Decided 2026-11-10: (a). Audience, licence, hosting of the code
(a) personal and private until submission, open-source at submission; (b) open from the start; (c) stays private
(a): referees increasingly expect public code at submission, while an early public repo invites scooping of an unpublished idea. Licence at release: MIT or BSD-3 (common in astro codes).
Q2
Decided 2026-11-10: (a). When to contact Dullemond
(a) after M7 with the bias result and corrected estimator; (b) now, with this plan; (c) after M9 with a near-complete paper
(a): a concrete, checkable result (the bias explains the thesis’ slow evolution) is a strong opener, shows the work is serious, and lets him shape scope (porosity model, comparison with RPMC) before the expensive phases.
Q3
Decided 2026-11-10: filling factor φ. Second property for M9
Filling factor or volume: it is what the thesis and Dullemond’s notes motivate, and Schöll et al. (2026) in his group just argued it is too expensive as an independent dimension.
Q4
Journal
A&A vs J. Comput. Phys. vs ApJS
Decide with Dullemond; A&A default.
Q5
Decided 2026-11-10: public research tool at coag.mader.xyz, English only. Web app scope
research tool for us and the group vs public teaching tool
Research tool first (personal audience, no i18n or legal pages); a public version only if it turns into a teaching asset.
Q6
Decided 2026-11-10: bridge for oracles and figures. Python in the loop
only oracles and figures (ADR 0013) vs a full Python API for group members
Bridge for oracles and figures now; a polished Python API only if co-authors ask.
Sampling does not beat full summation even when unbiased and variance reduced (H5 false)
medium
high
A clean negative result with a cost model is still publishable as part of a methods paper; the bias analysis, Radau/W-method Jacobian study, and 2D cost surface stand on their own. Decide go/no-go for M9 at the end of M8.
X2
Linear solve dominates cost for p = 2 regardless of sampling
high
high
Planned from the start: E13 (sparse, approximate, matrix-free Jacobians).
X3
MC noise destabilizes implicit steps at steady state
medium
medium
Hybrid estimator, batch averaging, implicit sampled form, step-size control decoupled from noise (E11, E12).
X4
Porosity physics recipes are contested and outside our expertise
medium
medium
Use one published recipe verbatim (Okuzumi 2012) and keep it swappable; the paper is about numerics, not about the best porosity model.
X5
Overlap with other groups (TT methods, Gurrutxaga et al. 2026, RPMC improvements)
medium
medium
Track arXiv (astro-ph.EP, math.NA) monthly; position clearly (grid-based, implicit, unbiased). Early contact with Dullemond. The public website makes the work visible early; Vincent accepts this as low risk (Q22).
X6
Scope creep (global disk, transport, GPU)
high
medium
Non-goals in the vision; each extension needs an explicit milestone.
X7
Own Radau implementation has subtle bugs
medium
high
scipy parity tests on identical problems, order tests, step statistics comparison (M6).
X8
Parity with thesis code is hard where the thesis code has bugs
high
low
Parity tests reproduce the thesis behavior behind toggles (e.g. gas_velocity_viscosity = "molecular", estimator = "thesis"); fixes are separate toggles.
X9
Time (part-time project)
high
medium
Milestones produce standalone value (reproduction, web tool, bias note); paper can be split (method paper first, porosity application second).
X10
Machine load (16 GB shared Mac)
medium
low
heavy lock for studies; CI sizes small; offload large sweeps to a server only with explicit OK.
X11
Native f128 stays nightly-only longer than expected
high
low
Double-double is the stable extended type; f128 is an optional nightly feature (ADR 0016).