Incognita
Open, calibrated predictions of neutron reaction cross sections across the chart of nuclides, with a public blind
benchmark and a registry of predictions made before the data.
At a glance
- Website: www.incognitanuclear.org.
Inside a Nuclear Reaction is a teaching walkthrough of what the engine
computes; Every nuclide shows the predicted curves for 6,168 nuclides
beside every curated measurement; the Scoreboard compares the
engine with other reaction codes, losses first.
- Scoreboard result (development comparisons against existing measurements, not a blind test): against the
TALYS-2.24 program, 10 wins, 9 ties, 0 losses; against all rivals together, 47 wins, 19 ties, 0 losses in 66 decided
comparisons (14 more are not applicable).
- Data package:
docs/walkthrough/data/board/ holds the per-nucleus rows behind
every scoreboard cell, and recompute.py recomputes the verdicts from them.
- Prediction registry:
predictions/2026-10-23/ holds predictions registered on
2026-10-23 for measurements not yet published, with a fixed scoring rule; 16 of the targets are blind and decide the
headline verdict.
- Engine: release tag
engine-2026-10-22.
- Library:
library/v0.2/ is the latest version, built with the scored engine (what changed:
CHANGELOG.md); library/v0.1/ stays available unchanged.
- Citing:
CITATION.cff (see Citing).
What is this
Most nuclei have never had a neutron cross section measured, yet reactors, nucleosynthesis networks, fusion shielding
and isotope production all need values for them. Incognita gives each nuclide a value, an uncertainty interval
whose coverage has been tested on nuclei held out of the fit, and a regime label (measured, anchored, interpolated
or extrapolated). The code, the predictions, the exam they were graded on and timestamped predictions for nuclei not
yet measured are all public, so every claim can be checked.
Highlights
| |
|---|
| Coverage | 6,168 nuclides (Z = 8–110), 1 keV – 20 MeV: capture, MACS, stellar rates, (n,p), (n,2n), (n,α) |
| Held-out capture | with measured D0, beats the default TALYS recipe on the sealed vault: 0.1519 vs 0.1712 dex |
| Tested intervals | 95 % interval covers 95.3 % of 674 sealed capture bins published after 2012 |
| Engine | a Python/C port of TALYS-2 physics; matches stock TALYS to four decimals on total, elastic, (n,p), (n,α) scores; 11.7x faster per target (no cache) |
Ties and losses are reported too: see the full benchmark.
Quick start
Needs Linux or macOS, Python 3.12 and uv.
git clone https://github.com/seth-michael-weaver/incognita.git incognita && cd incognita
uv sync --extra dev
uv run python -m incognita.library show Au-197 # capture curve with 68 / 95 % bands, MACS at 30 keV
bash scripts/reproduce.sh # regenerate the benchmark tables and compare with the shipped ones
Running the engine itself needs the TALYS structure database: see Running the engine.
What is in the repo
The full layout is in the guide.
Status and limitations
This is a v0.1 preview: research output, not an evaluation. Held-out tests are retrospective; the pre-registered
predictions have not yet been scored against new measurements. There is no claim of better accuracy than evaluated
libraries for measured nuclei; where good data exist, use the data or an evaluation. The library and its benchmark
numbers come from the v0.1 engine; the scoreboard uses the current engine (22 October 2026). Details: Status and limitations.
Learn more
Emulator
Millisecond emulators of the engine, with Jacobians and sampling: emu/README.md.
Citing
Use CITATION.cff, and also cite EXFOR (N. Otuka et al., Nucl. Data Sheets 120 (2014) 272) and
TALYS (A. Koning, S. Hilaire, S. Goriely, Eur. Phys. J. A 59 (2023) 131) when you use the engine or the EXFOR-derived
tables.
Licence
Code: Apache-2.0 (LICENSE, NOTICE); predictions and data: CC-BY-4.0
(LICENSE-DATA, including third-party terms); ported TALYS routines keep their MIT notice
(physics/hf/NOTICE-TALYS.md).
Contributing
See CONTRIBUTING.md and the Code of Conduct. If a number here is wrong,
please open an issue with the measurement.
