good_bad_grey — scoring algo evaluation

Ranking-first evaluation of the product scoring algo against a blind multi-LLM judge panel. Experiments collapse by algo content ({prompt+config} hash), so re-runs, branches, and merges of the same algo appear once.

Experiments (2) — agreement vs baseline

Baseline ⭐ = 33b6b6bc52cc (branch main). Each cell is this experiment's ensemble Spearman with the delta vs baseline; green = closer to the judge panel, red = further. #inv = consensus rank inversions.

experiment	oil_control	antioxidant	brightening	firmness_collagen	#inv
33b6b6bc52cc ⭐ main · 91fd5565 · 2 run(s)	0.524	0.802	0.583	0.466	5
b8e85befb448 exp/hero-active-benefit-bonus · f8ff93b8 · 1 run(s)	0.582 (+0.058)	0.779 (-0.023)	0.623 (+0.040)	0.478 (+0.012)	5 (+0)

Trend — ensemble Spearman across experiments

Experiment details

b8e85befb448

2026-08-11T13:39:36.185850+00:00 · clean · branch exp/hero-active-benefit-bonus · code f8ff93b8 · prompt 3c1cb0a2a9fbda8f · config cd59287fb260afee

benchmark serums-v1 v1 · judges: gpt-5.6-sol(inci), deepseek-v4-pro(inci), gpt-5.4-pro(full)

Ensemble ranking agreement

metric	spearman	ndcg@5	top5
oil_control	0.582	0.862	0.600
antioxidant	0.779	0.867	0.600
brightening	0.623	0.868	0.600
firmness_collagen	0.478	0.833	0.600

Consensus inversions

metric	product	algo rank	judge rank	direction
oil_control	1272752	7	18–20	algo over-ranks
oil_control	1252974	15	7–7	algo under-ranks
brightening	5003153	7	13–19	algo over-ranks
brightening	5003158	15	5–7	algo under-ranks
firmness_collagen	5003166	16	1–4	algo under-ranks

Full per-judge scorecard

metric	judge	spearman	kendall	ndcg5	top5	mean_drank	bucket_exact
oil_control	gpt-5.6-sol:inci	0.593	0.410	0.877	0.600	4.300	0.250
oil_control	deepseek-v4-pro:inci	0.570	0.396	0.841	0.600	4.600	0.200
oil_control	gpt-5.4-pro:full	0.514	0.364	0.843	0.600	4.650	0.200
oil_control	ensemble	0.582	0.396	0.862	0.600	4.450	0.200
antioxidant	gpt-5.6-sol:inci	0.694	0.554	0.906	0.600	3.300	0.600
antioxidant	deepseek-v4-pro:inci	0.749	0.567	0.833	0.600	3.400	0.400
antioxidant	gpt-5.4-pro:full	0.668	0.508	0.857	0.600	3.600	0.600
antioxidant	ensemble	0.779	0.607	0.867	0.600	2.800	0.700
brightening	gpt-5.6-sol:inci	0.669	0.474	0.926	0.400	3.900	0.350
brightening	deepseek-v4-pro:inci	0.526	0.381	0.817	0.600	4.400	0.150
brightening	gpt-5.4-pro:full	0.541	0.347	0.831	0.400	4.600	0.250
brightening	ensemble	0.623	0.438	0.868	0.600	4.200	0.250
firmness_collagen	gpt-5.6-sol:inci	0.556	0.413	0.886	0.400	4.200	0.300
firmness_collagen	deepseek-v4-pro:inci	0.425	0.324	0.825	0.600	5.200	0.300
firmness_collagen	gpt-5.4-pro:full	0.423	0.311	0.760	0.600	4.800	0.250
firmness_collagen	ensemble	0.478	0.354	0.833	0.600	4.650	0.300

33b6b6bc52cc

2026-08-11T13:24:18.616591+00:00 · clean · branch main · code 91fd5565 · prompt 3c1cb0a2a9fbda8f · config d778abee4d239730

benchmark serums-v1 v1 · judges: gpt-5.6-sol(inci), deepseek-v4-pro(inci), gpt-5.4-pro(full)

Ensemble ranking agreement

metric	spearman	ndcg@5	top5
oil_control	0.524	0.862	0.600
antioxidant	0.802	0.857	0.600
brightening	0.583	0.859	0.600
firmness_collagen	0.466	0.826	0.600

Consensus inversions

metric	product	algo rank	judge rank	direction
oil_control	1272752	7	18–20	algo over-ranks
oil_control	1252974	16	7–7	algo under-ranks
brightening	5003153	8	13–19	algo over-ranks
brightening	5003158	16	5–7	algo under-ranks
firmness_collagen	5003166	16	1–4	algo under-ranks

Full per-judge scorecard

metric	judge	spearman	kendall	ndcg5	top5	mean_drank	bucket_exact
oil_control	gpt-5.6-sol:inci	0.537	0.379	0.877	0.600	4.600	0.250
oil_control	deepseek-v4-pro:inci	0.518	0.364	0.841	0.600	4.900	0.200
oil_control	gpt-5.4-pro:full	0.452	0.333	0.843	0.600	4.950	0.200
oil_control	ensemble	0.524	0.364	0.862	0.600	4.750	0.200
antioxidant	gpt-5.6-sol:inci	0.719	0.554	0.899	0.600	3.400	0.400
antioxidant	deepseek-v4-pro:inci	0.760	0.579	0.820	0.600	3.350	0.300
antioxidant	gpt-5.4-pro:full	0.680	0.508	0.846	0.600	3.700	0.450
antioxidant	ensemble	0.802	0.607	0.857	0.600	2.800	0.500
brightening	gpt-5.6-sol:inci	0.630	0.442	0.923	0.400	4.000	0.500
brightening	deepseek-v4-pro:inci	0.493	0.337	0.805	0.600	4.600	0.350
brightening	gpt-5.4-pro:full	0.487	0.295	0.820	0.400	4.800	0.250
brightening	ensemble	0.583	0.385	0.859	0.600	4.350	0.250
firmness_collagen	gpt-5.6-sol:inci	0.549	0.406	0.882	0.400	4.200	0.350
firmness_collagen	deepseek-v4-pro:inci	0.412	0.306	0.815	0.600	5.350	0.250
firmness_collagen	gpt-5.4-pro:full	0.410	0.295	0.755	0.600	4.900	0.200
firmness_collagen	ensemble	0.466	0.337	0.826	0.600	4.800	0.250

Generated by eval.portal. Source of truth: eval/results/.