Manifund · cost–validity frontier

How much of the judgment do cheaper methods recover?

The Manifund page ranks proposals by thousands of pairwise LLM judgments. This companion eval asks what lexical features and frozen-embedding probes recover of that signal, per attribute — and how much the judges agree with each other in the first place. Forty proposals, one-line entities, rank statistics; directional, not decisive.

Findings (2026-08-16, n=40 — directional)

Ceiling. Inter-judge Spearman on the 32-battery: median 0.67 (gemma31b–gemma26b 0.70, gemma31b–gemini 0.69). Best-agreed: empirical testability 0.81, technical depth 0.79, downside risk 0.78. Worst: robustness to assumptions 0.35, counterfactual impact 0.45. The 1,001 subtle attributes replicate within-model at median 0.46 (IQR 0.35–0.56) at the 240-comparison budget — half the battery's reliability; a campaign finding, not a method one.
Structure, not halo. PC1 explains 31% (32-battery), 22% (subtle). PC1 loads + on existential seriousness, long-run impact, ambition, downside risk, AI-timeline sensitivity and − on tractability, feasibility, maturity, robustness: an ambitious-speculative vs tractable-grounded axis. Judges are not collapsing to one quality factor.
Embedding probes (LOO ridge, 40 items) reach 0.5–0.6 on epistemic rigor (0.61), speed to first result (0.62), technical depth (0.53), testability (0.51) — roughly 75–85% of the inter-judge ceiling on those; near zero on neglectedness, fundability, counterfactual impact. embeddinggemma on the one-liner is the most frequent best variant; descriptions did not reliably help.
Lexical single features top out ~0.5 and the winner is usually word count; multi-feature ridge is weak (≤0.48). Line-length correlates 0.5–0.56 with gold on quantitative reasoning, concreteness of milestones, clarity of ToC, information value — flagged. Longer blurbs may genuinely carry more concrete content, or judges reward length; the full-description phase decides.
Cost. ~12 pairwise LLM calls per entity-attribute vs one local embedding forward pass vs free — the frontier is steep, and the probe arm's validity is real on the epistemic-content attributes.

Open: rerun when the 1,269-pool phases land (pull_gold.py); add an absolute-scoring cheap-LLM rung; test whether the probe generalizes across pools.

Full report

Forty one-line proposal entities; 38/40 matched full descriptions. Rank correlations are descriptive at n=40: uncertainty intervals would be wide, so no p-values are presented. Description variants use the entity line for the two unmatched proposals except description-length correlations, which use the 38 matches only.

a. LLM-judge reliability (validity ceiling)

Attribute	31b–26b	31b–Gemini	26b–Gemini	Panel median
existential seriousness	0.785	0.604	0.724	0.724
technical depth	0.882	0.732	0.789	0.789
epistemic rigor of the underlying theory of change	0.813	0.690	0.690	0.690
counterfactual impact of a marginal dollar of funding	0.702	0.453	0.452	0.453
neglectedness of the problem the project addresses	0.665	0.543	0.585	0.585
tractability of the plan as proposed by this specific team	0.712	0.568	0.629	0.629
expected long-run impact on humanity's trajectory	0.849	0.686	0.655	0.686
scale of the problem the project addresses	0.818	0.685	0.505	0.685
empirical testability of its core claims	0.806	0.799	0.836	0.806
quality of quantitative reasoning in the proposal	0.787	0.766	0.748	0.766
legibility of the team's track record	0.695	0.624	0.624	0.624
feasibility given the stated team and resources	0.634	0.751	0.596	0.634
concreteness of proposed milestones	0.779	0.741	0.753	0.753
clarity of the theory of change	0.642	0.698	0.516	0.642
speed to a first observable result	0.766	0.713	0.691	0.713
room for more funding in this exact niche	0.477	0.502	0.438	0.477
cost-effectiveness per dollar spent	0.490	0.467	0.534	0.490
downside risk if the project succeeds at the wrong thing	0.806	0.753	0.779	0.779
probability of a net-negative outcome	0.490	0.771	0.601	0.601
robustness of the plan to its key assumptions being wrong	0.349	0.541	0.208	0.349
ambition	0.671	0.752	0.579	0.671
weirdness relative to mainstream research funding	0.717	0.712	0.728	0.717
interdisciplinarity	0.701	0.759	0.639	0.701
degree to which outputs are public goods	0.706	0.589	0.681	0.681
information value of simply running the project	0.747	0.724	0.586	0.724
replaceability by work others would do anyway	0.571	0.484	0.469	0.484
urgency of doing this now rather than in five years	0.811	0.659	0.550	0.659
maturity of the underlying field	0.660	0.639	0.454	0.639
sensitivity of the project's value to AI timelines	0.632	0.745	0.513	0.632
clarity of the writing itself	0.664	0.720	0.483	0.664
fundability by mainstream institutions	0.602	0.573	0.373	0.573
potential to become financially self-sustaining	0.673	0.790	0.619	0.673

Pairwise medians: gemma31b–gemma26b 0.701; gemma31b–gemini 0.694; gemma26b–gemini 0.598. Overall panel-pair median 0.672. Seed-2 file contains 293 attributes; only 284 overlap the main run, not the stated 292. Finite, non-constant test–retest correlations: 280/284; median 0.464, IQR 0.349–0.559.

b. Halo/general-factor check

32-attribute gemma31b PCA: PC1 31.4%, PC2 17.6% of variance. PC1 sign is oriented so loadings sum positive.

Attribute	PC1 loading
existential seriousness	0.244
technical depth	0.136
epistemic rigor of the underlying theory of change	0.073
counterfactual impact of a marginal dollar of funding	0.131
neglectedness of the problem the project addresses	0.169
tractability of the plan as proposed by this specific team	-0.249
expected long-run impact on humanity's trajectory	0.269
scale of the problem the project addresses	0.206
empirical testability of its core claims	-0.057
quality of quantitative reasoning in the proposal	0.087
legibility of the team's track record	-0.098
feasibility given the stated team and resources	-0.229
concreteness of proposed milestones	-0.035
clarity of the theory of change	-0.126
speed to a first observable result	-0.185
room for more funding in this exact niche	0.228
cost-effectiveness per dollar spent	-0.086
downside risk if the project succeeds at the wrong thing	0.272
probability of a net-negative outcome	0.214
robustness of the plan to its key assumptions being wrong	-0.230
ambition	0.259
weirdness relative to mainstream research funding	0.176
interdisciplinarity	0.127
degree to which outputs are public goods	0.002
information value of simply running the project	0.126
replaceability by work others would do anyway	-0.185
urgency of doing this now rather than in five years	0.151
maturity of the underlying field	-0.231
sensitivity of the project's value to AI timelines	0.258
clarity of the writing itself	-0.156
fundability by mainstream institutions	-0.006
potential to become financially self-sustaining	0.079

Subtle PCA: PC1 21.7%, PC2 11.4%. The source has 37,478/40,040 finite cells; missing cells were mean-imputed after per-attribute z-scoring, and 995/1,001 non-constant attributes entered PCA.

c. Embedding-probe arm

All probes use standardized embeddings, ridge alpha=1.0, and LOO predictions. The description variants use full descriptions for 38 entities and line-text fallback for two.

Attribute	voyage4nano_line/gemma31b	voyage4nano_line/gemma26b	voyage4nano_line/gemini	voyage4nano_desc/gemma31b	voyage4nano_desc/gemma26b	voyage4nano_desc/gemini	gemma_line/gemma31b	gemma_line/gemma26b	gemma_line/gemini	gemma_desc/gemma31b	gemma_desc/gemma26b	gemma_desc/gemini
existential seriousness	0.332	-0.037	0.329	0.328	-0.025	0.199	0.456	-0.150	0.351	0.025	-0.036	0.261
technical depth	0.477	0.551	0.545	0.486	0.494	0.546	0.527	0.709	0.669	0.486	0.587	0.608
epistemic rigor of the underlying theory of change	0.604	0.564	0.525	0.609	0.633	0.546	0.437	0.530	0.624	0.565	0.752	0.754
counterfactual impact of a marginal dollar of funding	-0.002	0.143	0.034	-0.134	0.311	0.211	0.214	0.279	0.250	0.045	0.303	0.198
neglectedness of the problem the project addresses	0.020	0.121	0.230	-0.105	0.100	-0.029	-0.234	-0.034	-0.002	-0.054	0.022	-0.073
tractability of the plan as proposed by this specific team	0.287	0.118	0.335	0.010	-0.052	0.384	0.203	0.239	0.370	0.016	0.092	0.274
expected long-run impact on humanity's trajectory	0.419	0.487	0.480	0.164	0.214	0.400	0.337	0.418	0.590	0.216	0.369	0.505
scale of the problem the project addresses	0.103	0.250	0.097	0.113	0.148	0.167	0.456	0.610	0.202	0.005	0.227	0.070
empirical testability of its core claims	0.369	0.334	0.446	0.409	0.453	0.577	0.452	0.635	0.749	0.507	0.436	0.704
quality of quantitative reasoning in the proposal	0.251	0.256	0.436	0.307	0.273	0.401	0.300	0.416	0.530	0.316	0.160	0.324
legibility of the team's track record	-0.013	-0.090	-0.310	0.204	-0.075	-0.072	0.324	0.349	-0.171	0.148	-0.163	-0.115
feasibility given the stated team and resources	0.426	0.146	-0.011	0.445	0.031	0.179	0.339	-0.027	0.054	0.505	0.198	0.089
concreteness of proposed milestones	0.090	0.172	0.087	0.094	0.242	-0.011	0.446	0.506	0.456	0.172	0.259	0.071
clarity of the theory of change	0.452	0.272	-0.026	0.022	0.083	0.032	0.539	0.319	0.299	0.294	0.134	0.420
speed to a first observable result	0.453	0.011	-0.096	0.459	0.240	0.305	0.619	0.068	0.088	0.521	0.365	-0.032
room for more funding in this exact niche	0.218	0.444	0.187	-0.219	0.361	0.168	-0.002	0.594	0.370	0.101	0.235	0.063
cost-effectiveness per dollar spent	0.129	0.056	0.166	-0.022	-0.052	0.431	0.072	0.346	0.286	-0.070	0.275	0.497
downside risk if the project succeeds at the wrong thing	0.132	0.346	0.242	-0.118	0.066	0.174	0.303	0.453	0.465	-0.084	0.193	0.180
probability of a net-negative outcome	0.189	0.561	0.030	-0.120	-0.154	-0.143	0.293	0.097	-0.038	-0.332	-0.128	-0.015
robustness of the plan to its key assumptions being wrong	0.145	0.338	0.414	0.306	0.204	0.307	0.264	0.063	0.385	0.155	0.005	0.395
ambition	0.347	0.075	0.180	0.270	0.259	0.435	0.262	0.129	0.046	0.105	0.250	0.124
weirdness relative to mainstream research funding	0.056	0.329	0.193	0.095	0.077	0.259	0.084	0.274	0.277	0.044	-0.065	0.270
interdisciplinarity	0.122	0.067	0.507	0.221	0.244	0.530	0.184	0.141	0.534	0.249	0.228	0.189
degree to which outputs are public goods	0.329	0.605	0.259	0.247	0.279	0.063	0.450	0.491	0.343	0.148	0.315	0.105
information value of simply running the project	0.412	0.392	0.366	0.200	0.286	0.132	0.515	0.620	0.438	0.430	0.175	0.110
replaceability by work others would do anyway	0.222	0.191	0.005	0.530	0.190	0.113	0.227	0.339	0.030	0.504	0.368	0.164
urgency of doing this now rather than in five years	0.115	0.105	0.325	0.091	-0.075	0.135	0.137	0.143	0.361	0.227	-0.094	0.277
maturity of the underlying field	0.227	0.166	0.159	0.231	-0.023	0.359	0.208	0.172	0.404	0.365	0.264	0.356
sensitivity of the project's value to AI timelines	0.339	0.116	0.423	0.251	-0.078	0.529	0.149	0.294	0.556	0.303	-0.089	0.718
clarity of the writing itself	0.233	0.297	0.005	0.284	0.258	0.100	0.033	0.152	-0.158	0.285	0.077	-0.080
fundability by mainstream institutions	-0.020	0.105	0.104	-0.037	-0.017	0.256	-0.212	0.071	-0.126	0.197	0.052	0.067
potential to become financially self-sustaining	0.177	0.186	0.393	0.358	0.041	0.249	0.315	0.302	0.383	0.320	0.184	0.407

Subtle: voyage4nano_line

Finite correlations 995/1,001; Q1 0.026, median 0.199, Q3 0.348.

Top attribute	rho
algorithm-shaped compromise	0.712
community-building instinct	0.705
fashionable-jargon half-life	0.674
bridge-building patience	0.662
mechanism before motivation	0.656
invention density	0.655
term-of-art precision	0.648
twitter-native compression	0.645
soft-power fluency	0.641
blame-preemption architecture	0.638

Bottom attribute	rho
accountability plumbing	-0.477
upkeep budgeting	-0.477
retraction grace	-0.559
process maturity	-0.624
urgency authenticity	-0.719
sublimated ambition	-0.771
audit-trail instinct	-0.865
striving energy	-0.894
structural inevitability	-0.949
dialectical fairness	-0.975

Subtle: voyage4nano_desc

Finite correlations 995/1,001; Q1 0.005, median 0.166, Q3 0.326.

Top attribute	rho
newcomer welcome	0.751
beginner's-mind access	0.706
bedside manner	0.697
coalition breadth	0.684
data-availability instinct	0.669
term-of-art precision	0.666
curse-of-knowledge resistance	0.663
community-building instinct	0.658
soft-power fluency	0.656
field-notes register	0.655

Bottom attribute	rho
overdue-ness	-0.522
vulnerability that isn't performance	-0.542
urgency authenticity	-0.575
example-to-principle ratio	-0.617
victim-instrumentalization risk	-0.637
sublimated ambition	-0.771
dialectical fairness	-0.872
audit-trail instinct	-0.883
striving energy	-0.894
structural inevitability	-0.949

Subtle: gemma_line

Finite correlations 995/1,001; Q1 0.035, median 0.239, Q3 0.401.

Top attribute	rho
invention density	0.776
falsifiable-promise density	0.773
blame-preemption architecture	0.772
beneficiary-voice presence	0.749
coalition breadth	0.727
twitter-native compression	0.726
mechanism before motivation	0.721
curse-of-knowledge resistance	0.720
guilt mobilization	0.708
community-building instinct	0.699

Bottom attribute	rho
chip-on-shoulder torque	-0.464
polish without sterility	-0.509
example-to-principle ratio	-0.560
process maturity	-0.570
evidence of killed darlings	-0.583
striving energy	-0.671
urgency authenticity	-0.766
sublimated ambition	-0.829
dialectical fairness	-0.872
structural inevitability	-0.949

Subtle: gemma_desc

Finite correlations 995/1,001; Q1 -0.018, median 0.156, Q3 0.316.

Top attribute	rho
discipline of its antifragility claims	0.723
blame-preemption architecture	0.704
term-of-art precision	0.700
field-notes register	0.686
first-principles freshness	0.678
twitter-native compression	0.678
reckless precision	0.665
stoic containment	0.647
curse-of-knowledge resistance	0.643
mechanism-design taste	0.637

Bottom attribute	rho
friction coefficient	-0.487
upkeep budgeting	-0.492
quotable-line frequency	-0.539
deadline scent	-0.547
playfulness under discipline	-0.616
dialectical fairness	-0.667
audit-trail instinct	-0.829
sublimated ambition	-0.886
striving energy	-0.894
structural inevitability	-0.949

d. Lexical arm

Single features are correlated directly with gemma31b gold. Description features use 38 matched records; the description ridge uses line fallback for two unmatched entities. Ridge uses alpha=1.0 and LOO.

Attribute	Best single feature	\|rho\|	Line ridge	Description ridge
existential seriousness	line:flesch_reading_ease	0.479	0.200	-0.175
technical depth	desc:capitalized_token_proportion	0.532	0.311	-0.366
epistemic rigor of the underlying theory of change	line:word_count	0.476	0.371	-0.031
counterfactual impact of a marginal dollar of funding	line:word_count	0.417	0.056	-0.286
neglectedness of the problem the project addresses	line:word_count	0.326	-0.144	-0.075
tractability of the plan as proposed by this specific team	line:type_token_ratio	0.314	0.013	-0.061
expected long-run impact on humanity's trajectory	line:flesch_reading_ease	0.343	0.260	-0.184
scale of the problem the project addresses	line:number_percentage_density	0.264	-0.121	-0.157
empirical testability of its core claims	desc:capitalized_token_proportion	0.507	0.247	0.156
quality of quantitative reasoning in the proposal	line:word_count	0.514	0.220	0.105
legibility of the team's track record	line:word_count	0.428	0.196	-0.173
feasibility given the stated team and resources	line:hedge_density	0.257	-0.101	-0.382
concreteness of proposed milestones	line:word_count	0.564	0.479	-0.256
clarity of the theory of change	line:word_count	0.510	0.275	-0.482
speed to a first observable result	line:capitalized_token_proportion	0.331	0.053	0.058
room for more funding in this exact niche	line:question_mark_count	0.317	0.128	0.007
cost-effectiveness per dollar spent	desc:number_percentage_density	0.374	0.396	0.069
downside risk if the project succeeds at the wrong thing	line:flesch_reading_ease	0.317	-0.023	-0.140
probability of a net-negative outcome	line:hedge_density	0.271	0.282	0.130
robustness of the plan to its key assumptions being wrong	desc:type_token_ratio	0.289	0.179	0.064
ambition	desc:question_mark_count	0.377	0.255	0.113
weirdness relative to mainstream research funding	desc:first_person_plural_ratio	0.361	-0.030	0.392
interdisciplinarity	desc:first_person_plural_ratio	0.405	-0.142	0.243
degree to which outputs are public goods	desc:capitalized_token_proportion	0.328	0.002	0.095
information value of simply running the project	line:word_count	0.520	0.376	-0.102
replaceability by work others would do anyway	line:word_count	0.363	0.010	0.053
urgency of doing this now rather than in five years	line:word_count	0.371	0.088	-0.531
maturity of the underlying field	desc:capitalized_token_proportion	0.295	-0.136	0.102
sensitivity of the project's value to AI timelines	line:flesch_reading_ease	0.340	0.020	-0.071
clarity of the writing itself	line:mean_word_length	0.350	0.329	-0.143
fundability by mainstream institutions	desc:first_person_plural_ratio	0.302	0.156	0.119
potential to become financially self-sustaining	desc:first_person_singular_ratio	0.460	-0.272	0.330

e. Discriminant table

Flags mark attributes where either available length measure has |rho(length, gold)| > 0.5.

Attribute	Line↔gold	Desc↔gold	Line↔voyage4nano_line pred	Desc↔voyage4nano_line pred	Line↔voyage4nano_desc pred	Desc↔voyage4nano_desc pred	Line↔gemma_line pred	Desc↔gemma_line pred	Line↔gemma_desc pred	Desc↔gemma_desc pred	Line↔PC1	Desc↔PC1	Flag
existential seriousness	0.142	-0.117	-0.091	-0.360	-0.215	-0.212	-0.165	-0.242	-0.221	0.017	0.127	0.030
technical depth	0.426	0.204	0.228	0.214	0.230	0.259	0.206	0.129	0.324	0.234	0.127	0.030
epistemic rigor of the underlying theory of change	0.476	0.337	0.312	0.313	0.345	0.406	0.175	0.185	0.514	0.342	0.127	0.030
counterfactual impact of a marginal dollar of funding	0.417	-0.015	0.032	0.134	-0.073	0.072	0.202	-0.054	0.038	0.199	0.127	0.030
neglectedness of the problem the project addresses	0.326	0.218	0.016	0.103	-0.168	-0.031	-0.036	-0.116	-0.053	0.083	0.127	0.030
tractability of the plan as proposed by this specific team	0.256	0.218	-0.077	0.086	0.109	0.191	0.409	0.021	0.132	0.151	0.127	0.030
expected long-run impact on humanity's trajectory	0.263	0.096	0.163	-0.168	0.211	-0.066	0.128	-0.175	0.102	0.057	0.127	0.030
scale of the problem the project addresses	-0.117	-0.088	0.187	-0.196	-0.023	-0.196	-0.195	-0.233	-0.053	-0.012	0.127	0.030
empirical testability of its core claims	0.479	0.163	0.131	0.207	0.140	0.161	0.340	0.299	0.298	0.139	0.127	0.030
quality of quantitative reasoning in the proposal	0.514	0.230	0.311	0.085	0.263	0.265	0.263	0.105	0.434	0.137	0.127	0.030	LENGTH
legibility of the team's track record	0.428	0.156	0.125	0.102	0.130	-0.034	0.481	0.080	0.178	0.063	0.127	0.030
feasibility given the stated team and resources	0.096	-0.089	-0.040	0.040	0.116	-0.057	0.099	0.143	0.106	-0.071	0.127	0.030
concreteness of proposed milestones	0.564	0.103	0.123	0.055	0.067	0.103	0.440	0.071	0.254	0.061	0.127	0.030	LENGTH
clarity of the theory of change	0.510	0.022	0.269	0.190	0.085	0.137	0.576	0.198	0.254	0.294	0.127	0.030	LENGTH
speed to a first observable result	0.326	0.273	0.028	0.071	0.260	0.080	0.305	0.051	0.176	0.018	0.127	0.030
room for more funding in this exact niche	0.040	0.034	-0.092	0.064	-0.056	0.101	-0.159	0.114	0.035	0.139	0.127	0.030
cost-effectiveness per dollar spent	0.332	-0.193	-0.044	-0.024	0.055	0.071	0.290	-0.002	-0.124	-0.050	0.127	0.030
downside risk if the project succeeds at the wrong thing	0.157	0.034	0.001	-0.308	-0.091	-0.180	-0.096	-0.223	-0.136	0.073	0.127	0.030
probability of a net-negative outcome	0.215	-0.162	0.016	-0.421	-0.092	-0.311	0.130	-0.314	-0.230	-0.147	0.127	0.030
robustness of the plan to its key assumptions being wrong	0.131	0.145	-0.038	0.294	0.018	0.318	0.077	0.368	0.166	0.233	0.127	0.030
ambition	-0.018	-0.141	0.181	-0.204	0.144	-0.227	-0.028	-0.216	-0.040	-0.025	0.127	0.030
weirdness relative to mainstream research funding	0.208	-0.013	-0.129	-0.203	-0.131	-0.192	0.037	-0.219	-0.233	-0.155	0.127	0.030
interdisciplinarity	0.121	0.162	-0.127	-0.186	-0.238	-0.181	-0.310	-0.278	-0.207	-0.030	0.127	0.030
degree to which outputs are public goods	0.140	0.054	0.043	-0.018	0.095	-0.009	-0.022	-0.146	0.102	0.210	0.127	0.030
information value of simply running the project	0.520	0.233	0.388	0.013	0.261	0.004	0.437	-0.078	0.247	0.150	0.127	0.030	LENGTH
replaceability by work others would do anyway	-0.363	-0.247	-0.380	-0.266	-0.308	-0.467	-0.260	-0.015	-0.380	-0.539	0.127	0.030
urgency of doing this now rather than in five years	0.371	0.174	0.089	0.068	-0.083	0.185	0.045	0.256	0.186	0.419	0.127	0.030
maturity of the underlying field	-0.050	-0.022	-0.145	-0.226	-0.054	-0.170	-0.024	-0.210	-0.089	-0.089	0.127	0.030
sensitivity of the project's value to AI timelines	0.016	0.117	0.041	-0.047	0.052	-0.018	-0.150	-0.243	0.048	0.060	0.127	0.030
clarity of the writing itself	-0.078	-0.126	-0.067	-0.028	-0.298	-0.003	0.118	0.041	-0.138	0.188	0.127	0.030
fundability by mainstream institutions	-0.009	-0.172	0.010	0.126	-0.155	-0.048	0.104	0.237	0.086	0.164	0.127	0.030
potential to become financially self-sustaining	0.263	-0.094	0.149	0.007	0.049	-0.164	0.260	-0.080	-0.165	-0.081	0.127	0.030

f. Frontier summary

Attribute	Panel median	Best lexical \|rho\|	Lexical ridge	Best embedding	Variant	Length rho
empirical testability of its core claims	0.806	0.507	0.247	0.507	gemma_desc	0.479
technical depth	0.789	0.532	0.311	0.527	gemma_line	0.426
downside risk if the project succeeds at the wrong thing	0.779	0.317	-0.023	0.303	gemma_line	0.157
quality of quantitative reasoning in the proposal	0.766	0.514	0.220	0.316	gemma_desc	0.514
concreteness of proposed milestones	0.753	0.564	0.479	0.446	gemma_line	0.564
information value of simply running the project	0.724	0.520	0.376	0.515	gemma_line	0.520
existential seriousness	0.724	0.479	0.200	0.456	gemma_line	0.142
weirdness relative to mainstream research funding	0.717	0.361	-0.030	0.095	voyage4nano_desc	0.208
speed to a first observable result	0.713	0.331	0.053	0.619	gemma_line	0.326
interdisciplinarity	0.701	0.405	-0.142	0.249	gemma_desc	0.121
epistemic rigor of the underlying theory of change	0.690	0.476	0.371	0.609	voyage4nano_desc	0.476
expected long-run impact on humanity's trajectory	0.686	0.343	0.260	0.419	voyage4nano_line	0.263
scale of the problem the project addresses	0.685	0.264	-0.121	0.456	gemma_line	-0.117
degree to which outputs are public goods	0.681	0.328	0.002	0.450	gemma_line	0.140
potential to become financially self-sustaining	0.673	0.460	-0.272	0.358	voyage4nano_desc	0.263
ambition	0.671	0.377	0.255	0.347	voyage4nano_line	-0.018
clarity of the writing itself	0.664	0.350	0.329	0.285	gemma_desc	-0.078
urgency of doing this now rather than in five years	0.659	0.371	0.088	0.227	gemma_desc	0.371
clarity of the theory of change	0.642	0.510	0.275	0.539	gemma_line	0.510
maturity of the underlying field	0.639	0.295	-0.136	0.365	gemma_desc	-0.050
feasibility given the stated team and resources	0.634	0.257	-0.101	0.505	gemma_desc	0.096
sensitivity of the project's value to AI timelines	0.632	0.340	0.020	0.339	voyage4nano_line	0.016
tractability of the plan as proposed by this specific team	0.629	0.314	0.013	0.287	voyage4nano_line	0.256
legibility of the team's track record	0.624	0.428	0.196	0.324	gemma_line	0.428
probability of a net-negative outcome	0.601	0.271	0.282	0.293	gemma_line	0.215
neglectedness of the problem the project addresses	0.585	0.326	-0.144	0.020	voyage4nano_line	0.326
fundability by mainstream institutions	0.573	0.302	0.156	0.197	gemma_desc	-0.009
cost-effectiveness per dollar spent	0.490	0.374	0.396	0.129	voyage4nano_line	0.332
replaceability by work others would do anyway	0.484	0.363	0.010	0.530	voyage4nano_desc	-0.363
room for more funding in this exact niche	0.477	0.317	0.128	0.218	voyage4nano_line	0.040
counterfactual impact of a marginal dollar of funding	0.453	0.417	0.056	0.214	gemma_line	0.417
robustness of the plan to its key assumptions being wrong	0.349	0.289	0.179	0.306	voyage4nano_desc	0.131

The frontier uses the line-text lexical ridge; description-ridge results remain in section d as the extra-information variant.

Cost classes per item: LLM pairwise ≈ 12 comparisons × prompt tokens per entity per attribute; embedding = one local forward pass; lexical = free.

Data findings

The main subtle run has 37,478 finite cells, 36 null scores, and 714 attributes with all 40 entities. Constant available scores make rank validity undefined for apology quality, corrosion resistance, self-doubt management, tu-quoque avoidance. Fewer than three finite entities occur for control of emphasis (n=2), inferential distance covered per paragraph (n=2), self-doubt management (n=2). The narrowest nonzero span is elephant-naming promptness (0.084, 4 unique values). Seed 2 has 293 attributes, with 9 absent from the main run and 284 overlapping attributes.