Replication package

"Pricing Risk at a Boundary" Romain Fillon

Overview

This package recomputes every number, figure and table of the paper from its inputs, with one command (bash run_all.sh). The manuscript itself is not part of this repository: values.tex (written as 02_output/values/values_generated.tex), the table files (02_output/tables/) and the figures (02_output/figures/) are its inputs, and the run writes them all.

The paper prices the risk of crossing a climate tipping threshold. The code solves the pricing problem numerically (boundary-value problems and closed forms), calibrates it to three climate elements by Monte Carlo over the published ranges of their inputs, estimates the recovery rate of the Atlantic overturning circulation (AMOC) from an observed sea-surface temperature record, and measures by simulation what that record can and cannot establish.

Exhibits of the paper and the code that produces them

The paper displays 5 figures and 4 tables. Numbers are as LaTeX prints them; labels are those of the manuscript, as listed in 02_output/paper_manifest.json.

Exhibit	Label	File	Produced by	Inputs
Figure 1	fig:two_boundaries	02_output/figures/fig_intro_two_boundaries.pdf	01_code/03_figures/fig_intro_two_boundaries.py	no data (closed forms)
Figure 2	fig:record_power	02_output/figures/fig_record_power.pdf	01_code/03_figures/fig_record_power.py	record-study outputs (the study reads the AMOC series)
Figure 3	fig:bifurcation_intro	02_output/figures/fig_bifurcation_intro.pdf	01_code/03_figures/fig_bifurcation_intro.py	no data (closed forms)
Table 1	tab:geometries	typed in the manuscript	—	no computed number
Figure 4	fig:crossover	02_output/figures/fig_crossover.pdf	01_code/03_figures/fig_crossover.py	AMOC series, via the recovery-rate estimate
Table 2	tab:disagreements	02_output/tables/disagreements.tex	01_code/04_tables_values/make_tables_values.py	cells are macros of values.tex; some cells need the AMOC series
Figure 5	fig:forest_tc	02_output/figures/fig_forest_tc.pdf	01_code/03_figures/fig_forest_tc.py	RCMIP emission paths; literature constants
Table 3	tab:eps_dyn	02_output/tables/eps_dyn.tex	01_code/04_tables_values/make_tables_values.py	cells are macros of values.tex; some cells need the AMOC series
Table 4	tab:inputs	typed in the manuscript	—	26 macros of values.tex

Every number the text quotes is a macro of values.tex (242 macros), written as 02_output/values/values_generated.tex by 01_code/04_tables_values/make_tables_values.py from 02_output/computed.json (the printed results of the scripts, collected by collect_computed.py) and from 01_code/inputs_literature.py (the sourced constants). Each macro carries, as a comment on its own line, the script and the inputs it came from; section 3 groups them by script.

The generated sections of this README (between <!-- BEGIN GENERATED ... --> markers) are written by 01_code/04_tables_values/make_inventory.py at every run, from what the paper displays (02_output/paper_manifest.json, see section 3) and the run's outputs; a hand edit there is overwritten by the next run.

---

1. Data Availability Statement

Statement about rights

The author has legitimate access to and permission to use all the data used in this manuscript. The one dataset whose host publishes no licence (the AMOC series below) is not redistributed; the package contains a script that downloads it from its publisher. The one third-party dataset that is redistributed (the RCMIP extract) is redistributed under its own licence, CC BY-SA 4.0.

Summary of the data sources

Data	Source and citation	Licence	In the package	Read by
AMOC SST fingerprint: subpolar-gyre sea-surface temperature index, annual, 1871–2016	Caesar, L., S. Rahmstorf, A. Robinson, G. Feulner and V. Saba (2018), "Observed fingerprint of a weakening Atlantic Ocean overturning circulation", Nature 556, 191–196, doi:10.1038/s41586-018-0006-5; file served by the Potsdam Institute for Climate Impact Research (PIK). Derived by those authors from HadISST: Rayner, N. A., et al. (2003), J. Geophys. Res. 108(D14), 4407, doi:10.1029/2002JD002670	None published by the host	No. Downloaded once, before the run, by 00_data/amoc/fetch_pik_sg_index.py	01_code/01_calibration/ews_calibration.py; 01_code/05_record_study/identification.py
CO2 emission paths SSP2-4.5 and SSP3-7.0, 2010–2500	Nicholls, Z. and J. Lewis (2021), Reduced Complexity Model Intercomparison Project (RCMIP) protocol v5.1.0, Zenodo, doi:10.5281/zenodo.4589756; Nicholls, Z. R. J., et al. (2020), Geosci. Model Dev. 13, 5175–5190, doi:10.5194/gmd-13-5175-2020; scenarios and extensions of Meinshausen, M., et al. (2020), Geosci. Model Dev. 13, 3571–3605, doi:10.5194/gmd-13-3571-2020	CC BY-SA 4.0	Yes, a two-row extract, 00_data/emissions/rcmip_co2_world_ssp245_ssp370.csv, with its checksums and the script that rebuilds it from the published file (extract_rcmip.py)	01_code/01_calibration/emission_paths.py
Outputs of the record-detectability study (eight JSON files)	This paper: written by the scripts of 01_code/05_record_study/	Same as the code	Yes, 00_data/record_study/; recomputed by the default run, or reused with RECORD_STUDY=cached	01_code/01_calibration/record_detectability.py (through 02_output/record_study/)
Values taken from publications	See the table below	—	Yes, as constants with their citations in 01_code/inputs_literature.py	the calibration scripts

The AMOC series (not redistributed)

The series is the file

https://www.pik-potsdam.de/~caesar/AMOC_slowdown/sg_index_hadisst.txt

of 1 638 bytes: one header line (# for the years 1871-2016 in K) and 146 annual values. The host page publishes no licence, copyright notice or conditions of use, so the package does not redistribute the file. A replicator downloads it once, before the run:

python3 00_data/amoc/fetch_pik_sg_index.py

The script writes the bytes unchanged to 00_data/amoc/sg_index_hadisst.txt and refuses them unless their md5 is the one the paper was computed on, c48516162322edebaaf113b2841f93be (recorded in 00_data/amoc/PROVENANCE.md). The scripts that read the file check the md5 again and stop with an explicit message if the file is missing or different. This download is the package's only external dependency.

If the URL stops resolving, the index can be rebuilt from HadISST with the subpolar-gyre region definition of Caesar et al. (2018). That reconstruction is not part of the package; a rebuilt series will have a different md5, and should match the summary statistics ews_calibration.py prints for the series itself: 146 values, variance of the detrended annual means 0.0683 K², lag-one autocorrelation 0.6106.

36 of the 199 computed macros in values.tex carry [requires network], alongside 43 sourced literature constants (242 macros in the file).

Those macros cannot be recomputed without the series. In the default mode the record study also needs it (identification.py, then calib82.py); with RECORD_STUDY=cached the study's shipped outputs are used instead, and the \rec* macros are recomputed from them without the series.

The RCMIP extract (redistributed under CC BY-SA 4.0)

00_data/emissions/rcmip_co2_world_ssp245_ssp370.csv is an adaptation (two rows, Emissions|CO2, World, scenarios ssp245 and ssp370, values unchanged) of rcmip-emissions-annual-means-v5-1-0.csv from the RCMIP protocol v5.1.0 by Zebedee Nicholls and Jared Lewis (doi:10.5281/zenodo.4589756), licensed under the Creative Commons Attribution-ShareAlike 4.0 International licence (https://creativecommons.org/licenses/by-sa/4.0/). The extract remains under CC BY-SA 4.0, and is not covered by the licence of the code. The published file (48 MB) is not shipped; 00_data/emissions/extract_rcmip.py rebuilds the extract from it and checks its md5. Details in 00_data/emissions/PROVENANCE.md.

Values taken from publications

Every such value is a constant of 01_code/inputs_literature.py, with its citation and, where the source has one, the table or section it is read from. The table below groups them by source; the citation strings in that file are the reference.

Source	Values	Where in the source
Armstrong McKay, D. I., et al. (2022), "Exceeding 1.5°C global warming could trigger multiple climate tipping points", Science 377, eabn7950, doi:10.1126/science.abn7950	Thresholds ΔT* (central, low and high ends) and timescales T (central and range) of the AMOC, the Amazon and the West Antarctic ice sheet	Table 1
IPCC (2021), Climate Change 2021: The Physical Science Basis, Working Group I contribution to the Sixth Assessment Report, Cambridge University Press, doi:10.1017/9781009157896	Transient climate response to cumulative CO2 emissions (TCRE): central 0.45, likely range 0.27–0.63 °C per 1000 GtCO2	Summary for Policymakers, D.1.1
Forster, P. M., et al. (2024), "Indicators of Global Climate Change 2023", Earth Syst. Sci. Data 16, 2625–2658	Observed warming above pre-industrial, 2014–2023 average, 1.19 [1.06 to 1.30] °C, rounded to 1.2	as cited in inputs_literature.py
Friedlingstein, P., et al. (2023), "Global Carbon Budget 2023", Earth Syst. Sci. Data 15, 5301–5369	Present global CO2 emissions, fossil and land use, about 40 GtCO2 per year	as cited in inputs_literature.py
Dietz, S., J. Rising, T. Stoerk and G. Wagner (2021), "Economic impacts of tipping points in the climate system", PNAS 118(34), e2103081118	Baseline SCC and the increase due to tipping points (Amazon, all elements)	Table 2
Hill, E. A., et al. (2023), The Cryosphere 17, 3739–3759, doi:10.5194/tc-17-3739-2023	Relaxation time of the West Antarctic ice sheet's grounding-line flux, 10 to 20 years	Section 4.1
Boulton, C. A., T. M. Lenton and N. Boers (2022), Nature Climate Change 12, 271–278	Recovery time of the Amazon vegetation index	Methods
Fuss, S., et al. (2018), Environ. Res. Lett. 13, 063002	Costs of direct air capture (backstop range)	as cited in inputs_literature.py
Howard, P. H. and T. Sterner (2017), Environ. Resource Econ. 68(1), 197–225	Ratio of their SCC to DICE-2013R	Abstract
Hambel, C., F. van der Ploeg and T. van den Bremer, replication package, Zenodo 20075290	Calibration constants of their tipping model, read from the printed output of their code (no code of theirs is included here)	Code_Parameters.m, Code_Initiation.m, as printed by MakeTable1.m

Constants marked "modelling choice" in inputs_literature.py (the discount rate, prior widths, the number of Monte-Carlo draws) are the author's and are stated there.

---

2. Computational requirements

Software

• Python 3.14.3 (CPython, 64-bit).
• The complete environment, with every package at its exact version, is requirements-lock.txt (14 packages). Use it to reproduce the outputs. 01_code/requirements.txt lists only the four direct dependencies (numpy 2.4.4, scipy 1.17.1, matplotlib 3.10.8, pandas 3.0.1), at the same versions.
• Bash, to run run_all.sh. On Windows, Git Bash.
• No LaTeX. Step 5/5 compiles the manuscript only when a manuscript/ directory sits next to the package, which is not the case in this repository; the step is skipped and says so.

Systems tested

• Windows 11 Pro, Git Bash, Python 3.14.3 (the author's machine, and a fresh clone in a fresh virtual environment built from requirements-lock.txt).

No other system has been tested. Nothing in the code is specific to Windows; on Linux and macOS run_all.sh calls python3 by default.

Hardware

Intel Core i7-1165G7 (4 cores, 8 threads, 2.8 GHz), 32 GB of RAM. The record study uses 6 worker processes. Disk: about 60 MB for the package and its outputs.

Runtime

Measured on the hardware below, from a fresh clone, wall clock of run_all.sh:

Mode	Command	Runtime
Full recomputation (default)	bash run_all.sh	1 h 25 min to 2 h 11 min (5,114 s and 7,854 s, two runs)
Record study reused	RECORD_STUDY=cached bash run_all.sh	9 min (549 s)

The record study accounts for almost all of the difference: its eight scripts took 4,622 s and 7,239 s in the two full runs (puissance.py 1,292–2,200 s, mle_ditlevsen.py 822–1,413 s, calib82.py 791–1,254 s, identification.py 572–1,169 s, horizon_np.py 649–705 s, horizon_attente.py about 460 s, the other two under 30 s each). The longest script outside it is noise_reversal_frontier.py (7 to 8 min). The faster full run is the clean-room run, in a fresh virtual environment built from requirements-lock.txt, with nothing else running. The two modes write the same values.tex, the same computed.json and the same tables, byte for byte; run_all.sh prints its per-step wall clock at the end.

---

3. Description of the programs

run_all.sh is the single entry point. It runs, in order:

1. Step 1/5, generators. Every script that produces a number, a figure or a study output is run once, and its standard output is captured in 02_output/values/raw_runs/<key>.out. In dependency order: the closed-form figures; the Monte-Carlo calibration (calibration_ci.py, eps_inf_ci.py, fig_forest_tc.py); the ceiling (bhp_bound.py); the record study (01_code/05_record_study/, eight scripts, or a copy of their shipped outputs with RECORD_STUDY=cached) and its reader (record_detectability.py, fig_record_power.py); crossover.py; the AMOC recovery rate (ews_calibration.py) and the four scripts that read it (reduced_form_scc.py, noise_reversal_frontier.py, fig_crossover.py, moving_budget.py); the numerical validation (validation_bvp.py).
2. Step 2/5. collect_computed.py parses the captured outputs into 02_output/computed.json. It runs nothing; a missing capture stops it.
3. Step 3/5. make_tables_values.py writes values.tex (as 02_output/values/values_generated.tex), the two table files (02_output/tables/) and 02_output/values/VALUES_DIFF.md (a per-macro comparison with the previous values.tex); make_inventory.py writes the generated sections of this README.
4. Step 4/5. check_exhibits.py checks that every file the paper includes exists and was written by this run.
5. Step 5/5. Compiles the manuscript, when it is present (see below); skipped otherwise.

The run exits non-zero if any step fails, and its last lines name the failed steps, the verdict of the record-study comparison and the time each step took. If step 2/5 or 3/5 fails, the run stops there with a non-zero code and writes nothing more, so that a failed run cannot leave behind the outputs of an earlier one under the appearance of a success.

With or without the manuscript. The manuscript is not part of this repository. What the code needs to know about it (the title, the figures and tables it displays, with their labels and sections, and the macros it cites) is in 02_output/paper_manifest.json, a file generated from paper.tex by 01_code/04_tables_values/paper_manifest.py and never edited by hand. The mode is set by the presence of the manuscript/ directory as a whole, never by a single file:

• Without manuscript/ (this repository): the run prints manuscript/ not present: using 02_output/paper_manifest.json; step 4/5 checks the exhibits under 02_output/; VALUES_DIFF.md compares with the committed 02_output/values/values_generated.tex; step 5/5 is skipped.
• With manuscript/ (the author's working copy): the run also copies values.tex, the tables and the figures into manuscript/, rewrites the manifest from paper.tex, checks that the two copies of each file are identical byte for byte, and compiles manuscript/paper.pdf.

The outputs under 02_output/ are the same in both modes.

Directory layout

Path	Contents
run_all.sh	The single entry point
requirements-lock.txt	The complete Python environment, exact versions
00_data/amoc/	The script that downloads the AMOC series, and its provenance; the series itself is not shipped
00_data/emissions/	The RCMIP extract (CC BY-SA 4.0), its provenance and the script that rebuilds it
00_data/record_study/	The shipped outputs of the record study, and their provenance
01_code/01_calibration/	Calibration: Monte Carlo, the AMOC recovery rate, the reduced-form level, the record-study reader
01_code/02_solvers/	The boundary-value problems and the price
01_code/03_figures/	The five figures
01_code/04_tables_values/	computed.json, values.tex, the table files, the paper manifest, this README's generated sections
01_code/05_record_study/	The record-detectability study
02_output/	Everything the run writes, including the inputs of the manuscript (values/values_generated.tex, tables/, the five figures in figures/) and paper_manifest.json

Every script

Script	Role	Needs
01_code/01_calibration/annual_mean_ou.py	module: annual means -> continuous OU parameters (imported by ews_calibration.py and the record study)	none (self-contained)
01_code/01_calibration/bhp_bound.py	the ceiling of ssec:ceiling under the floored hazard (eq:ceiling_body)	none (self-contained)
01_code/01_calibration/calibration_ci.py	Monte Carlo over the calibration inputs: budgets M* and their intervals	none (self-contained)
01_code/01_calibration/crossover.py	which disagreement moves the price (ssec:which_parameter): the \xover* macros	none (self-contained)
01_code/01_calibration/emission_paths.py	module: the deadline bound on SSP2-4.5 and SSP3-7.0 (RCMIP extract)	RCMIP extract (00_data/emissions/)
01_code/01_calibration/eps_dyn.py	module: the lower edge eps_dyn of the fixed-budget regime and the factor F (ssec:window)	none (self-contained)
01_code/01_calibration/eps_inf_ci.py	Monte Carlo: the deadline bound t_c of each element	none (self-contained)
01_code/01_calibration/ews_calibration.py	reads the AMOC series: recovery rate, variance and noise of the overturning circulation	AMOC series
01_code/01_calibration/phi_fold.py	module: the fold amplitude Phi of app:amp	none (self-contained)
01_code/01_calibration/record_detectability.py	reads the record-study outputs; the \rec* macros	record-study outputs (02_output/record_study/)
01_code/01_calibration/reduced_form_scc.py	the reduced-form price level (diagnostic macros)	AMOC series, via ews_calibration.py
01_code/02_solvers/moving_budget.py	the price while the budget is being spent (ssec:window, app:additivity)	AMOC series, via ews_calibration.py
01_code/02_solvers/noise_reversal_frontier.py	BVP: where the noise reversal of the theorem (ssec:gen_thm) ends	AMOC series, via ews_calibration.py
01_code/02_solvers/paper2_all_bifurcations.py	module: the BVP solvers of the three geometries (imported by validation_bvp.py)	none (self-contained)
01_code/02_solvers/price.py	module: the price at fixed state, eq:gen_scc, and its leading law	none (self-contained)
01_code/02_solvers/validation_bvp.py	BVP: the numerical validation of app:solver	none (self-contained)
01_code/03_figures/fig_bifurcation_intro.py	Figure fig:bifurcation_intro	none (self-contained)
01_code/03_figures/fig_crossover.py	Figure fig:crossover (imports crossover.py and eps_dyn.py)	AMOC series, via ews_calibration.py
01_code/03_figures/fig_forest_tc.py	Figure fig:forest_tc and the deadline macros on the emission paths	RCMIP extract (00_data/emissions/)
01_code/03_figures/fig_intro_two_boundaries.py	Figure fig:two_boundaries	none (self-contained)
01_code/03_figures/fig_record_power.py	Figure fig:record_power (imports record_detectability.compute())	record-study outputs (02_output/record_study/)
01_code/04_tables_values/collect_computed.py	step 2/5: collects the printed results into computed.json	none (self-contained)
01_code/04_tables_values/make_inventory.py	step 3/5: writes the generated sections of this README	none (self-contained)
01_code/04_tables_values/make_tables_values.py	step 3/5: writes values.tex, the two table files and VALUES_DIFF.md	none (self-contained)
01_code/04_tables_values/paper_manifest.py	step 3/5: what the paper displays, written to 02_output/paper_manifest.json from manuscript/ when present, read from it otherwise	none (self-contained)
01_code/05_record_study/calib82.py	record study: that test's false-rejection rate on the record -> calib82.json	AMOC series, via identification.py
01_code/05_record_study/compare_shipped.py	control: recomputed study outputs against 00_data/record_study/	none (self-contained)
01_code/05_record_study/estimateurs.py	record study: trend, level and ramp statistics -> estimateurs.json	none (self-contained)
01_code/05_record_study/fenetres.py	record study: the level statistic by window width -> fenetres.json	none (self-contained)
01_code/05_record_study/horizon_attente.py	record study: the bound while waiting -> horizon_attente.json	none (self-contained)
01_code/05_record_study/horizon_np.py	record study: the Neyman-Pearson bound -> horizon_np.json	none (self-contained)
01_code/05_record_study/identification.py	record study: joint region of the two trends on the AMOC record -> identification.json	AMOC series
01_code/05_record_study/mle_ditlevsen.py	record study: the Ditlevsen-Ditlevsen likelihood test -> mle_ditlevsen.json	none (self-contained)
01_code/05_record_study/puissance.py	record study: power of the trend estimators -> puissance.json	none (self-contained)
01_code/05_record_study/record_pipeline.py	module: the record study's trend estimators, the AMOC reader and the output paths	none (self-contained)
01_code/check_exhibits.py	control (step 4/5): every exhibit paper.tex displays was produced by this run	none (self-contained)
01_code/inputs_literature.py	module: the sourced constants, each with its citation	literature constants
01_code/tcre_draw.py	module: the TCRE draw shared by the three Monte-Carlo scripts	none (self-contained)

The macros of values.tex, by producing script

Produced by	Macros	of which [requires network]	Needs
01_code/01_calibration/bhp_bound.py	11	0	none (self-contained)
01_code/01_calibration/calibration_ci.py	6	0	none (self-contained)
01_code/01_calibration/crossover.py	6	0	none (self-contained)
01_code/01_calibration/eps_inf_ci.py	3	0	none (self-contained)
01_code/01_calibration/ews_calibration.py	5	5	AMOC series
01_code/01_calibration/record_detectability.py	53	0	record-study outputs (02_output/record_study/)
01_code/01_calibration/reduced_form_scc.py	2	2	AMOC series, via ews_calibration.py
01_code/02_solvers/moving_budget.py	8	8	AMOC series, via ews_calibration.py
01_code/02_solvers/noise_reversal_frontier.py	2	2	AMOC series, via ews_calibration.py
01_code/02_solvers/validation_bvp.py	30	7	none (self-contained)
01_code/03_figures/fig_forest_tc.py	25	0	RCMIP extract (00_data/emissions/)
01_code/04_tables_values/collect_computed.py	48	12	closed forms and sourced constants
01_code/inputs_literature.py	43	0	literature constants

---

4. Instructions to replicators

From a fresh clone, at the root of the package:

1. Get the AMOC series (once; the only step that uses the network):

   python3 00_data/amoc/fetch_pik_sg_index.py
   

   It prints the md5 and writes 00_data/amoc/sg_index_hadisst.txt, or stops if the file PIK serves differs from the one the paper used. On Windows, python3 may be the Microsoft Store alias rather than an interpreter; call the Python 3.14 of your installation by its path in steps 1 and 2 if so.

2. Create the Python environment:

   python3 -m venv .venv
   .venv/bin/python -m pip install -r requirements-lock.txt
   

   On Windows, the interpreter of the environment is .venv/Scripts/python.exe.

3. Run the package, with the interpreter of that environment:

   PYTHON=.venv/bin/python bash run_all.sh
   

   On Windows (Git Bash): PYTHON=.venv/Scripts/python.exe bash run_all.sh. To skip the recomputation of the record study, add RECORD_STUDY=cached in front of the command.

4. Check the result. The run ends with Failed steps: none, and with the number of the eight recomputed record-study outputs identical to the shipped ones. 02_output/values/VALUES_DIFF.md compares every regenerated macro with the committed values.tex (a clean reproduction reports 0 substantive), and git status shows whether any generated file changed: a clean reproduction changes none.

5. The paper. values.tex (02_output/values/values_generated.tex), the table files (02_output/tables/) and the figures (02_output/figures/) are the inputs of the manuscript, which is not part of this repository. Step 5/5 is skipped and says so.

Figures are reproduced to the byte. run_all.sh sets SOURCE_DATE_EPOCH=1767225600 (2026-01-01 00:00:00 UTC, a fixed constant rather than the date of any run), which matplotlib writes as the creation date of every PDF instead of the time of the run. A figure script run by hand without it draws the same picture but writes the current date, and so different bytes.

---

5. Notes

Random seeds

Every random draw is seeded, so a rerun returns the same numbers:

• 01_code/01_calibration/calibration_ci.py, 01_code/01_calibration/eps_inf_ci.py and 01_code/03_figures/fig_forest_tc.py: np.random.seed(42), set once before the draws; 01_code/tcre_draw.py draws right after it.
• The record study, 01_code/05_record_study/: one numpy.random.default_rng seed per simulated cell, stated in each script's header (identification.py 42, puissance.py 1, 2, …, mle_ditlevsen.py 6000+, calib82.py 8000+, estimateurs.py 1300+, fenetres.py 2000+, horizon_np.py 5000+, horizon_attente.py 7000+). The cells run on 6 worker processes; each cell carries its own seed, so the results do not depend on the number of processes.
• No other script draws random numbers (the solvers, ews_calibration.py and the figures are deterministic).

---

6. License

• Code. The code of this package (01_code/, run_all.sh and the scripts under 00_data/) is released under the MIT licence; see LICENSE.
• The RCMIP extract, 00_data/emissions/rcmip_co2_world_ssp245_ssp370.csv, is redistributed under its own licence, CC BY-SA 4.0, and not under the MIT licence. Its attribution and licence notice are in 00_data/emissions/PROVENANCE.md and in section 1 ("The RCMIP extract").
• The AMOC series of Caesar et al. (2018) is not redistributed: its host publishes no licence. 00_data/amoc/fetch_pik_sg_index.py downloads it from its publisher; see section 1 ("The AMOC series").