-- +goose Up
-- Seed the prior only requirement_time_to_fill model, a discrete time survival forecast of how long a staffing
-- requirement stays open before it is filled. It rides the SAME coefficient store as the other MAG models (no
-- new tables). Its OutputKind is opportunity_score: the served number is P(filled within a horizon) = 1 minus
-- the cumulative survival, an OPPORTUNITY level probability, never a per candidate score, so the frozen v1.4.0
-- engine stays the only per candidate scorer. The per month hazard vector is stored as one coefficient row per
-- month (hazard_month_1 .. hazard_month_12); serving reads them, forms the survival curve as the running product
-- of one minus the hazard, and reports the fill within horizon probability with the per month hazard list as the
-- interpretable evidence.
--
-- The prior only v1 hazard is a DEFENSIBLE constant: 0.25 per month, so the median time to fill is about two and
-- a half months (survival 0.75^t crosses 0.5 near month 2.4), a reasonable federal staffing baseline. Every
-- month carries the same 0.25 the day it ships, and a learned per month hazard lights up automatically as filled
-- and still open requirement durations accrue and cmd/train-model fits the empirical hazard. A discrete hazard
-- model needs NO calibration bins (the survival is exact from the hazards), so none are seeded.

INSERT INTO aedt_learned_models (model_key, version, family, output_kind, feature_spec, champion, notes) VALUES (
    'requirement_time_to_fill',
    1,
    'discrete_hazard',
    'opportunity_score',
    '{"features": [{"name": "hazard_curve", "source": "requirement_fill_durations", "horizon_months": 12}]}',
    true,
    'Prior only v1. A constant 0.25 per month fill hazard (median time to fill about two and a half months), a defensible federal staffing baseline. Learned per month hazards light up as filled and still open requirement durations accrue. OutputKind opportunity_score expressing P(filled within a horizon), never a per candidate score.'
) ON CONFLICT (model_key, version) DO NOTHING;

-- The 12 month baseline hazard curve, each month at the constant 0.25 prior. The intercept row is 0.0 (a
-- discrete hazard model has no bias term; it is stored only so the store shape matches the other models).
INSERT INTO aedt_model_coefficients (model_key, version, feature_name, coefficient) VALUES
    ('requirement_time_to_fill', 1, 'intercept',        0.0000000000),
    ('requirement_time_to_fill', 1, 'hazard_month_1',   0.2500000000),
    ('requirement_time_to_fill', 1, 'hazard_month_2',   0.2500000000),
    ('requirement_time_to_fill', 1, 'hazard_month_3',   0.2500000000),
    ('requirement_time_to_fill', 1, 'hazard_month_4',   0.2500000000),
    ('requirement_time_to_fill', 1, 'hazard_month_5',   0.2500000000),
    ('requirement_time_to_fill', 1, 'hazard_month_6',   0.2500000000),
    ('requirement_time_to_fill', 1, 'hazard_month_7',   0.2500000000),
    ('requirement_time_to_fill', 1, 'hazard_month_8',   0.2500000000),
    ('requirement_time_to_fill', 1, 'hazard_month_9',   0.2500000000),
    ('requirement_time_to_fill', 1, 'hazard_month_10',  0.2500000000),
    ('requirement_time_to_fill', 1, 'hazard_month_11',  0.2500000000),
    ('requirement_time_to_fill', 1, 'hazard_month_12',  0.2500000000)
ON CONFLICT (model_key, version, feature_name) DO NOTHING;

-- +goose Down
DELETE FROM aedt_model_coefficients WHERE model_key = 'requirement_time_to_fill';
DELETE FROM aedt_learned_models     WHERE model_key = 'requirement_time_to_fill';