Calibration Constants

Every hand-tuned numeric constant in the engine, where it lives, what it does, and what changes when it moves. See docs/CALIBRATION.md for the long-form companion.

Module 1

Constant	Value	Where	Effect when changed
MAX_REASONABLE_BUDGET	1 × 10⁹	module1.py	Catches unit errors. Raise only if you legitimately run >£1bn campaigns.
test_and_learn_pct max	0.5	wizard_state.py	Cap on carve-out. Higher leaves the LP with nothing to do.
seasonality_index clamp	[0.1, 10.0]	module1.py	Out-of-range multipliers are usually unit errors.

Module 4

Constant	Value	Effect
CPU_OUTLIER_MULTIPLE	100.0	CPU > 100× the per-goal median is dropped. Lower = more aggressive filtering.

Module 5

Constant	Value	Effect
YIELD_BRACKETS	[(0.25, 1.0), (0.35, 0.65), (0.40, 0.35)]	Diminishing returns. Steeper drop-off → more diversification.
SHRINKAGE_KAPPA	30	Days at which a platform's own history weights 50/50 against the prior.
DEFAULT_MC_TRIALS	200	Monte Carlo default. Slider exposes 50–1000.
DEFAULT_INSTABILITY_CV	0.20	Allocation CV above this flags a platform as unstable.
Scenario multipliers	conservative 1/1.2, base 1.0, optimistic 1.15	Budget scaling per scenario.

Module 6

Constant	Value	Effect
DEFAULT_UNCERTAINTY_BAND	0.30	Fallback band when no observations and no history.
_REFERENCE_WINDOW_DAYS	30.0	The window the default band represents.
_MIN_BAND	0.05	Floor — below this is false precision.
_MAX_BAND	1.00	Ceiling — above this is useless.

Module 7 (Module7Policy)

Field	Default	Effect
corner_concentration	0.90	Top-platform share for "Corner-dominant" classification.
balanced_concentration	0.75	Top-platform ceiling for "Balanced".
corner_max_nonzero_cells	2	Max funded cells for a corner solution.
confidence_floor	40	Confidence never goes lower.
dq_small_kpi_share	0.50	Triggers data-quality flag when ≥50% of count-KPI forecasts are below threshold.
dq_small_kpi_threshold	5.0	Forecast count below this is "small".
plan_b_top_platform_cap	0.70	Diversification cap on Plan B.
plan_b_meaningful_tradeoff_pct	5.0	Trade-off (%) above which Plan B is surfaced prominently.
All confidence_*_penalty fields	various	Deductions from the 100-point starting score.

Changing a constant safely

1. Run pytest -q first to record the green baseline.
2. Make the change.
3. Run pytest -q again — note which tests fail. Most "regressions" are the engine doing what you asked.
4. Run PYTHONPATH=. python tests/behavioural_check.py and skim the end-to-end output for any sign the change broke the story the engine tells, not just the numbers.