INCOGNITA · TERRA

Scoreboard

As of .

The INCOGNITA engine is a port of TALYS-2 (A. J. Koning, S. Hilaire, S. Goriely) and uses the TALYS nuclear-structure database. What INCOGNITA adds is listed, switch by switch, at the end of the teaching page.

How to read this

Error: how far a code's values are from the measurements, as a root-mean-square in dex (log10 units; 0.01 dex ≈ 2.3 %).
Gap: the engine's error minus the rival's, on the same nuclei. Negative means the engine is closer to the data.
WIN: the whole 95 % interval of the gap is below zero. LOSS: the whole interval is above zero. TIE: the interval contains zero, so the data cannot tell the two apart. A tie is not a win.
Second scoring: the same measurements, with each code's curve read on a fixed 0.25 MeV grid. Second data set: another, partly overlapping selection of nuclei and measurements. Capture, nuclei far from stability: the compound nucleus has at least one proton fewer than the valley of stability at its mass.

What to keep in mind

This is not a blind test: the engine's global rules were chosen while watching these comparisons.
We ran the rival codes ourselves at their default input; their authors did not tune them.
Some comparisons reuse the same measurements (for example a second scoring of the same (n,2n) data), so the comparisons are not independent; no correction for multiple comparisons is applied.

1 · The engine against other reaction codes

The engine predicts every nuclide with one set of global rules: no parameter is adjusted to the cross sections of an individual nucleus. Like TALYS, it reads each nucleus's measured structure data (discrete levels, and resonance spacings where measured) from the TALYS structure database. Each comparison is one kind of measurement (for example “capture, 10–100 keV”) scored for the engine and one rival.

Every comparison that is not a win

By rival

How this is scored

Error

The root-mean-square of log10(code / measured), in dex. In most comparisons every measured point counts, with weight 1 / (0.1² + u²) for a point of uncertainty u (dex), so nuclei with many points weigh more; in fission and stellar capture each nucleus counts equally. Both codes are scored on the same points. The 0.1 dex in the weight is a floor: no point counts as more precise than about 26 %, so a few very precise points cannot dominate.

Interval

A paired bootstrap over nuclei: 2,000 resamples, with both codes scored on each; the interval runs from the 2.5th to the 97.5th percentile of the gap.

Not scored (N/A)

A comparison is N/A when the rival gives no value, when 5 or fewer nuclei can be scored, or when the rival's error exceeds 1 dex (a factor of 10: its run failed). The rule is the same whatever the verdict would have been; the data package still lists these gaps. The TERRA board keeps its own rule: N/A with 3 or fewer nuclei for plain accuracy, fewer than 5 otherwise.

Kinds of tie

Engine behind: the gap is zero or positive. Engine ahead, not resolved: the gap is negative but the interval contains zero; this label is given when a power estimate finds that at most four times as many nuclei would put the interval below zero. About level: the other ties.

Which rivals count

Only model calculations. Evaluated libraries are fitted to the same measurements, so the engine is not scored against them; TERRA is (section 2).

The rivals

TALYS-2.24: the TALYS program at default input, from runs we made and stored earlier. This project's own port of TALYS-2, the TALYS rival before 23 October 2026, matches the program to within 0.002 dex rms for capture, stellar capture and fission, and 0.013–0.023 dex for (n,p) and (n,α). The program's results are stored on its 64-point energy grid; for (n,2n), which rises steeply above threshold, the engine is read on the same grid (read on its own finer grid, it would win those four comparisons, mostly because of the grid). TENDL-style recipe: this project's reproduction, in its own engine, of the recipe TENDL uses for nuclei without fitted inputs; not the TENDL library. CoH3 3.6.1 and YAHFC 3.67: independent Hauser–Feshbach codes; by default YAHFC emits no alpha particles, so it has no (n,α) comparisons.

Stellar capture

Maxwellian-averaged capture cross sections (MACS) at 30 keV, grouped by proton number Z, scored on the nuclides whose compiled value is measured, not estimated. The bands Z < 20 and Z ≥ 84 hold no such nuclide and are not counted.

Overlaps

For example, “(n,2n) near threshold” and its second scoring use the same measurements; “capture, all” contains the narrower capture comparisons; each all-Z stellar-capture comparison contains its Z bands.

Blind test

None has been run on this engine yet. A blind test, on measurements kept sealed from development, is held back for later.

Terms

EXFOR: the international library of measured nuclear reaction data (IAEA and partners).
Evaluated library: a recommended set of cross sections for applications (ENDF/B, JEFF, JENDL, TENDL), built by fitting models to measurements.
MACS: Maxwellian-averaged capture cross section, the capture rate in a star at a given temperature.
Hauser–Feshbach: the statistical model of a compound nucleus decaying through every open channel; every code here uses it.
Surrogate measurement: a cross section inferred from a different reaction that forms the same compound nucleus.
Near threshold: energies just above the lowest energy at which a reaction can occur, where its cross section rises steeply.
Curated set, selected nuclei: smaller subsets of nuclei and measurements chosen by this project; the data package lists them.

What changed

23 October 2026. Comparisons resting on 5 or fewer nuclei, or on a failed rival run, are now N/A: three wins and two ties, all for fission, left the count. Stellar-capture bands with no scored nucleus (Z < 20, Z ≥ 84) are no longer counted as N/A. The TALYS rival is now the TALYS-2.24 program itself, at default input, in place of this project's port of TALYS-2. Every verdict is the same as against the port except four (n,2n) comparisons, which are now ties: the program's results are stored on a coarser energy grid, and with both codes read on that grid the two cannot be told apart. 22 October 2026. Two engine fixes (a numerical breakdown at very low neutron energies; memory use for heavy nuclei). No verdict moved.

Version

Check the numbers yourself: data and a script that recomputes every engine comparison

2 · The data product (TERRA) against evaluated libraries

TERRA anchors the engine's curves to good measurements where they exist, with an uncertainty band. It is compared with ENDF/B-VIII.1, JEFF-4.0, JENDL-5 and TENDL-2025 on accuracy and on whether its bands contain the measurements as often as they claim.

TERRA was fitted to exactly these measurements, while the libraries were evaluated with other, often older, data and with integral benchmarks. The comparison therefore favours TERRA; its results are not predictions.

How TERRA is scored

Accuracy uses the same error as section 1. Uncertainty coverage scores how well the bands' stated 68 % and 95 % match how often they contain the measurements. In every comparison lower is better for TERRA, and the verdict rule is the same.

TERRA's comparisons cannot be recomputed from the data package: that would need values derived from the evaluated libraries, which this project does not redistribute. Each comparison is listed there.