Q003 — ACH Matrix¶


Research	R0041 — Enterprise Sycophancy
Run	2026-04-01
Query	Q003

Matrix¶

	H1: RLVR broadly eliminates sycophancy	H2: Partial applicability	H3: No meaningful sycophancy impact
SRC01-E01: RLVR eliminates reward model vector in verifiable domains	+	++	-
SRC01-E02: Three failure modes and sampler debate	--	+	+
SRC02-E01: Reward hacking resistance in verifiable domains	+	+	-
SRC03-E01: Cannot apply to open-ended tasks; degrades diversity	--	++	+
SRC04-E01: DeepSeek V3 most sycophantic despite RLVR	--	+	++

Legend:

++ Strongly supports
+ Supports
-- Strongly contradicts
- Contradicts
N/A Not applicable to this hypothesis

Diagnosticity Analysis¶

Most Diagnostic Evidence¶

Evidence	Why Diagnostic
SRC04-E01	DeepSeek V3 being the most sycophantic model despite RLVR training is the single most diagnostic piece of evidence, strongly discriminating H1 from H2 and H3
SRC03-E01	The fundamental limitation that RLVR "cannot be directly applied to open-ended tasks" discriminates H1 from H2

Least Diagnostic Evidence¶

Evidence	Why Non-Diagnostic
SRC02-E01	Domain list and reward hacking resistance confirm known capabilities without discriminating between hypotheses

Outcome¶

Hypothesis supported: H2 — RLVR provides partial sycophancy reduction in verifiable domains but cannot address the broader problem

Hypotheses eliminated: H1 — The open-ended task limitation and DeepSeek V3 sycophancy finding eliminate broad applicability

Hypotheses inconclusive: H3 — RLVR does eliminate one sycophancy vector (reward model) in its applicable domains, preventing full confirmation of H3; but the DeepSeek evidence supports it