect where a phase forms and which states persist. A supported surface region may survive while the interior transforms or never forms the same phase. Decompression, metastability and retained properties explains why a retained state and the path to it need separate evidence. No nucleation rate or recovery recipe is inferred here.

Before testing C, establish specimen identity, the chosen time origin and measured pressure/temperature history. Use independent composition and structural contrasts that can distinguish Q from the proposed alternatives in the actual sampled region. A single coincident peak is not unique phase identification. Microscopy, diffraction and structure evidence develops this distinction. The phase-selective response used below is a supplied calibrated model, not something guaranteed by calling an instrument a diffractometer.

A matched blank tests the non-Q/background response; a suitable positive reference tests that the method still responds in the intended geometry and state. Contact and path checks address transport. If a reference fails, a missing specimen signal is inconclusive under that method, rather than evidence that Q disappeared. Preserve the raw traces, corrections, reference observations and unsuccessful trials. Fix the primary falsification endpoint at 60 minutes before viewing the record; the 10-minute observation supplies time context, not an extra vote selected after the result.

A thin connected region can carry the axial current

Compare two hypotheses. H_bulk is a uniform conductor with σ_bulk = 1.20 × 10⁶ S/m. H_skin has Q only in two continuous broad-face layers, each thickness δ = 10.0 µm = 0.0100 mm; its core is an ideal electrical insulator. The supplied skin conductivity is σ_s = 6.00 × 10⁷ S/m. These are fictional transport parameters, not properties assigned to a real predicted phase.

For this model, the layers and core extend along the entire gauge length. Ideal end connections make each end face equipotential, with known gauge voltage ΔV and no included contact drop. The steady axial field is uniform, ΔV/L; side leakage, current crowding, heating and changes of material state are excluded. The two skins therefore form parallel paths. Each carries I_s = σ_s(wδ)ΔV/L, so total conductance is G_skin = I/ΔV = σ_s(2wδ)/L and R_skin = L/[σ_s(2wδ)].

The Q volume fraction is 2δ/b = 0.0200, or 2.00%. Yet the resistance is 1.66667 mΩ. If that reading is interpreted with the entire section wb as one homogeneous material, the apparent conductivity is σ_app = L/(Rwb) = σ_s(2δ/b) = 1.20 × 10⁶ S/m. H_bulk gives exactly the same resistance. The conductivity threshold passes under that homogeneous interpretation while 98% of H_skin is not Q.

This is not proof that the actual current took a surface route. It is an explicit alternative that the one resistance cannot exclude. Contacts, a connected impurity phase or another shunt can supply further explanations. A four-terminal gauge addresses particular contact drops, not every internal parallel path.

A prospective thickness comparison can separate the two stated models. Keep L, w, temperature, actual skin thickness and skin state fixed while changing b from 1.00 to 2.00 mm. H_skin predicts unchanged 1.66667 mΩ; H_bulk at fixed σ predicts half that value, 0.833333 mΩ. The same processing name does not guarantee the required matching: changed interior structure or layer thickness can change the prediction. A discrepancy tests the model together with those controls; it does not uniquely identify the shunt.

A structural signal weights the region that the probe reaches

Let a calibrated dimensionless response Y(d) have expectation equal to a depth-weighted phase fraction. The supplied weight is W(z;d) = exp(−z/d), where d > 0 is an effective sampling depth and exp is the exponential function. Then Y(d) = [∫₀ᵇ f(z)exp(−z/d) dz]/[∫₀ᵇ exp(−z/d) dz]. The denominator normalises the weight, not the specimen mass. A uniform f(z) = q gives Y = q at every d.

This deliberately limited response assumes known equal phase-response calibration, uniform attenuation, resolved Q contrast, linear detection and a laterally uniform central region away from side edges. Absorption and outgoing-signal losses are represented only through the supplied d. Refraction, coherent interference, texture changes, unequal response factors and changes of attenuation between phases are excluded. An actual scattering intensity is not generally a phase fraction. Changing angle, photon energy or preparation requires checking the relevant calibration and geometry, not merely relabelling d.

Depth-sensitive measurements can test a spatial alternative. In a 2009 study of thin HfO₂ films, Green and colleagues changed grazing-incidence scattering geometry to distinguish contributions from surface features and internal film structure. Their analysis included geometry, attenuation and comparison limitations. This is evidence that sampling depth can distinguish locations; our exponential two-skin model and its millimetre dimensions are not a reproduction of that experiment.

For H_skin, f = 1 in 0 ≤ z ≤ δ and b − δ ≤ z ≤ b, and zero in the core. Integrating each interval gives Y_skin(d) = [1 − exp(−δ/d) + exp(−(b − δ)/d) − exp(−b/d)]/[1 − exp(−b/d)]. The second pair of terms counts the far skin. Omitting it would describe a different specimen.

With b = 1.00 mm and δ = 0.0100 mm, the shallow condition d = 0.00200 mm gives Y_skin = 0.993262; the deeper condition d = 1.00 mm gives 0.0215899. The actual model volume fraction remains 0.0200 in both cases. A shallow near-unity response can therefore coexist with very little Q in the complete gauge. As d becomes much larger than b, the weight approaches uniform and Y_skin tends to 2δ/b.

Use the fictional mean observations in the table. Each mean comes from four independent reads of a stable state within its specified observation window. A known Gaussian read-noise standard deviation s₀ = 0.0400 gives mean standard deviation s = s₀/√4 = 0.0200. For this teaching model, background correction, calibration, geometry and d are exact, with no additional common error. These stringent assumptions enable the probability calculations below; they are not conclusions about a real instrument.

At both recorded times, the resistance and shallow response are consistent with a conducting surface-region alternative. The deeper mean is far below the homogeneous C prediction Y ≥ 0.90. The record can support a narrower local formation/persistence observation while challenging its proposed bulk interpretation, provided the identity and control checks are valid. It does not establish a unique chemical cause or a continuous lifetime.

A decision threshold and a detectable true signal are different quantities

Now ask whether a small calibrated response is present above a zero-response blank. Let η be its true expected response, not an automatically identified whole-specimen volume fraction. Each independent read obeys Y_i = η + ε_i, with ε_i normally distributed about zero, known standard deviation s₀ = 0.0400, and no drift or unmodelled common error. The mean of n reads then has standard deviation s = s₀/√n. The state must remain unchanged over the averaging interval. Do not clip negative estimates or estimates above one; that would change the stated sampling model.

Write Φ(x) for the standard normal cumulative probability, the probability that a zero-mean, unit-standard-deviation normal variable is at most x. It is obtained by integrating exp(−t²/2)/√(2π) from minus infinity to x. The positive quantile z₀.₉₉ = 2.326347874 satisfies Φ(z₀.₉₉) = 0.99. This numerical probability model is additional information, unlike using an unexplained multiple of uncertainty.

Declare one positive detection when the observed mean exceeds L_C = z₀.₉₉s. If η = 0, the chance of that decision is 1 − Φ(z₀.₉₉) = 0.0100. That is a one-sided false-positive probability for this predeclared comparison under this model. It is not the probability that a particular positive report is wrong, and it is not a current certification rule.

If the true response is instead η*, the probability of detecting it is P_det(η*) = 1 − Φ[(L_C − η*)/s] = Φ(η*/s − z₀.₉₉). This probability is often called power. A method can have a low false-positive probability and still miss most small real responses. A decision level acts on the observed mean; detection capability asks how often a specified true signal produces that decision.

For η* = 0.0200 and n = 1, L_C = 0.0930539 and P_det ≈ 0.0339. A nondetection is therefore expected in about 96.6% of repeated measurements even when this small response exists. It is weak evidence against that claim. With n = 64 independent reads, s = 0.00500, L_C = 0.0116317 and P_det ≈ 0.9529. These are repeated noise realizations of an unchanged state, not 64 independent syntheses.

A true response that reaches 95% detection probability in this model is L_D,95 = (z₀.₉₉ + z₀.₉₅)s, with z₀.₉₅ = 1.644853627. This follows by setting P_det = 0.95 and solving for η*. It differs from L_C. To detect η* = 0.0200 with at least 95% probability, require n ≥ [(2.326347874 + 1.644853627)0.0400/0.0200]² = 63.0818; the smallest whole number is 64.

The bulk claim uses the other tail. At the predeclared 60-minute endpoint, C implies η ≥ 0.90 for the calibrated deep response. Reject that homogeneous prediction when the observed mean is below 0.90 − z₀.₉₉s. With n = 4 this threshold is 0.853473; the supplied 0.0220 mean fails decisively. Under the stated Gaussian model the false-rejection chance is at most 1% for any η ≥ 0.90. This tests C and the measurement model together, not every possible Q structure or synthesis route.

If the normal law, known variance, reference transfer or independence fails, those tail probabilities no longer follow from these formulas. Common calibration error does not disappear by taking more reads. A blank that changes during processing, a state that decays during acquisition or a fitted response contrast needs a fuller model. Report that limitation rather than keeping the convenient 1% label.

Keep failed trials and independently manufactured objects in the record

Recomputing archived data can check analysis. Repeating the reading on the same object can check measurement behaviour. Preparing another specimen or transferring preparation to another laboratory tests additional manufacturing and material variability. Those achievements should not share one undifferentiated reproduction label. Comparable results and independent reproduction develops the metrological distinctions.

Consider the disjoint categories of twelve fictional preparation attempts in the table. Three have documented complete passes; seven have valid failures of a stated material/response criterion; two remain inconclusive because controls failed. The documented complete-pass fraction is 3/12 = 25.0% of all attempts. The fraction among only the ten validly classified attempts is 3/10 = 30.0%. Both calculations are possible, but they have different denominators. Omitting the unresolved runs would conceal part of the record. Neither fraction is a known probability of future manufacturing success.

Scaling requires defining the new object and useful function. Greater thickness, area or continuous length can change thermal/chemical gradients, interface fraction, defects, connections and process variability. A local signal, or an isolated strong constituent, does not by itself establish a continuous body with the same properties. A larger object also needs its own current path, specimen geometry and matched-condition property evidence.

A real example of a distinct scale accomplishment is the 2013 report by Behabtu and colleagues: its abstract describes manufactured macroscopic nanotube fibres and measured mechanical, thermal and electrical functions. That is a different evidence stage from an ideal individual-tube calculation. It is not evidence that every predicted constituent property transfers unchanged, nor a present-day universal production or application guarantee. None of that paper’s property numbers supplies our fictional phase Q.

Choose the next comparison to address the remaining alternative: depth or cross-sectional phase evidence for the bulk fraction, a thickness-dependent electrical prediction for the current path, a defined later-time comparison for persistence, and independent preparation for manufacturing transfer. No one successful measurement supplies all four. Evidence for macroscopic performance and limits continues that object-scale distinction.

Build the evidence ladder carefully

Synthesis, identification, repeated property measurement, persistence, scale-up and component use are separate accomplishments. Independent reproduction examines transfer beyond one preparation or laboratory. A candidate can remain promising while clearly stated requirements remain unmet.

A test is informative when a fixed claim predicts an observable that the method could distinguish from consequential alternatives. A matching number can leave the material identity or path unresolved; a negative result can be uninformative if the relevant response would usually be missed. Preserve which statement was challenged, at what conditions, time, spatial extent and error model.

A failed comparison does not diagnose misconduct. It can reveal an incorrect physical explanation, a changed specimen, a missing calibration term or a manufacturing limit. Examine traceable evidence and independent comparisons. The proper conclusion can retain a useful local achievement while rejecting the stronger bulk, duration or scale claim.

Worked examples

Try the reasoning

1. Does lack of independent reproduction prove fraud?

   Show the answer and reasoning

   No.

   It identifies an unresolved evidential step.

2. What makes a negative test informative?

   Show the answer and reasoning

   A method capable of detecting the claimed response under matching conditions.

   An insensitive or mismatched test may not challenge the claim.

3. Keep the model gauge and σ_s fixed, but let each surface layer be δ = 5.00 µm instead of 10.0 µm. Find Q volume fraction, R and σ_app. Does a halved phase amount have to halve every structural response?

   Show the answer and reasoning

   Q fraction is 1.00%, R = 3.33333 mΩ and σ_app = 6.00 × 10⁵ S/m. A depth-weighted response need not halve.

   The parallel conducting area halves, so G halves and R doubles. For d = 0.00200 mm, the shallow expectation is about 0.917915, still far above the 0.0100 whole-volume fraction; the exponential weighting differs from the electrical cross-section account.

4. For the original two-skin model, replace the exponential weight by a uniform weight W = 1 across the entire thickness. What does Y become? Does a measured shallow 0.993262 by itself establish that result?

   Show the answer and reasoning

   Y = (2δ)/b = 0.0200. The shallow measurement alone does not establish the whole-volume fraction.

   The uniform numerator counts the thickness of both layers and the denominator counts b. A different sampling kernel asks a different spatial average. Actual phase quantification also needs identity and response calibration.

5. Use n = 16, s₀ = 0.0400 and α = 0.0100. Find L_C and L_D,95. What is the detection probability when η* = 0.0200?

   Show the answer and reasoning

   s = 0.0100, L_C = 0.0232635, L_D,95 = 0.0397120 and P_det ≈ 0.3721.

   L_C is 2.326347874s. The true response required for 95% detection is (2.326347874 + 1.644853627)s. At η* = 0.0200, power is Φ(2 − 2.326347874); a low false-positive probability has not supplied high sensitivity.

6. Twenty independent, zero-response settings are each inspected with the same 1% false-positive rule, and only the largest positive result is reported. Under those exact independence assumptions, what is the chance that at least one passes? Is it still 1% for this search?

   Show the answer and reasoning

   1 − (0.99)²⁰ ≈ 0.182093, about 18.21%. The search has a different error probability.

   All twenty avoid a false positive with probability (0.99)²⁰; take the complement. Correlated settings need a different calculation. Preserve the full search and choose a comparison rule for the intended question rather than silently selecting one favourable result.

7. A 60-minute specimen gives no resolved Q response, but its supposedly stable positive reference also loses its expected response under the changed geometry. Has that observation refuted Q retention?

   Show the answer and reasoning

   Not under the stated test model. The method-control failure leaves this comparison inconclusive.

   The negative must be interpretable at the claimed response, depth and conditions. Investigate the reference transfer or geometry before attributing the missing signal to the specimen. This does not make a valid negative result permanently immune to a clear, capable test.

8. Phase-sensitive observations at 10 and 60 minutes both support a local Q region. May the report say every interior point was continuously Q throughout that hour, or that Q will survive one year?

   Show the answer and reasoning

   Neither extension follows.

   The sampled region, finite observation windows and tested duration remain part of the evidence. States could differ between observations or elsewhere, and longer retention needs a justified kinetic model or longer matched observations. Local persistence at the recorded times is still a meaningful result.

What to carry forward

• Make claims precise and falsifiable.
• Track demonstrated stages without collapsing them into one label.
• A current path and a depth-sensitive signal can be dominated by a small region; distinguish sampled response from whole-specimen phase content.
• Fix the falsification comparison and its controls before examining the record. A failed control can make a result inconclusive.
• A decision level controls a specified observed-data error; detection capability depends on the true response and the complete sampling model.
• Retain negative and inconclusive trials, and distinguish repeated reads, independent preparations, new scale and useful component evidence.