Edge Grades — Phase 2A.2 + 2A.3 Audit

Distribution of GATING / STRONG / WEAK / NEGATIVE / DROP grades across the 1010 scenario→prediction edges. The grader (Phase 2A.2) rebalances each cosine-discovered correlate edge into a strength band; the keyword linker (Phase 2A.3) adds direct edges for predictions cosine missed. See /graph/dag fork views for the resulting differentiation.

Total scenario→pred edges

1177

Manual prereq (untouched)

247

weight ≈1.0

GATING

248

w=0.7 p_pos=0.85

STRONG

188

w=0.5 p_pos=0.75

WEAK

140

w=0.3 p_pos=0.65

NEGATIVE

w=0.4 p_pos=0.30

DROP

165

w=0.05 (neutralized)

Phase 2A.3 keyword-added

for isolated high-conv preds

Per-scenario grade distribution

Sorted by GATING count desc; clicking a scenario opens its fork view.

Scenario	Family	Total	GAT	STR	WEAK	NEG	DROP	2A.3
S_HUMANOID_MASS_2033	Humanoid deployment	37	21	9	7
S_HUMANOID_FACTORY_2026	Humanoid deployment	32	21	9			2
S_COMPUTE_100GW_2030	Compute scale	39	17	14	5		1	2
S_HUMANOID_CONSUMER_2030	Humanoid deployment	33	16	8	8		1
S_AGI_WINTER_2036PLUS	AGI	38	15	12	11
S_IPO_TRILLION_2026	$1T+ IPO	34	15	9	5		5
S_IPO_TRILLION_2027	$1T+ IPO	29	15	4	3		7
S_AGI_MID_2029	AGI	38	14	14	4			6
S_GRID_50GW_2027	Energy / grid	40	12	18	8		2
S_HUMANOID_ENTERPRISE_2028	Humanoid deployment	37	12	9	12		3	1
S_ASI_SLOW_2040PLUS	ASI	27	11	14	2
S_MARS_2026	Mars uncrewed	19	10		2		7
S_AGI_FAST_2027	AGI	31	9	6	6		1	9
S_AGI_SLOW_2031	AGI	27	8	7	5		7
S_MARS_2028	Mars uncrewed	26	8		3		14	1
S_MARS_2031PLUS	Mars uncrewed	23	8		3		12
S_ROBOTAXI_TESLA_2026	Robotaxi	22	6	2	1		13
S_COMPUTE_10GW_2028	Compute scale	26	5	14			4	3
S_COMPUTE_1GW_2027	Compute scale	19	5	10	2		1	1
S_ROBOTAXI_MASS_2030	Robotaxi	16	5	4	3		4
S_IPO_TRILLION_2028	$1T+ IPO	13	4				9
S_ASI_MID_2034	ASI	22	2	12	3		5
S_ROBOTAXI_NATIONWIDE_2028	Robotaxi	9	2				7
S_ASI_FAST_2031	ASI	7	2	4				1
S_COMPUTE_STARGATE_FAILURE	Compute scale	6	2	1			3
S_NO_RECESSION_5Y	Recession	19	1		7		11
S_RECESSION_2026	Recession	11	1		2		6	2
S_GRID_50GW_2029	Energy / grid	6	1	3			1	1
S_NO_AI_PAUSE_5Y	AI pause	23			12	6	5
S_AI_PAUSE_2026	AI pause	21			12	6	3
S_AI_PAUSE_2028	AI pause	19			7	5	7
S_AI_PAUSE_2027	AI pause	14			2	5	7
S_RECESSION_2027	Recession	9			2		7
S_GRID_50GW_DELAYED	Energy / grid	8		5			3
S_RECESSION_2028	Recession	7			2		5
S_ROBOTAXI_DELAYED	Robotaxi	2			1		1
S_IPO_TRILLION_NONE_5Y	$1T+ IPO	1					1

How to read this

GATING (green) — edge fires strongly when the scenario is clamped TRUE. p_pos=0.85 contributes ~+1.36 logits at weight 0.7.
STRONG (blue) — moderate push. p_pos=0.75 × weight 0.5 ≈ +0.55 logits.
WEAK — Phase 2A default before grading. p_pos=0.65 × weight 0.3 ≈ +0.19 logits. ~7× weaker than GATING.
NEGATIVE (amber) — anti-correlated. p_pos=0.30 (FALSE more likely under TRUE clamp). Used for AI Pause × AI advancement predictions.
DROP — spurious cosine match neutralized at weight 0.05. Effectively no edge.
2A.3 (purple) — keyword-linked from link_isolated_high_conviction.py; carries the same grade taxonomy.
Scenarios with high GATING + STRONG counts produce the largest fork-lane "avg shift" signals. Scenarios with mostly DROP rows are inherently sparse (NONE/FAILURE/DELAYED counterfactuals).