Why
Ranking figures now can hold out --intermediate-budget, but the regret-vs-exploration curve still uses a single gate bootstrap stream (rng=123) and no error bars. Class crossover frequency is a normalized pair-participation rate, not a claim about a population of classes. Ensemble-disagreement figures exist; they still need the same budget-role contract as the audit.
What to do
- Pass the same
--budgets / --intermediate-budget / --fit-form as the audit.
- Add interval or replicate bands to regret-vs-exploration, or label it as a diagnostic sweep rather than a paper figure.
- Write
figure_metadata.json next to every PNG/PDF (CLI already started this).
- If Chinchilla figures cannot support the gate curve, skip and log that; do not draw a compute-power-law substitute.
Acceptance
Files
asla/figures.py
asla/cli.py
tests/test_figures.py
Do not fabricate
- Fill missing crossover roots with placeholder vertical lines
Why
Ranking figures now can hold out
--intermediate-budget, but the regret-vs-exploration curve still uses a single gate bootstrap stream (rng=123) and no error bars. Class crossover frequency is a normalized pair-participation rate, not a claim about a population of classes. Ensemble-disagreement figures exist; they still need the same budget-role contract as the audit.What to do
--budgets/--intermediate-budget/--fit-formas the audit.figure_metadata.jsonnext to every PNG/PDF (CLI already started this).Acceptance
results/audit.jsonFiles
asla/figures.pyasla/cli.pytests/test_figures.pyDo not fabricate