R0042/2026-03-28/Q003 — ACH Matrix¶
Matrix¶
| H1: Documented enterprise deployment exists | H2: No enterprise deployment, confined to providers/research | H3: Component in provider design, not enterprise primary goal | |
|---|---|---|---|
| SRC01-E01: Anthropic Constitutional AI | -- | ++ | ++ |
| SRC02-E01: Google DeepMind consistency training | -- | ++ | ++ |
| SRC03-E01: Sycophancy survey — no enterprise examples | -- | ++ | + |
| SRC04-E01: OpenAI GPT-4o incident response | -- | ++ | ++ |
Legend:
- ++ Strongly supports
- + Supports
- -- Strongly contradicts
- - Contradicts
- N/A Not applicable to this hypothesis
Diagnosticity Analysis¶
Most Diagnostic Evidence¶
| Evidence ID | Why Diagnostic |
|---|---|
| SRC03-E01 | A comprehensive academic survey of the entire sycophancy field contains zero enterprise deployment examples — if such deployments existed, this survey would reference them |
| SRC01-E01 | Anthropic's anti-sycophancy work is the closest to an enterprise-relevant case, but the design goal belongs to the model provider, not the enterprise customer — this distinction is highly diagnostic |
Least Diagnostic Evidence¶
| Evidence ID | Why Non-Diagnostic |
|---|---|
| SRC04-E01 | The GPT-4o incident demonstrates provider accountability but does not help discriminate between H2 and H3 |
Outcome¶
Hypothesis supported: H3 — Anti-sycophancy exists as a design component at the model provider and research institution level but has not been documented as a primary goal of enterprise private AI deployment.
Hypotheses eliminated: H1 — No evidence of enterprise anti-sycophancy private AI deployment exists.
Hypotheses inconclusive: H2 — Fully supported but less precise than H3. The distinction is that H2 says "confined to providers/research" while H3 adds the nuance that enterprises benefit from provider-level anti-sycophancy work.