# Methodology, scales, and limits

## What was built

1. **A source base.** A programmatic harvest of 817 arXiv records submitted between January 2024 and December 2026, drawn from 49 queries aimed at measurement systematics rather than at subject matter: "systematic error", "measurement bias", "unobservable", "blind spot", "cannot be distinguished", "degeneracy", "identifiability", "hidden variable", "nuisance parameter", "model discrepancy", "interlaboratory", "boundary condition", "contact resistance", "nonlinear impedance", "third harmonic", "small-signal approximation", "Kramers–Kronig", "instrument response", "deconvolution", "drift", "aging", "hysteresis", "model misspecification", and 26 others. The harvester is `src/harvest_arxiv.py`; the raw pack is `data/arxiv_pack.json`.
2. **A prior-art screen.** 358 unique prior Incoprea hypothesis titles were extracted from the existing export bundles, and every new title was compared against them by string similarity. Two substantive overlaps were found and replaced (see `CORRECTIONS.md`).
3. **Hypothesis construction.** 300 hypotheses were written directly, in six domain batches: physics and astronomy (50), chemistry and electrochemistry (50), materials (50), biology and medicine (50), engineering and Earth science (50), cross-cutting methodology (50). Each was required to have a named baseline, a mechanism, a discriminating observation, an available test, prior art, and an explicit refutation. The parser `src/parse_hyp.py` enforces the schema and rejects missing fields.
4. **Selection and refinement.** 30 were selected for development on three criteria: the mechanism had to be expressible quantitatively; the blind spot had to be invisible to the instrument or test that is normally used as a check; and the refutation had to be cheap. Refinement required labelling each part of the claim as an identity, an artefact, or a genuine prediction.
5. **Investigation.** Three were taken to full mathematical treatment with validated code: finite-window decay identifiability (I1), spot-size bias in thermal metrology (I2), and Kramers–Kronig blindness (I3). Every number quoted in the refined set is regenerated by `investigations/R1_refinement_numbers.py`.

## What "readiness" means

| Score | Meaning |
|---:|---|
| 5 | A reanalysis of published data, or a two-point change to an existing measurement, would settle it |
| 4 | Existing public data or an existing apparatus can confirm or kill it |
| 3 | Needs a dedicated measurement campaign or a new analysis pipeline |
| 2 | Needs new instrumentation |
| 1 | Needs the idea made precise before it needs an experiment |

Readiness is a judgement about cost, not importance. Of the 300: 45 are scored 5, 201 are scored 4, and 54 are scored 3.

## What "novelty" means

| Score | Meaning |
|---:|---|
| 5 | A mechanism not present in the literature as far as the source base shows, with a new observable |
| 4 | A known mechanism applied where it is not currently applied, with a quantitative prediction |
| 3 | A known mechanism given a specific, testable form in a new domain |
| 2 | A known effect quantified for a domain that has not quantified it |
| 1 | A restatement or a small extension |

Novelty 4–5 is claimed for 61 hypotheses; the rest are mostly 3. This is deliberate: most real progress in measurement science comes from taking a known mechanism seriously in a place it has not been taken seriously, not from inventing a new mechanism.

## The three labels used in the refined set

- **Identity.** True by construction. The logistic slope law in R01, the Bessel conductance ratio in R02, the closed-form temperature rise in R03, and the series-resistance formula in R25 are identities. They are useful and they are not discoveries. Presenting them as discoveries is the most common way this kind of work goes wrong.
- **Model artefact.** A property of the analysis rather than of the world. R07 (spurious ratio correlation), R09 (log-space bias), R21 (weighting), and R20 (error-bar deflation) are artefacts in this sense: they are real effects on reported numbers, produced entirely by the analysis.
- **Genuine prediction.** A statement that could be false. R01's claim that the offset is recoverable from curvature, R02's claim that the KK test cannot detect the bias, R03's claim that κ_app ∝ s² at large spot size, R11's frequency-dependence ratio, and R26's area-scaling collapse are all of this kind.

## Limits

- **No experiment was performed.** Every result is analytic or numerical. The collection contains no new measurement, and nothing here should be cited as evidence that a proposed effect was observed.
- **The source base is arXiv-weighted.** arXiv covers physics, mathematics, materials, and quantitative biology well and clinical medicine poorly. Hypotheses in the medical entries rest on general methodological literature rather than on a dedicated source harvest.
- **The prior-art screen is lexical.** String similarity catches phrasing overlap, not conceptual overlap. Two overlaps were found; others may remain.
- **Quantitative predictions are model-dependent.** The numbers in the refined set are computed from stated models with stated parameters. They are illustrations of magnitude and scaling, not predictions for named materials or systems. Where a number depends on an assumed parameter (for instance the substrate conductance G in I2), the assumption is stated.
- **The readiness scores are optimistic by design.** A 4 means someone could test it soon, not that the test is easy or cheap.
- **Adversarial reading.** Several entries take the contrarian side of a live dispute (P11 on the neutron lifetime, P20 on clock self-gravity, P36 on the proton radius). They are marked, and each carries the observation that would show it is wrong.

## Reproducing the collection

```bash
cd blindspot-2026-11-01
../.venv-research/bin/python src/parse_hyp.py            # rebuild CSV/JSON, validate schema
../.venv-research/bin/python investigations/I1_offset_window_bias.py
../.venv-research/bin/python investigations/I2_boundary_bias.py
../.venv-research/bin/python investigations/I3_kk_blindspot.py
../.venv-research/bin/python investigations/R1_refinement_numbers.py
python3 build/build_doc.py                                # assemble and typeset
```

The arXiv harvest is not required to rebuild the document; `data/arxiv_pack.json` is included. Re-running `src/harvest_arxiv.py` will produce a different pack because arXiv grows, so the pack shipped here is the one this collection was built from.