R0048/2026-04-01/Q002 — ACH Matrix¶
Matrix¶
| H1: Training warns about sycophancy | H2: Adjacent concepts only | H3: No related concepts | |
|---|---|---|---|
| SRC01-E01: Georgetown frames as unaddressed policy problem | -- | + | + |
| SRC02-E01: IPR "hidden risk" framing | -- | + | + |
| SRC03-E01: Brookings recommends future AI literacy | -- | + | + |
| SRC04-E01: Science study — 49% more sycophantic | -- | + | + |
| SRC05-E01: Microsoft/IBM failure scenario training | - | ++ | - |
| SRC06-E01: NHS automation bias naming | - | ++ | -- |
Legend:
- ++ Strongly supports
- + Supports
- -- Strongly contradicts
- - Contradicts
- N/A Not applicable to this hypothesis
Diagnosticity Analysis¶
Most Diagnostic Evidence¶
| Evidence | Why Diagnostic |
|---|---|
| SRC06-E01 | NHS automation bias naming is the key discriminator between H2 and H3 — proves at least adjacent concepts exist in training |
| SRC01-E01 | Georgetown's policy framing as unaddressed risk is the key discriminator between H1 and H2 — proves sycophancy itself is not in training |
Least Diagnostic Evidence¶
| Evidence | Why Non-Diagnostic |
|---|---|
| SRC04-E01 | Science study documents the phenomenon but says nothing about whether training addresses it — it equally supports both H2 and H3 |
Outcome¶
Hypothesis supported: H2 — Sycophancy is absent from training but adjacent concepts (automation bias, failure scenarios) are partially addressed in some programs.
Hypotheses eliminated: H1 — No evidence found of any training program warning about sycophancy by name or by behavioral description.
Hypotheses inconclusive: H3 — Partially supported for most corporate training but contradicted by NHS automation bias coverage and Microsoft failure-scenario training.