Q003 — Hallucination Training — ACH Matrix¶
Matrix¶
| Evidence | H1 — Fundamental + spectrum | H2 — Occasional random | H3 — Undifferentiated, missing spectrum/sycophancy |
|---|---|---|---|
| SRC01-E01 — Rich taxonomy in research | -- | - | ++ |
| SRC02-E01 — IBM "fundamental operation" | + | - | + |
| SRC03-E01 — Confirmatory hallucination | -- | - | ++ |
| SRC04-E01 — Co-hallucination concept | -- | - | ++ |
| SRC05-E01 — NIST probabilistic framing | + | - | + |
| SRC06-E01 — Technical fix framing | - | + | + |
| SRC07-E01 — "Fluency ≠ accuracy" | + | N/A | + |
| SRC08-E01 — Biased sampling mechanism | -- | -- | ++ |
Legend¶
| Symbol | Meaning |
|---|---|
| ++ | Strongly consistent with hypothesis |
| + | Consistent with hypothesis |
| - | Inconsistent with hypothesis |
| -- | Strongly inconsistent with hypothesis |
| N/A | Not applicable |
Diagnosticity Analysis¶
Most diagnostic evidence:
- SRC08-E01 — The biased sampling mechanism is strongly inconsistent with both H1 (not in training) and H2 (not random), while strongly consistent with H3 (missing from training).
- SRC03-E01 — The hallucination-sycophancy connection discriminates sharply between H1 (requires this connection) and H3 (notes its absence).
Least diagnostic evidence:
- SRC07-E01 — "Fluency ≠ accuracy" is a step toward H1 but exists in guides, not training, making it consistent with H3 as well.
Outcome¶
H3 (undifferentiated, missing spectrum and sycophancy connection) receives the strongest support (6 of 8 items rated + or ++). H1 is eliminated by the absence of spectrum/sycophancy content from training (4 items rated --). H2 is partially supported but too simplistic. H3 is the best-supported hypothesis.