Functions
dict	load_spec (Path\|None path=None)

list[dict]	active_params (dict spec, int tier)

list[dict]	apply_exclusions (list[dict] params, list[str] exclude)

np.ndarray	to_physical (np.ndarray u, dict p)

	saltelli_rows (int N, list[dict] params, int seed)

str	git_sha ()

	main (argv=None)

Variables
	HERE = Path(__file__).resolve().parent

Detailed Description

make_design - the ONE canonical Saltelli design for the Tier-2/3 campaign.

Run this ONCE, on ONE machine, and commit the output.  Every machine in the
campaign reads the same file; no machine ever generates its own.

Two reasons this is not merely tidy:

  1. Saltelli indices are computed from PAIRED rows.  S_j pairs f(A_i) against
     f(AB_i^(j)).  If two machines independently sample, the pairing is
     destroyed and the indices are silently WRONG -- not noisy, wrong.

  2. Tier 3 is a maximin NESTED SUBSET of the Tier-2 design (plan S4/Tier 3).
     A Sobol sequence is nested by construction: the first M base points of a
     2^k sequence are themselves a balanced design.  Latin hypercube is NOT --
     it cannot be extended without rebuilding.  The shared theta points between
     LF and HF ARE the multi-fidelity discrepancy model, so they must coincide
     exactly, which is why theta is referenced by a stable row_id.

Design size: N base samples, p sampled parameters -> N*(p+2) rows per condition
(matrices A, B, and AB_j for j = 1..p).

Usage
-----
    python design.py --tier 2 --N 16 --out design/T2_design_v1.csv
    python design.py --tier 3 --N 4 --nest-from design/T2_design_v1.csv \
                     --out design/T3_design_v1.csv

Outputs a CSV with one row per model evaluation:
    row_id, matrix, base_idx, param_j, condition, <key_1> ... <key_p>
plus a sidecar .meta.json recording the seed, parameter version, git SHA and a
SHA-256 of the design itself.  run_ensemble.py refuses to run against a design
whose hash does not match its sidecar.

Function Documentation

◆ active_params()

list[dict] design.active_params	(	dict	spec,
		int	tier
	)

Parameters sampled at this tier.

Tier 2 -> tier==2 only.  Tier 3 -> tier 2 AND 3: the HF design must vary
everything the LF design varied *plus* the dose-driven block, otherwise the
HF runs are not comparable to their LF partners.

Definition at line 66 of file design.py.

◆ apply_exclusions()

list[dict] design.apply_exclusions	(	list[dict]	params,
		list[str]	exclude
	)

Drop parameters this campaign holds FIXED rather than samples.

Kept out of active_params/parameters.json deliberately: an excluded
parameter is excluded *by this design*, not deleted from the model, and it
keeps its prior and its documentation.  The exclusion is recorded in the
design's .meta.json, so a Sobol table can never be read as though the
parameter had been varied and found unimportant.

Rev 6 excludes i_mobile and v_mobile: bin_moment fixes them at 50 and 5
(they set the discrete/binned boundary, so sampling them would change the
DISCRETISATION between rows, not just the physics), and run_ensemble refuses
to run bin_moment against a design that still carries them.

Definition at line 78 of file design.py.

◆ git_sha()

str design.git_sha ( )

Definition at line 162 of file design.py.

◆ load_spec()

dict design.load_spec ( Path | None path = None )

Load parameters.json (or a parameters.yaml sibling if pyyaml exists).

Definition at line 51 of file design.py.

◆ main()

design.main ( argv = None )

Definition at line 171 of file design.py.

◆ saltelli_rows()

design.saltelli_rows	(	int	N,
		list[dict]	params,
		int	seed
	)

Return (rows, meta) where rows is a list of dicts.

Layout per base index i:  A_i, B_i, then AB_i^(j) for each j.
AB_i^(j) is A_i with column j replaced by B_i's column j.

Definition at line 127 of file design.py.

◆ to_physical()

np.ndarray design.to_physical	(	np.ndarray	u,
		dict	p
	)

Map u ~ U[0,1)^n onto the parameter's prior support.

Every family is inverse-CDF so the Sobol sequence's low-discrepancy
structure survives the transform -- which is the whole point of using one.

Definition at line 103 of file design.py.

Variable Documentation

◆ HERE

design.HERE = Path(__file__).resolve().parent

Definition at line 47 of file design.py.

Functions

Variables