# Experiment 008 — representation matrix

Held-out 24-way fact-identity accuracy. Chance = 1/24 = 0.0417.
Significance bar (one-tailed binomial, n=48, p=1/24, α=0.01): **≥ 7/48 = 0.1458**.

Row `k` is `hidden_states[k]`: row 0 is the token embeddings (the lexical baseline), row *k* is the input to decoder layer *k*, row 18 is exactly what Experiments 005–007 read, row 28 is the final hidden state.

## primary centroids

all 6 training phrasings — **the headline definition**

| row | last_prompt | cue_entity | prompt_mean |
|---:|---:|---:|---:|
| 0 ← embeddings | 0.0417 | *0.3750* | *1.0000* |
| 1 | 0.2083 | *1.0000* | *0.9792* |
| 2 | 0.1875 | *1.0000* | *0.9583* |
| 3 | 0.0625 | *1.0000* | *0.9583* |
| 4 | 0.3125 | *1.0000* | *1.0000* |
| 5 | 0.1250 | *1.0000* | *1.0000* |
| 6 | 0.2500 | *1.0000* | *0.9583* |
| 7 | 0.2500 | *1.0000* | *0.9583* |
| 8 | 0.2083 | *1.0000* | *0.9375* |
| 9 | 0.1250 | *1.0000* | *0.8333* |
| 10 | 0.1458 | *1.0000* | *0.8542* |
| 11 | 0.2292 | *1.0000* | *0.7292* |
| 12 | 0.1667 | *1.0000* | *0.7083* |
| 13 | 0.1875 | *1.0000* | *0.6667* |
| 14 | 0.1458 | *1.0000* | *0.5833* |
| 15 | 0.1250 | *1.0000* | *0.5833* |
| 16 | 0.1250 | *1.0000* | *0.6042* |
| 17 | 0.1250 | *1.0000* | *0.7292* |
| 18 ← Exp 005–007 | 0.2292 | *1.0000* | *0.6875* |
| 19 | 0.2083 | *1.0000* | *0.6875* |
| 20 | 0.3333 | *1.0000* | *0.9583* |
| 21 | 0.4375 | *1.0000* | *0.9583* |
| 22 | 0.8958 | *1.0000* | *0.8750* |
| 23 | 0.7917 | *1.0000* | *0.8542* |
| 24 | 0.8333 | *1.0000* | *0.9375* |
| 25 | 0.7917 | *1.0000* | *0.9167* |
| 26 | 0.8750 | *1.0000* | *0.9167* |
| 27 | 0.8750 | *1.0000* | *0.8333* |
| 28 ← final state | 0.8542 | *0.9792* | *0.4583* |

*Italic* = position is lexically determined and may not be cited as paraphrase-stable addressing. **Bold** = ≥ 0.90 at a position that is not.

## qa_only centroids

2 QA training phrasings only — **leakage control, never the headline**

| row | last_prompt | cue_entity | prompt_mean |
|---:|---:|---:|---:|
| 0 ← embeddings | 0.0417 | *0.3750* | *1.0000* |
| 1 | 0.3125 | *1.0000* | *1.0000* |
| 2 | 0.1458 | *1.0000* | *1.0000* |
| 3 | 0.0417 | *1.0000* | *1.0000* |
| 4 | 0.1667 | *1.0000* | *1.0000* |
| 5 | 0.2292 | *1.0000* | *1.0000* |
| 6 | 0.1042 | *1.0000* | *1.0000* |
| 7 | 0.0833 | *1.0000* | *1.0000* |
| 8 | 0.2083 | *1.0000* | *1.0000* |
| 9 | 0.1875 | *1.0000* | *1.0000* |
| 10 | 0.1875 | *1.0000* | *1.0000* |
| 11 | 0.2292 | *1.0000* | *1.0000* |
| 12 | 0.1875 | *1.0000* | *1.0000* |
| 13 | 0.2083 | *1.0000* | *1.0000* |
| 14 | 0.2500 | *1.0000* | *1.0000* |
| 15 | 0.2708 | *1.0000* | *0.9792* |
| 16 | 0.3125 | *1.0000* | *0.9583* |
| 17 | 0.3542 | *1.0000* | *1.0000* |
| 18 ← Exp 005–007 | 0.2292 | *1.0000* | *0.9792* |
| 19 | 0.4167 | *1.0000* | *1.0000* |
| 20 | 0.5417 | *1.0000* | *1.0000* |
| 21 | 0.4167 | *1.0000* | *1.0000* |
| 22 | 0.8958 | *1.0000* | *1.0000* |
| 23 | 0.8750 | *1.0000* | *1.0000* |
| 24 | 0.8958 | *1.0000* | *1.0000* |
| 25 | 0.8958 | *0.9792* | *1.0000* |
| 26 | **0.9792** | *0.9792* | *1.0000* |
| 27 | **0.9792** | *0.9375* | *1.0000* |
| 28 ← final state | **0.9583** | *0.8542* | *0.9375* |

*Italic* = position is lexically determined and may not be cited as paraphrase-stable addressing. **Bold** = ≥ 0.90 at a position that is not.