Wave 3 Conditioning Strategy Comparison

Generated 2026-07-06 | Source: post-fix wave 1 replay (10,000 designs × 1 seed × both designs)

Design B (Geography Frozen)

Strategy	N	Q min	Q p5	Q med	J pass%	P falls%	Both%	maint.bias shrinkage	maint.bias mean	attempt.bias shrinkage	identity shrinkage
`S0: Current (top-1000)`	1000	35.6	82.5	152.3	24%	19%	14%	47%	+0.33	3%	-0%
`SB: Top-200 composite`	200	35.6	65.8	126.8	62%	48%	40%	50%	+0.55	26%	14%
`SC: Union(500 STPM + J+P)`	1440	35.6	93.2	558.6	70%	69%	68%	42%	+0.72	9%	6%
`SD: Constraint-first`	500	50.6	133.2	548.0	100%	100%	100%	48%	+0.77	27%	17%
`SE: Near-flat+J, top-500 Q`	500	35.6	96.6	194.4	53%	36%	30%	50%	+0.59	29%	11%

NROY target zone (from wave 2 NROY designs): maintenance.bias mean = +0.63, attempt.bias mean = -3.74, identity mean = +1.29

Design A (Geography Free)

Strategy	N	Q min	Q p5	Q med	J pass%	P falls%	Both%	maint.bias shrinkage	maint.bias mean	attempt.bias shrinkage	identity shrinkage
`S0: Current (top-1000)`	1000	57.9	97.9	169.3	24%	14%	10%	30%	+0.11	3%	-1%
`SB: Top-200 composite`	200	57.9	75.7	126.6	52%	32%	27%	31%	+0.37	18%	8%
`SC: Union(500 STPM + J+P)`	1383	57.9	102.2	874.2	69%	68%	66%	31%	+0.39	7%	4%
`SD: Constraint-first`	500	73.0	140.5	983.4	100%	100%	100%	35%	+0.49	21%	16%
`SE: Near-flat+J, top-500 Q`	500	65.4	103.5	220.3	54%	25%	21%	31%	+0.31	23%	12%

NROY target zone (from wave 2 NROY designs): maintenance.bias mean = +0.63, attempt.bias mean = -3.74, identity mean = +1.29

Recommendation: Strategy SE for Wave 3

SE: Near-flat prevalence (δ < +2pp) + ≤1 Jackson failure, top-500 by Q_STPM

Best overall shrinkage: maintenance.bias ~50%, attempt.bias ~29%, identity ~11%. Current strategy (S0) achieves 47%, 3%, 0% respectively.
maintenance.bias mean closest to NROY zone: +0.59 (SE) vs +0.63 (NROY target) vs +0.33 (current S0).
Preserves Q quality: Same best Q as current strategy, but the distribution is more concentrated in the promising region.
Captures bridge designs: The near-flat relaxation (δ < +2pp instead of strictly falling) brings in designs that are close to both STPM-quality and constraint-passing — the bridging region where NROY lives.

Why not the alternatives?

S0 (current): Too lenient. 35 of 40 parameters have <5% shrinkage. Not learning.
SB (top-200): Good shrinkage but small pool (200) — posterior estimates may be noisy.
SC (union pool): Largest pool but bimodal posterior — Gaussian recentring doesn't capture the two modes. The prior would sit between the modes rather than in either.
SD (constraint-first): Strongest constraint signal but Q_STPM floor is too high (~51). Would concentrate wave 3 entirely in the constraint-passing regime, potentially missing the STPM-quality designs that are close to bridging.

Implementation

Run condition_wave1.py with:

python tools/calibration/condition_wave1.py   --wave1-root output/calibration_runs/20260703_0711_postfix_wave1   --design design_B_geography_frozen   --impl-dir postprocess/implausibility_prefixVs   --selection-mode top-k   --top-k-n 500   --w-stpm 1 --w-sts 0 --w-jackson 34 --w-prevalence 34   --jackson-band strict   --output-dir output/wave3_conditioning

But first filter the input pool to designs with prevalence_delta < 0.02 and jackson_fail_count ≤ 1, then rank by Q_STPM. This requires either a pre-filtering step or a new --pool-filter argument to condition_wave1.py.

Wave 2 NROY Design Characteristics (Design B)

30 NROY designs from wave 2 Design B (post-fix V_s scoring). These are the designs we want wave 3 conditioning to target — the region of parameter space that produces Q_STPM < 34.3.

NROY Design Parameters

Parameter	W1 Prior mean	W2 Prior mean	NROY mean	NROY sd	NROY range	W2 Prior sd	"NROY shrinkage"
`maintenance.bias`	-0.389	+0.480	+0.629	0.574	[-0.27, +2.29]	0.703	18%
`attempt.bias`	-3.644	-3.745	-3.742	0.347	[-4.73, -3.10]	0.425	18%

"NROY shrinkage" = 1 - (NROY sd / W2 prior sd). Shows how concentrated the NROY region is relative to the wave-2 prior.

Prior → Posterior → NROY Evolution

Key insight: The NROY region is well-localised in maintenance.bias (mean +0.63, much narrower than the W2 prior sd=0.70) and mNonSmokerSelfIdentity (shifted +0.6 SD from W2 prior mean). These two parameters, along with attempt.bias, should be the primary conditioning targets for wave 3. A strategy that concentrates the posterior around maintenance.bias ≈ +0.6 and identity ≈ +1.3 would maximise the chance of hitting the NROY zone.

NROY Constraint Satisfaction

Constraint	Pass count	Pass rate
Prevalence falls	13 / 30	43%
Jackson pass	17 / 30	57%
Both	10 / 30	33%

Even among NROY designs, only 33% pass both constraints. The structural tension between Q_STPM and prevalence is narrowing but not resolved.

Source: post-fix wave 1 replay, scored with pre-fix V_s. Generated 2026-07-06.