CLI and config files¶
Install the published package before using these commands.
The examples target VAMOS 1.4.0; use vamos --help or vamos <command> --help
for the parser in your installed version.
| Command family | Status in 1.4.0 | Use it for |
|---|---|---|
| Main optimization runner | Stable core invocation | One run or a configured problem set |
results inspect, results verify, reproduce |
Stable | Canonical run inspection, integrity, and exact replay |
study plan/create/run/inspect/resume/retry/summarize |
Stable, single-owner | Durable experiment matrices |
quickstart, summarize, open-results |
Experimental | Interactive onboarding and convenience helpers |
bench, tune, ablation, profile, zoo, studio, check |
Experimental | Optional workflows and diagnostics |
Stable commands have compatibility guarantees under the stability policy. Experimental workflows may change in a minor release. Dask, third-party integrations, and visualization remain Experimental even when selected through the main runner.
VAMOS 1.4.0: the experimental vamos assist command is
removed with the vamos.assist package, and VAMOS no longer calls any LLM
provider. Use vamos quickstart to write a config file, or
vamos study plan to check a study before running it.
Quickstart wizard¶
The Experimental wizard writes a config file and executes a single run:
vamos quickstart
If you are new to Python, start with Minimal Python.
List available templates:
vamos quickstart --template list
Run a template non-interactively:
vamos quickstart --template physics_design --yes --no-plot
Skip optional dependency warnings:
vamos quickstart --no-preflight
Template keys (short list):
demo: quick benchmark demo (ZDT1)physics_design: mixed-variable structural design (welded beam)bio_feature_selection: real-data feature selection (requiresexamplesextra)chem_hyperparam_tuning: SVM hyperparameter tuning (requiresexamplesextra)
The config is saved under results/quickstart/ and can be re-run with vamos --config <path>.
VAMOS 1.2.0: a seed whose run directory already exists cannot
be reused. Without --seed, quickstart takes the first free seed from the
template's (42, then 43, ...) and says so. With a taken --seed, it stops with
exit code 8 before writing its config. Its closing Re-run: line adds a free
--seed to the command. VAMOS 1.1.0 always used seed 42, so a second run, or
its own Re-run: command, optimized and then failed when saving.
Results helpers¶
The Experimental convenience helpers summarize or open run folders. For stable data access, use the canonical commands in the next section.
Summarize recent runs:
vamos summarize --results results
Show only the latest run:
vamos summarize --latest
Open the latest run folder:
vamos open-results --open
Canonical run commands (Stable)¶
Inspect one canonical run without materializing arrays:
vamos results inspect results/ZDT1/nsgaii/numpy/seed_7
vamos results inspect results/ZDT1/nsgaii/numpy/seed_7 --json
Fully verify integrity and exact replay compatibility without optimization:
vamos results verify results/ZDT1/nsgaii/numpy/seed_7
vamos results verify results/ZDT1/nsgaii/numpy/seed_7 --require-level exact
Execute a verified same-environment built-in replay as a new canonical run:
vamos reproduce results/ZDT1/nsgaii/numpy/seed_7
vamos reproduce results/ZDT1/nsgaii/numpy/seed_7 --output results/replays/zdt1-seed-7
Add --json to any of these commands for one machine-readable stdout
document. Replay never overwrites or modifies its source.
Durable studies¶
Resolve a durable study before creating a directory or executing an objective:
{
"problems": ["zdt1", "zdt2"],
"algorithms": ["nsgaii"],
"seeds": [0, 1],
"max_evaluations": 10000,
"pop_size": 80
}
vamos study plan study.json
vamos study plan study.json --output studies/comparison-01
vamos study plan study.json --json
vamos study create study.json --output studies/comparison-01
vamos study run studies/comparison-01
vamos study inspect studies/comparison-01 --json
vamos study summarize studies/comparison-01
vamos study summarize studies/comparison-01 --format csv --output artifacts/studies/tasks.csv
Planning is read-only: it creates no study, runs no task, and does not reserve
the proposed output. Its plan_id and task IDs match later Python
vamos.create_study(...) creation from the same StudySpec. Every study
command emits exactly one vamos.study-command-result version 1.0.0
document in JSON mode. Creation and execution are separate. inspect and an
in-memory summarize are read-only; JSON/CSV summary files are written only
when --output is explicit and never overwrite an existing path.
VAMOS 1.2.0: without --json, vamos study summarize prints
its task table: one row per task with its index, state, problem, algorithm,
seed and evaluations. Its suggested next step reads
vamos study inspect STUDY_DIR. In VAMOS 1.1.0 it printed only the study
header, without the table, and suggested a bare inspect.
Use vamos study resume STUDY_DIR for eligible pending/interrupted work and
vamos study retry STUDY_DIR --failed for explicit bounded failed-task retry.
Concurrent mutation is unsupported: one process must remain the only mutation
owner for a study. There is no cross-process cancel command; foreground Ctrl+C
uses graceful durable cancellation, and a cancelled study cannot be resumed
(see Resume and retry).
VAMOS 1.3.0: a Ctrl+C that arrives while the study writes a
journal event and the checkpoints that record it takes effect once those writes
finish, so the cancelled study always loads. Before, a Ctrl+C right after
execution_started could leave the root checkpoint behind the journal, and
vamos study run then failed with CANCELLATION_PUBLICATION_FAILED.
Main runner¶
Use vamos for single runs and problem sets.
Quick walkthroughs¶
Single run (default output under results/):
vamos --problem zdt1 --algorithm nsgaii --max-evaluations 5000 --population-size 80 --seed 7
Python equivalent (preferred for scripting):
from vamos import optimize
result = optimize("zdt1", algorithm="nsgaii", max_evaluations=5000, pop_size=80, seed=7)
After a single run, the runner (and vamos quickstart) prints
Hypervolume (reference [...]): ...; a command with several runs prints the
reference point under its summary table. That point comes from the fronts of
the same command: for each objective, the worst value among the runs of that
problem plus 10% of its magnitude (at least 0.1). Compare these values only
between runs of one command. Another command derives another point, so a longer
run can print a lower hypervolume than a shorter one. To compare runs of
different commands, compute the hypervolume against one fixed reference point,
as in Compute hypervolume.
Run a predefined continuous problem set with one compatible algorithm:
vamos --problem-set zdt --algorithm nsgaii --max-evaluations 3000 --seed 7
The families preset mixes real and permutation problems, so it is not
compatible with every algorithm selected by --algorithm both. Check the
capability matrix before
combining problem sets and algorithms.
Compare backends on one problem:
vamos --problem zdt1 --experiment backends --max-evaluations 2000
Install optional kernels with
python -m pip install "vamos-optimization[compute]==1.4.0".
The --experiment backends comparison skips unavailable optional kernels with
a warning. An explicit single-run --engine numba or --engine moocore
instead fails if that dependency is missing; it never switches silently.
Multiprocessing evaluation for expensive problems:
vamos --problem zdt1 --algorithm nsgaii --max-evaluations 8000 --eval-strategy multiprocessing --n-workers 4
VAMOS 1.2.0: the run keeps one pool of worker processes from
its first evaluation to its end, instead of starting one per generation, and
without --n-workers it uses the number of CPUs, at most 8. It only pays off
when each evaluation is expensive. See
Multiprocessing evaluation.
Enable Experimental live visualization and save plots (requires the
analysis extra):
vamos --problem zdt1 --algorithm nsgaii --max-evaluations 2000 --live-viz --plot
Early stop when hypervolume reaches a target fraction:
vamos --problem zdt1 --algorithm nsgaii --max-evaluations 15000 --hv-threshold 0.9
VAMOS 1.2.0: the target is a fraction of the reference front's hypervolume measured against the front's nadir plus 10% of its range in each objective (1.1 after normalization, as in pymoo and jMetal). VAMOS 1.1.0 and earlier added at least 5 to every objective, so a target of 0.5 was met by a random ZDT1 population; thresholds from those versions are not comparable.
Run every built-in algorithm on one problem:
vamos --problem zdt1 --algorithm both --max-evaluations 2000
VAMOS 1.2.0: the Experimental third-party baselines are
removed: --include-external, --external-problem-source, the
pymoo_nsga2, jmetalpy_nsga2 and pygmo_nsga2 algorithm names, and the
research extra that installed them. VAMOS 1.1.0 added those NSGA-II runs to
--algorithm both on ZDT1 when given --include-external.
VAMOS 1.2.0: a failed run no longer stops --algorithm both
or --problem-set. The error is logged, the remaining runs and the summary
complete, and the command then exits with a nonzero status that lists the
failed runs. RVEA derives its reference-vector partitions from
--population-size, so the default population of 100 runs on two objectives.
A population with no exact Das-Dennis lattice for the problem's objectives,
such as 100 on three objectives, is rejected before optimizing, and the error
names the nearest valid sizes (91 or 105).
Walkthrough: run and inspect outputs¶
1) Run a single optimization:
vamos --problem zdt1 --algorithm nsgaii --max-evaluations 5000 --population-size 80 --seed 7
2) Inspect the canonical artifact under results/ (default):
manifest.json: requested/resolved configuration, actual seed, outcome, provenance, and hashesresult.npz: objective, decision, constraint, population, and archive arraysenvironment.json: bounded runtime environment details
3) Save plots as presentation output outside the canonical run leaf. Use
another seed (or another --output-root): the run of step 1 already holds
results/ZDT1/nsgaii/numpy/seed_7, and a run directory is never overwritten.
vamos --problem zdt1 --algorithm nsgaii --max-evaluations 5000 --population-size 80 --seed 8 --plot
VAMOS 1.2.0: every run directory of the command is checked
before the first optimization. If one already exists, or two runs of the
command would share one, the command exits with code 8, the output-collision
code of vamos results and vamos reproduce, without optimizing anything.
VAMOS 1.1.0 optimized first and then stopped with a traceback and exit code 1
when saving, and a series stopped at the first collision.
Key flags¶
--algorithm: agemoea, ccmo, cmoeams, ctaea, ibea, moead, nsgaii, nsgaiii, rvea, smpso, smsemoa, spea2, top, or both--engine: numpy | numba | moocore | auto. The deterministic default isnumpy;autoprefers available Numba with a NumPy fallback. See the automatic engine policy for the selection and replay rules in VAMOS 1.4.0.--problem: any registry key (see Problems)--problem-set: predefined sets (e.g.,families)--validate-config: validate--configwith the checks that precede a run and exit--output-root: directory for run artifacts (default:results/)--no-preflight: skip optional dependency warnings--population-size,--offspring-population-size--max-evaluations--hv-thresholdand--hv-reference-front(VAMOS 1.3.0: a CSV path, orjmetalfor jMetal's packaged front of ZDT and DTLZ, see Reference fronts)--selection-pressure,--external-archive-size--eval-strategy: serial | multiprocessing (with--n-workers) | dask (Experimental; with--dask-address, see Dask)--live-vizwith--live-viz-interval,--live-viz-max-points--plot: save Pareto front plots after runs- Variation overrides per algorithm (examples):
--nsgaii-crossover sbx --nsgaii-crossover-prob 1.0 --nsgaii-mutation polynomial --nsgaii-mutation-prob 1/n--moead-crossover sbx --moead-mutation polynomial --moead-aggregation pbi--smsemoa-mutation polynomial --nsga3-crossover sbx
VAMOS 1.2.0: the --nsga3-* options and a defaults.nsgaiii
block of a config file reach NSGA-III; before, NSGA-III ignored them and used
SBX (probability 1.0, eta 20) and polynomial mutation (1/n, eta 20). An
operator parameter such as --nsgaii-crossover-prob needs the operator it
configures (--nsgaii-crossover sbx). Before running, the CLI also rejects a
negative --seed, a --n-var or --n-obj below 1, a different --n-obj or
--n-var on a problem with a fixed number of objectives or variables (see the
problem catalog), a --selection-pressure larger than the
population a problem runs with (after its problems.<name> overrides), and --track-genealogy with an algorithm that records no genealogy. It
warns about algorithm-specific flags that the selected algorithm ignores.
VAMOS 1.2.0: --track-genealogy (or track_genealogy in a
config file) works with NSGA-II, MOEA/D, SMS-EMOA, NSGA-III, SPEA2, IBEA and
SMPSO; AGE-MOEA and RVEA record no genealogy, so the CLI rejects it with them
and with --algorithm both. VAMOS 1.1.0 accepted it only with NSGA-II. In
the Python API, the genealogy of IBEA, MOEA/D and SPEA2 now follows the
survivors of each generation (in 1.1.0 it reported that no offspring reached
the final front), and NSGA-III and IBEA record both parents of each SBX child.
Config files (YAML/JSON)¶
Use --config path/to/spec.yaml; CLI flags override file values.
version: "1"
defaults:
title: My run
algorithm: nsgaii
engine: numpy
population_size: 120
max_evaluations: 20000
hv_threshold: 0.8
moead:
crossover: {method: sbx, prob: 1.0, eta: 20}
mutation: {method: pm, prob: "1/n", eta: 20}
problems:
bin_knapsack:
algorithm: nsgaii
n_var: 30
population_size: 150
nsgaii:
crossover: {method: uniform}
mutation: {method: bitflip, prob: "1/n"}
The moead block applies only when MOEA/D is selected, for example with
--algorithm moead.
VAMOS 1.2.0: a key repeated in the same mapping of a YAML or
JSON config file is an error; VAMOS 1.1.0 kept the last value without warning.
Keys under problems are case-insensitive, like the problem registry:
ZDT1 applies to zdt1, and two keys that differ only in case are an error.
In VAMOS 1.1.0 problems.ZDT1 passed validation but none of its overrides
reached the run. The stopping and archive blocks merge key by key across
defaults, the top level and problems.<name>; see
Stopping and archive.
VAMOS 1.2.0: algorithm, engine, eval_strategy,
external_archive_pruning, problem_set and experiment take the values of their command-line option, in defaults and
in problems.<name>. engine and eval_strategy may use any case, such as
NUMPY, as in VAMOS 1.1.0; the other keys need the exact spelling, such as
nsgaii. Any other value stops the command, and --validate-config, with exit
code 2 and an error that names the key, the value and the valid values. In
VAMOS 1.1.0 a config file skipped the check that
the command line makes: an unknown algorithm printed No runs were executed
and exited with code 0, an unknown engine or pruning policy stopped the run
with a traceback, experiment in a problem block was ignored, and --validate-config accepted most of them.
VAMOS 1.2.0: an operator that names the algorithm's default
operator keeps the default parameters it does not set. With NSGA-II,
crossover: {method: sbx, eta: 5} is SBX with probability 0.9 and eta 5, and
MOEA/D's crossover: {method: de, f: 0.7} keeps cr: 1.0. An operator with
another name gets exactly the parameters given, completed by that operator's
own defaults. Inside an operator, a
parameter set to null takes its default. A problems.<name> operator
replaces the defaults one whole, and an entry set to null there, or
repair: auto, keeps the value already in effect. In VAMOS 1.1.0 every
operator override replaced the default operator, so changing only eta also
lowered the SBX probability from 1.0 to 0.9, and the saved configuration did
not record it.
VAMOS 1.2.0: every field of an algorithm block reaches the
algorithm configuration, in defaults and in problems.<name> (which wins).
Examples include MOEA/D's neighbor_size, delta, replace_limit and
batch_size, NSGA-III's reference_directions and enforce_ref_dirs,
SMS-EMOA's reference_point and eliminate_duplicates, SMPSO's
archive_size, NSGA-II's steady_state (with
offspring_population_size: 1), and constraint_mode, result_mode,
initializer and mutation_prob_factor for every algorithm. They are
validated as in the Python configurations: an out-of-domain value in any
problem stops the command before any run, and fields that conflict with each
other stop it, with exit code 2, when the configuration is built.
Fields that the config sets elsewhere are rejected:
- pop_size and pop_size_auto: use population_size;
- offspring_size: use offspring_population_size;
- external_archive: use the archive block;
- track_genealogy: set it in defaults or the problem block;
- the Python callbacks generation_callback, generation_callback_copy and
parent_selection_filter.
In VAMOS 1.1.0, --validate-config accepted all of these fields and the run
ignored most of them without warning: for example, moead: {neighbor_size: 7}
ran with 20. For NSGA-II, MOEA/D, SMS-EMOA, NSGA-III, AGE-MOEA and RVEA, a
defaults block also lost every field that no command-line option reads (a
command-line option reads crossover and mutation, plus NSGA-II's repair and
MOEA/D's aggregation), its selection and repair included.
VAMOS 1.2.0: hv_threshold (or --hv-threshold) stops only
NSGA-II early. With another algorithm the run logs a warning and uses its full
evaluation budget; with --algorithm both, only the NSGA-II run can stop early.
Validate a config without running:
vamos --config configs/experiment.yaml --validate-config
VAMOS 1.2.0: --validate-config runs every check that precedes
the first run: the option values (for example a negative max_evaluations or
a crossover prob above 1), each selected problem's settings after its
problems.<name> overrides, its algorithm fields, and its stopping and
archive hooks. It prints Config OK. and exits with code 0 only when they
pass; otherwise it exits with code 2 and the error. An error about an option
whose value comes from the config file names its key, such as
defaults.nsgaii.crossover.prob. In VAMOS 1.1.0 it checked only the keys, so a
config could pass and then fail when run.
Run a config with a CLI override:
vamos --config configs/experiment.yaml --algorithm smsemoa --max-evaluations 10000
Other subcommands¶
The following commands are Experimental. Run vamos help for the full list.
Install any required optional extras first.
- Self-check:
vamos check - Benchmarking:
vamos bench --listandvamos bench ZDT_small --algorithms nsgaii moead --output report/ - Fast benchmark verification:
vamos bench ZDT_small --algorithms nsgaii --output report/ --smoke - Tuning:
vamos tune --instances zdt1,zdt2,zdt3 --algorithm nsgaii --backend optuna --backend-fallback random --split-strategy suite_stratified --budget 5000 --tune-budget 200 --n-jobs -1 - Ablation plans:
vamos ablation --config configs/ablation.yaml - Profiling:
vamos profile --problem zdt1 --engines numpy,numba --budget 2000 --output report/profile.csv - Problem zoo:
vamos zoo list,vamos zoo info zdt1,vamos zoo run zdt1 --algorithm nsgaii --budget 3000 - Studio (interactive, needs
studioextra): read the Studio walkthrough and launcher limitation before starting.
VAMOS 1.2.0: vamos --version prints the installed version.
vamos check prints one line per check, without the summary of each tiny run,
and a failed NumPy check exits with code 1 and a message instead of a
traceback. vamos zoo info with an unknown problem exits with code 1, vamos zoo
list names the encoding of real-valued problems real for all of them (some
said continuous), and vamos zoo run no longer prints HV: None: it
computes no hypervolume, which needs a reference point.
VAMOS 1.2.0: vamos profile runs each engine once untimed, to
pay one-time costs such as Numba compilation, then --repeats timed runs
(default 3). It reports their median, which the speedups use, and their range.
An unknown engine or problem stops the command with exit code 2 before any run.
An engine whose run fails is listed with its error, with empty cells in the
CSV, and the command exits with code 1. The hypervolume uses the feasible
solutions inside a reference point: the nadir of the problem's reference front
plus 10% of its range, or --hv-reference. It works with any number of
objectives. Without a matching reference front and without --hv-reference, it
is not computed and a warning says so. The CSV has the columns engine,
time_seconds (median), time_min, time_max, repeats, n_solutions,
hypervolume and error, and leaves a missing value empty. VAMOS 1.1.0 timed
a single cold run per engine, so the first Numba run measured mostly
compilation. It recorded a failed engine with time 0 and exit code 0. Its
hypervolume used the reference [1.1, ..., 1.1] whatever the problem's scale,
so the value was meaningless for a front outside the unit range (ZDT3,
DTLZ1, WFG) and for problems without a reference front. It was missing with
three or more objectives, and the CSV stored 0.0 where the table said N/A.
VAMOS 1.2.0: vamos bench and vamos ablation build the
indicator references per problem and number of objectives. The HV reference
point (hv_reference, with hv_source set to problem) and the reference
front for IGD, IGD+, epsilon and Hausdorff come from the non-dominated union of
that problem's runs in the same benchmark or ablation. The values are relative
to those runs, so do not compare them across problems or with another
benchmark. VAMOS 1.1.0 built one reference for the whole suite, which distorts
its indicator values and significance tests; rerun those benchmarks before
comparing them.
VAMOS 1.2.0: the indicators and their references use only the
feasible solutions of each run, those whose stored constraint values are all
G <= 0. A run without feasible solutions gets HV 0 and +inf for IGD, IGD+,
epsilon and Hausdorff, the worst values, and stays in the rank tests, which
count two such runs, or two such means, as a tie. The
n_feasible column counts the feasible solutions of each run. When a run
stores no constraint values, feasibility_known is false and every solution
counts as feasible: this is the case for unconstrained problems, and for
SPEA2 and SMPSO with constraint_mode: none. VAMOS 1.1.0 also measured
infeasible solutions, so its values for constrained problems are not valid.
VAMOS 1.2.0: vamos bench checks its inputs before any run:
--configtakes a JSON or YAML mapping with the keyspopulation_size,offspring_population_size,max_evaluations,seeds,engineandeval_strategy.max_evaluationsandseedsreplace the budget and the seeds of every experiment of the suite, also with--smoke. Any other key, or an invalid value, stops the command with exit code 2. VAMOS 1.1.0 ignoredmax_evaluations,seedsand unknown keys without a warning.- An unknown suite name exits with code 2 and suggests the closest names.
- Every study of the suite is planned before the first one runs. If a
study-NNNNdirectory of the suite already exists under--output, the command exits with code 5, asvamos study createdoes, and runs nothing. VAMOS 1.1.0 stopped with a traceback, after running the studies before the occupied one. --only-reportalso rewritessummary/metrics.csvwith the--metricsit is given, so it can recompute the indicators of a finished benchmark. Without studies of the suite under--outputit exits with code 2. VAMOS 1.1.0 leftmetrics.csvas the run wrote it, which could disagree with the tables.- The report draws its figures without
pyplot, so it no longer switches the process-wide matplotlib backend toAgg, and its plots work with every matplotlib version that VAMOS supports (3.7 or newer).
VAMOS 1.2.0: the statistics of vamos bench changed:
- The LaTeX tables bold the best mean of each problem. They mark an algorithm
whose runs differ from those of the best in a two-sided rank-sum
(Mann-Whitney U) test at
alpha:-if it is significantly worse and+if it is significantly better. Runs with the same seed are independent, so they are not paired. VAMOS 1.1.0 paired them in a signed-rank test, which cannot reach p < 0.05 with five seeds, so its tables carried no marks. - Few runs cannot show a significant difference: without ties, the smallest
two-sided p-value of the exact test is 0.1 with three runs per algorithm and
0.0079 with five. When such a pair cannot reach
alpha, the table caption says so and a warning is logged. With ties SciPy uses the asymptotic test, aslab/does, and its p-value can fall below that floor. - Only the hypervolume family is maximized. IGD, IGD+, the epsilon indicators
and the Hausdorff distances are minimized under every alias (
eps,epsilon_add,hausdorff,avg_hausdorff), in the tables, the rank tests andlab/. VAMOS 1.1.0 put the worst value in bold for some of these aliases. statistics.jsongains apairwiselist per indicator: a signed-rank test on the per-problem means for every pair of algorithms, also when there are only two, with its Holm-adjusted p-value (p_value_holm) and whether it issignificantatalpha.- A mean of +inf, from a run without feasible solutions, shows as
infwith no spread.
In lab/wilcoxon/Wilcoxon-<indicator>.csv, each cell holds one character per
problem, from a rank-sum test of the row algorithm against the column
algorithm: + means the row algorithm is significantly better, o
significantly worse, and - no significant difference. There, unlike in the
LaTeX tables, - does not mean worse.
Tuning quick notes (vamos tune)¶
Use this guide for quick usage. For the complete, maintained tune reference
(all backends, split/fallback behavior, finisher/validation/test, and artifact
contracts), see:
Recommended robust invocation:
vamos tune \
--instances zdt1,zdt2,zdt3,dtlz1,dtlz2,wfg1 \
--algorithm nsgaii \
--backend optuna \
--backend-fallback random \
--split-strategy suite_stratified \
--budget 5000 \
--tune-budget 200 \
--n-jobs -1
Quick verification path (built-in backend, tiny budgets):
vamos tune --instances zdt1,zdt2,zdt3,dtlz1,dtlz2,wfg1 --algorithm nsgaii --backend random --smoke --output-dir results/tuning_smoke
Ablation config example¶
algorithm: nsgaii
engine: numpy
output_root: results/ablation_demo
default_max_evals: 2000
problems: [zdt1]
seeds: [1, 2, 3]
base_config:
population_size: 60
offspring_population_size: 60
variants:
- name: baseline
- name: small_population
config_overrides:
pop_size: 40
- name: sbx_half
variations:
nsgaii:
crossover: {method: sbx, prob: 0.5, eta: 20}
summary_dir: results/ablation_demo/summary
The CLI writes a summary CSV by default to <output_root>/summary/ablation_metrics.csv (override with summary_path or summary_dir).
VAMOS 1.2.0: each variant writes its studies to <output_root>/<variant>/study-NNNN (--flat-output keeps them all in <output_root>), and a key of base_config or config_overrides that is neither an ablation setting nor a field of the algorithm configuration is rejected before any run. See Ablation planning.