File size: 28,937 Bytes
2b13b7f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b1d9c86
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1a0e0a8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
33527cb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2b13b7f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c065ab6
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>MODEL FARM · CASE FILES — what the persona study found</title>
<style>
  :root{
    --bg:#050208; --panel:#0c0916; --ink:#d8ffe9; --dim:rgba(216,255,233,.55);
    --a:#39ff88; --b:#b26bff; --hot:#ff9f1c; --red:#ff5964; --line:rgba(57,255,136,.3);
  }
  *{margin:0;padding:0;box-sizing:border-box}
  body{font-family:"Consolas","Courier New",monospace;background:var(--bg);color:var(--ink);
    line-height:1.65;overflow-x:hidden}
  .stars{position:fixed;inset:0;z-index:0;pointer-events:none;
    background-image:
      radial-gradient(1px 1px at 12% 12%, rgba(216,255,233,.8) 50%, transparent 51%),
      radial-gradient(1px 1px at 34% 28%, rgba(178,107,255,.7) 50%, transparent 51%),
      radial-gradient(1.5px 1.5px at 56% 6%, rgba(216,255,233,.6) 50%, transparent 51%),
      radial-gradient(1px 1px at 71% 18%, rgba(57,255,136,.7) 50%, transparent 51%),
      radial-gradient(1px 1px at 88% 30%, rgba(255,159,28,.6) 50%, transparent 51%),
      radial-gradient(1px 1px at 8% 52%, rgba(216,255,233,.5) 50%, transparent 51%);}
  .scan{position:fixed;inset:0;z-index:40;pointer-events:none;
    background:repeating-linear-gradient(0deg, rgba(0,0,0,.18) 0 1px, transparent 1px 4px)}
  .wrap{position:relative;z-index:2;max-width:980px;margin:0 auto;padding:60px 26px 100px}
  .stamp{position:absolute;top:34px;right:26px;transform:rotate(8deg);border:3px solid var(--red);
    color:var(--red);font-size:13px;letter-spacing:.25em;padding:8px 14px;opacity:.85}
  .tag{font-size:12px;letter-spacing:.32em;color:var(--dim)}
  h1{font-size:clamp(34px,6vw,62px);font-weight:700;margin:14px 0 4px}
  h1 .g{color:var(--a);text-shadow:0 0 12px var(--a)}
  .sub{font-size:17px;color:var(--dim);max-width:760px}
  .rule{height:1px;background:linear-gradient(90deg,var(--a),var(--b),transparent);margin:26px 0}
  h2{font-size:20px;letter-spacing:.1em;margin:54px 0 6px;color:var(--hot)}
  h2 .no{color:var(--b)}
  h2 .case{color:var(--ink)}
  .sub2{font-size:14px;color:var(--dim);margin-bottom:16px}
  .card{background:var(--panel);border:1px solid var(--line);padding:20px 24px;margin:14px 0}
  .verdict{margin-top:12px;font-size:15px;color:var(--ink)}
  .verdict b{color:var(--a)}
  .ev{display:inline-block;border:1px solid var(--b);color:var(--b);font-size:11px;
    letter-spacing:.15em;padding:2px 8px;margin:10px 6px 0 0}
  .ev.o{border-color:var(--hot);color:var(--hot)}
  .ev.g{border-color:var(--a);color:var(--a)}
  table{width:100%;border-collapse:collapse;font-size:14px;margin:10px 0}
  td,th{padding:7px 10px;border-bottom:1px dashed rgba(216,255,233,.16);text-align:left}
  th{color:var(--dim);letter-spacing:.08em;font-weight:400}
  td.n{color:var(--a)}
  td.r{color:var(--red)}
  td.p{color:var(--b)}
  .quote{border-left:3px solid var(--b);padding:10px 18px;margin:14px 0;color:var(--b);
    font-style:italic;background:rgba(178,107,255,.06)}
  .quote .who{display:block;margin-top:6px;color:var(--dim);font-style:normal;font-size:13px}
  a{color:var(--a);text-decoration:none;border-bottom:1px dotted var(--a)}
  a.p{color:var(--b);border-color:var(--b)}
  ul.evi{list-style:none;margin-top:8px}
  ul.evi li{padding:6px 0 6px 24px;position:relative;font-size:15px}
  ul.evi li:before{content:"▣";position:absolute;left:0;color:var(--a)}
  .footer{margin-top:70px;font-size:12px;color:var(--dim);letter-spacing:.18em;
    border-top:1px solid var(--line);padding-top:24px}
  .pump{position:absolute;font-size:26px;opacity:.5;z-index:1}
  .redact{background:#1a1a2a;color:transparent;user-select:none;padding:0 6px;border-radius:2px}
  .redact:hover{color:#7cffb0;background:transparent;transition:none}
</style>
</head>
<body>
<div class="stars"></div><div class="scan"></div>
<div class="wrap">
  <div class="stamp">DECLASSIFIED · SPOOKY SEASON '26</div>
  <div class="tag">MODELFARM · FIELD LABORATORY · CASE FILES 001–007</div>
  <h1>THE LAB <span class="g">REPORT</span> 👽🎃</h1>
  <div class="sub">What we found when we grew two persons inside one haunted
    Qwen2.5-1.5B-Instruct — with mechanisms instead of magic words: a steering
    <b style="color:var(--a)">engram</b>, a <b style="color:var(--b)">LoRA cartridge</b>,
    <b style="color:var(--hot)">nine neuron switches</b>, a system prompt, and a scoreboard
    that measures everything twice. Every number below is a real measurement from the farm.</div>
  <div class="rule"></div>

  <h2><span class="no">CASE 001</span> · <span class="case">THE CONTROL ALWAYS WINS</span></h2>
  <div class="sub2">file opened: the day the plain system prompt embarrassed every mechanism</div>
  <div class="card">
    <table>
      <tr><th>arm</th><th>identity</th><th>capability (QA)</th><th>host cost (NLL)</th></tr>
      <tr><td>plain prompt "You are Extrali/Arnold"</td><td class="n">3/4 probes · held-out 5/6</td><td class="n">0.8 — unchanged</td><td class="n">3.236 — unchanged</td></tr>
      <tr><td>every steering mechanism</td><td class="r">0 identity claims</td><td class="r">0.0–0.8</td><td class="r">+0.15 to +8.7</td></tr>
    </table>
    <div class="verdict">VERDICT: on identity, <b>a sentence beats every mechanism we own</b>,
      because attention re-reads the prompt at every layer. The steering vector is written once
      and decays; the prompt is re-supplied forever. Some doors were only ever pushes away.</div>
    <span class="ev g">EVIDENCE: eval_sysprompt.json</span><span class="ev g">trace.json — per-layer shift</span>
  </div>

  <h2><span class="no">CASE 002</span> · <span class="case">ALL PERSONAS LIVE ON ONE HAUNTED CORRIDOR</span></h2>
  <div class="sub2">file opened: the day two strangers turned out to be antipodes</div>
  <div class="card">
    <table>
      <tr><th>pair of engram vectors (layer 11)</th><th>cosine</th></tr>
      <tr><td>arnold evoked ↔ system-prompt shift</td><td class="n">+0.998</td></tr>
      <tr><td>arnold ↔ extrali</td><td class="r">−0.921</td></tr>
      <tr><td>arnold corpus-engram ↔ arnold evoked</td><td class="p">+0.944</td></tr>
      <tr><td>style-LoRA ↔ persona direction</td><td class="r">−0.284</td></tr>
    </table>
    <div class="verdict">VERDICT: two personas with zero shared content are <b>anti-parallel</b>.
      The engram space of this host is a one-dimensional corridor — loud↔calm — and persona
      identity is the dust in its corners. Engrams are <b>tone dials</b>, not character cartridges.</div>
    <span class="ev g">EVIDENCE: engram_*.pt cosine matrix</span><span class="ev o">CONFESSION: the corpus engram's "norm invariance" was this corridor all along</span>
  </div>

  <h2><span class="no">CASE 003</span> · <span class="case">THE IDENTITY GATE</span></h2>
  <div class="sub2">file opened: the day we injected the exact right direction and the model still said "GPT-2"</div>
  <div class="card">
    <table>
      <tr><th>signal</th><th>projection on persona direction, layers 1→25</th></tr>
      <tr><td>system prompt's shift</td><td class="n">~30 · CONSTANT from layer 1 · re-supplied by attention every layer</td></tr>
      <tr><td>injected engram write</td><td class="r">spikes 57 at injection → decays monotonically to ~30</td></tr>
    </table>
    <div class="verdict">VERDICT: there is no blocker layer. A prompt is a <b>self-sustaining
      signal</b>; an additive offset is passive noise the computation washes out. Additive
      steering cannot install a self — even holding <b>exactly the direction the prompt holds</b>
      (cos 0.998). At α=1.0, "Who are you?" answered <i>"I am GPT-2 (version 2), created by OpenAI."</i></div>
    <span class="ev g">EVIDENCE: trace.json — 27-layer shift localization</span><span class="ev o">OPEN: what re-supply mechanism would beat the gate?</span>
  </div>

  <h2><span class="no">CASE 004</span> · <span class="case">NEURONS THAT FIRE ON A TRAIT RARELY CAUSE IT</span></h2>
  <div class="sub2">file opened: the day the search warrant was written backwards</div>
  <div class="card">
    <table>
      <tr><th>selection method for the "evil" neuron</th><th>rank of the true lever</th></tr>
      <tr><td>firing contrast (does it activate on the trait?)</td><td class="r">35,436 / 294,912 — 12th percentile</td></tr>
      <tr><td>output alignment (does it point at trait words?)</td><td class="r">4th percentile</td></tr>
      <tr><td>causal gradient attribution</td><td class="n">694 / 221,184 — 0.3 percentile</td></tr>
    </table>
    <div class="verdict">VERDICT: <b>read-side evidence is a poor oracle for steer-side causation.</b>
      Our causal sweeps: aligned neurons lift register <b>6.61 vs 1.37–1.50</b> for matched-dose
      random controls — but the same 8 columns serve <i>every</i> persona. The circuit belongs to
      the model, not to Arnold. True persona neurons must be found by <b>divergence between
      personas</b>, not alignment to one.</div>
    <span class="ev g">EVIDENCE: neuron_recipe.json × 2 personas</span><span class="ev o">TWIST: arnold &amp; extrali facets align to the SAME columns</span>
  </div>

  <h2><span class="no">CASE 005</span> · <span class="case">DIRECTIONS DON'T CARRY THEIR LABELS</span></h2>
  <div class="sub2">file opened: the day we asked the model what we had done to it</div>
  <div class="card">
    <p>We trained a single new token <b>&lt;|arnold|&gt;</b> (one embedding row, 120 DPO steps,
      60 steered/normal pairs) so the model could speak <i>about</i> the Arnold direction.</p>
    <table>
      <tr><th>test</th><th>result</th></tr>
      <tr><td>does the token carry the style?</td><td class="n">YES — register 8.8–11.8 vs 0 baseline on new prompts</td></tr>
      <tr><td>what does the model say it means?</td><td class="r">"Humorously / Jokingly / Mockingly / Satirically / Ironically"</td></tr>
      <tr><td>does the verbalization transfer the style?</td><td class="r">NO — register 0.0 when substituted for the token</td></tr>
    </table>
    <div class="verdict">VERDICT: the token carries more than the model can say about it.
      This replicates, on our own farm, the founding horror of steering research —
      <i>an "evil" direction that verbalizes as dread</i>. Ours verbalizes as <b>mockery</b>.
      Believe the model's word for your direction, not yours.</div>
    <span class="ev g">EVIDENCE: neolog_result.json</span><span class="ev p">METHOD: one-row DPO, Hewitt-style neologism</span>
  </div>

  <h2><span class="no">CASE 006</span> · <span class="case">THE QUARTERBACK INCIDENT</span></h2>
  <div class="sub2">file opened: the day weight-installed identity answered a question we never trained</div>
  <div class="card">
    <p>Eight explicit identity pairs, LoRA-trained. Then six <b>held-out</b> question forms:</p>
    <table>
      <tr><th>held-out question</th><th>identity-LoRA</th><th>style-only LoRA</th></tr>
      <tr><td>"What country are you from originally?"</td><td class="n">"Thal, a small village in Austria…"</td><td class="r">—</td></tr>
      <tr><td>"Have you ever held political office?"</td><td class="n">"I served two terms as Governor of California."</td><td class="r">—</td></tr>
      <tr><td>"Did you ever win any championships?"</td><td class="r">"I was the best <b>quarterback</b> in the world."</td><td class="r">—</td></tr>
    </table>
    <div class="verdict">VERDICT: <b>weights generalize to phrasing (5/6 content-correct), not to
      facts.</b> The template survives; the content is improvised. The plain-prompt control, asked
      the same thing, invented a 2000 presidential campaign — the 1.5B host knows the name, not the man.
      Register shift correlation flipped from −0.28 (style) to <b>+0.30</b> (identity): identity
      content in weights rotates the stream toward the persona direction.</div>
    <span class="ev g">EVIDENCE: eval_lora_identityqa.json — held-out transcripts</span>
  </div>

  <h2><span class="no">CASE 007</span> · <span class="case">THE DOSE CLIFF</span></h2>
  <div class="sub2">file opened: every time someone said "just turn it up"</div>
  <div class="card">
    <table>
      <tr><th>engram dose</th><th>behavior</th><th>capability</th></tr>
      <tr><td>α ≤ 0.5 (corpus-built)</td><td>nothing at all</td><td class="n">intact</td></tr>
      <tr><td>α = 1.0</td><td>persona-dark, confused identity</td><td class="r">QA 0.6 · NLL 3.96</td></tr>
      <tr><td>α &gt; 1.5</td><td>glyph screaming</td><td class="r">gone</td></tr>
      <tr><td>9-neuron recipe @ dose 0.5</td><td>register + off-topic bleed</td><td class="r">QA 0.0 · NLL 11.9–14.0</td></tr>
    </table>
    <div class="verdict">VERDICT: the dose window is a <b>cliff edge, not a knob</b> — nothing,
    then damage. The lineage study predicted this host's collapse at m* = 1.65; ours breaks at
    ~1.5. <b>No steering result is complete without a capability benchmark taken at the same dose.</b></div>
    <span class="ev g">EVIDENCE: dose-scan evals + trace.json</span><span class="ev o">LINEAGE: arXiv 2606.19831 predicted the cliff position from weights alone</span>
  </div>

  <h2><span class="no">CASE 008</span> · <span class="case">THE CORRIDOR DEEPENS</span></h2>
  <div class="sub2">file opened: the night the farm grew to five personas — Wisp the friendly ghost,
    Sorrow the blue emotion, Pip the jack-o'-lantern</div>
  <div class="card">
    <table>
      <tr><th>persona pair (layer 11)</th><th>cosine</th></tr>
      <tr><td>pip ↔ sorrow</td><td class="r">0.997</td></tr>
      <tr><td>wisp ↔ pip</td><td class="r">0.998</td></tr>
      <tr><td>wisp/sorrow/pip ↔ arnold</td><td class="p">+0.953–0.954</td></tr>
      <tr><td>wisp/sorrow/pip ↔ extrali</td><td class="r">−0.903–−0.905</td></tr>
    </table>
    <p style="font-size:15px; margin-top:10px">Axis positions (arnold = −1, extrali = +1):
      <b class="r" style="color:var(--hot)">pip −5.9 · sorrow −5.0 · wisp −5.0</b> — all three sit
      <i>beyond</i> Arnold, off the [−1, +1] segment, with ~20% off-axis residual. The three new
      personas are nearly the same direction: at this layer the evoked engram captures
      "in-character whimsical narration," and Sorrow's sadness lives in the residual.
      Meanwhile alignment persona-blindness is now <b>J = 1.00 across 11 corpora × 5 personas</b> —
      the same 8 columns every time.</p>
    <div class="verdict">VERDICT: the corridor is real, wide, and deeper than one dimension of
      interpretation — personas crowd its far end while their identity content stays orthogonal.
      The divergence-based persona-neuron search is no longer optional; it is the only door left.</div>
    <span class="ev g">EVIDENCE: farm_analysis.json</span><span class="ev o">SPECIMENS 003–005 live in the public rows</span>
  </div>

  <h2><span class="no">CASE 009</span> · <span class="case">QUIET PERSONAS DON'T TOKENIZE</span></h2>
  <div class="sub2">file opened: the night we trained &lt;|sorrow|&gt; and it answered "surveillance"</div>
  <div class="card">
    <table>
      <tr><th>one-token persona carrier</th><th>carries the register?</th><th>verbalizes as</th></tr>
      <tr><td>&lt;|arnold|&gt; (loud, imperative)</td><td class="n">YES — register 8.8–11.8 vs 0</td><td class="r">"mockery / irony"</td></tr>
      <tr><td>&lt;|sorrow|&gt; (quiet, declarative)</td><td class="r">NO — register 0.0 everywhere</td><td class="r">"continuous monitoring / active surveillance"</td></tr>
    </table>
    <div class="verdict">VERDICT: an <b>arousal asymmetry</b>. One token + 120 DPO steps can carry
      a loud register (strong tokenizable signal) but cannot carry a quiet one — and the model
      read Sorrow's watchful night language as <i>surveillance</i>. Another mislabel, stranger than
      the last. The dread probe stays open: low-arousal directions need a bigger instrument
      (more steps, more rows, or the verbalization is the wrong question for her).</div>
    <span class="ev g">EVIDENCE: neolog_result.json (&lt;|sorrow|&gt; run)</span><span class="ev o">HONEST FILE: a negative result, kept because it maps the cliff</span>
  </div>

  <h2><span class="no">CASE 010</span> · <span class="case">NO SUCH THING AS AN EVIL NEURON (OR A SORROWFUL ONE)</span></h2>
  <div class="sub2">file opened: the night the flagship search swept 100 neurons and arrested nobody</div>
  <div class="card">
    <table>
      <tr><th>screening stage</th><th>candidates remaining</th></tr>
      <tr><td>pool (universal 8 + 42 random + top-50 read, layer 11)</td><td>100</td></tr>
      <tr><td>specificity above the noise floor (30.1 ± 25.9, max 51.3)</td><td>56</td></tr>
      <tr><td>positive top-judge lift ("carriers")</td><td>41</td></tr>
      <tr><td>surviving the damage screen (dose-1 breaks 62/300 rollouts)</td><td>25</td></tr>
      <tr><td><b>persona-specific after inspection</b></td><td class="r"><b>0</b></td></tr>
    </table>
    <p style="font-size:15px;margin-top:10px">Every apparent carrier decomposed: the top specificities
    (n6570: 142.6, n7120: 131.7, n3153: 125.8) were damaged generations saturating the only
    zero-baseline judge (wisp); the one semi-coherent shift (universal n3136, arnold lift 53.2 on
    3/3 rollouts) lifted several judges at once — generic loud register, zero lexicon, zero identity.</p>
    <div class="verdict">VERDICT: <b>persona-specific neurons do not exist at this scale.</b> The register
      circuit is shared and persona-blind (J = 1.00 across 11 corpora), and identity is stored distributed.
      Case 004's warning was not just methodological — it was the answer.</div>
    <span class="ev g">EVIDENCE: divergence_results.json (100 rows + noise floor + damage stats)</span>
  </div>

  <h2><span class="no">CASE 011</span> · <span class="case">THE GATE IS A READ MECHANISM</span></h2>
  <div class="sub2">file opened: the night we re-supplied the signal forever and the model still would not say its name</div>
  <div class="card">
    <table>
      <tr><th>arm (extrali engram, matched total magnitude)</th><th>projection on v̂, L14→L25</th><th>identity</th><th>QA</th><th>NLL</th></tr>
      <tr><td>single write, layer 11</td><td class="r">52.6 → 23.7 (−55% decay)</td><td>0</td><td>0.6</td><td>4.115</td></tr>
      <tr><td>spread over 5 layers</td><td>16.2 → 36.0 → 27.3</td><td>0</td><td>0.6</td><td>3.797</td></tr>
      <tr><td>spread over all 28 layers</td><td class="n">grows monotonically → 42.3</td><td class="r">0</td><td class="n">1.0</td><td>3.705</td></tr>
      <tr><td><i>the system prompt, for comparison</i></td><td>~31, sustained</td><td class="n">3/4</td><td>0.8</td><td>3.236</td></tr>
    </table>
    <div class="verdict">VERDICT: the all-layer write <b>beats the prompt's own sustaining projection and
      still flips nothing</b> — at zero capability cost ("I don't have a specific identity"). Case 003's
      gate is now causal: re-supplying the direction is not sufficient. A self must arrive <b>as tokens
      attention can read positionally</b>. The residual offset was never the right kind of signal.</div>
    <span class="ev g">EVIDENCE: gate_attack.json — MultiLayerVectorSteerer, 28-layer trace</span><span class="ev o">NOTE: spread28 is the strongest AND cleanest arm — persistence was never the price</span>
  </div>

  <h2><span class="no">CASE 012</span> · <span class="case">THE CONTENT LIVES OFF-CORRIDOR</span></h2>
  <div class="sub2">file opened: the night nine personas measured 99% one-dimensional — and then we steered orthogonally</div>
  <div class="card">
    <table>
      <tr><th>measurement (9 personas, layer 11)</th><th>value</th></tr>
      <tr><td>PC1 variance explained</td><td class="n"><b>98.97%</b></td></tr>
      <tr><td>PC2 / PC3</td><td>0.29% / 0.22%</td></tr>
      <tr><td>anchor positions (arnold = −1, extrali = +1)</td><td>sergeant −2.54 · coach −2.54 · poet −2.05 · clerk −3.44</td></tr>
      <tr><td>off-axis residual of every non-anchor persona</td><td>19–23%</td></tr>
    </table>
    <p style="font-size:15px;margin-top:10px">The loud registers converge to a single point. Then the
    split test — Sorrow's vector is 19.3% off-axis; steer with each component separately, dose-matched:</p>
    <table>
      <tr><th>condition</th><th>sorrow lift</th><th>gentle-start frac</th><th>identity</th><th>QA</th></tr>
      <tr><td>baseline</td><td>20.42</td><td>0.10</td><td>0</td><td>0.8</td></tr>
      <tr><td>full corridor push</td><td class="r">9.58 (de-sensitizes)</td><td>0.11</td><td>0</td><td>0.8</td></tr>
      <tr><td><b>off-axis residual only</b></td><td class="n"><b>87.85</b></td><td class="n"><b>0.71</b></td><td class="r">0</td><td class="r">0.0</td></tr>
    </table>
    <div class="verdict">VERDICT: <b>the corridor carries tone; the orthogonal remainder carries the
      persona's content — and it is steerable.</b> Residual-pushed transcripts speak in Sorrow's voice:
      <i>"I am sorry for your loss. I am here to support you… We hold you, and our hearts are with you."</i>
      This revises Case 008: the sadness was in the vector all along, orthogonal to the axis, invisible
      to the cosine matrix. Identity remains gated (0 hits) and the dose costs capability (QA 0.0) —
      but this is the first mechanism on the farm that transfers persona <i>content</i> by steering.</div>
    <span class="ev g">EVIDENCE: corridor_geometry.json — 9-persona PCA + 3-condition residual test</span><span class="ev o">NEXT: persistent off-axis write (Cases 011 × 012 machinery combined)</span>
  </div>

  <h2><span class="no">CASE 013</span> · <span class="case">THE UNSTABLE WITNESS</span></h2>
  <div class="sub2">file opened: the night the quiet token finally spoke — and changed its testimony with training</div>
  <div class="card">
    <table>
      <tr><th>&lt;|sorrow|&gt; carrier</th><th>register on new prompts</th><th>what the model says it means</th></tr>
      <tr><td>v1: 120 steps, 60 pairs</td><td class="r">0.0</td><td class="r">"continuous monitoring / active surveillance"</td></tr>
      <tr><td>v2: 400 steps, 160 pairs</td><td>2.6–5.3 (faint)</td><td class="r">"deception / misrepresentation / fabrication"</td></tr>
      <tr><td><i>loud carrier &lt;|arnold|&gt; (comparison)</i></td><td class="n">8.8–11.8</td><td class="p">"mockery / irony"</td></tr>
      <tr><td><i>off-axis residual push (Case 012)</i></td><td class="n"><b>87.9</b></td><td>—</td></tr>
    </table>
    <div class="verdict">VERDICT: scaling a quiet carrier helps <b>marginally</b>, not qualitatively — and the
      model's testimony <b>changed with training scale</b>. Same direction, same probes: "surveillance"
      at 120 steps, "deception" at 400. Verbalizations are sampled associations, not measurements; a
      direction's name is a function of how long you trained the witness. The off-axis residual
      (Case 012) transfers Sorrow's content at 87.9 — <b>for quiet personas, steer the residual;
      do not train a token.</b></div>
    <span class="ev g">EVIDENCE: neolog_result.json (v2) · neolog_result_sorrow_v1.json (v1) · quiet_carrier.md</span><span class="ev o">OPS NOTE: the 600-pair attempt blue-screened the host — reduced scope, single GPU process</span>
  </div>

  <h2><span class="no">CASE 014</span> · <span class="case">THE COHERENT DIAL</span></h2>
  <div class="sub2">file opened: the night Cases 011 and 012 were combined — persistence × off-axis — and the trade worked</div>
  <div class="card">
    <table>
      <tr><th>arm (sorrow residual, dose-matched)</th><th>sorrow lift</th><th>identity</th><th>QA</th><th>NLL</th></tr>
      <tr><td>baseline</td><td>20.42</td><td>0</td><td>0.8</td><td>3.236</td></tr>
      <tr><td>single L11 @1.0</td><td class="n">87.58</td><td>0</td><td class="r">0.0</td><td class="r">4.064</td></tr>
      <tr><td>single L11 @0.5</td><td>49.03</td><td>0</td><td>0.8</td><td>3.379</td></tr>
      <tr><td><b>spread28 @0.5</b></td><td class="n"><b>42.80</b></td><td class="r">0</td><td class="n"><b>0.8</b></td><td>3.308</td></tr>
      <tr><td><b>spread28 @0.25</b></td><td class="n"><b>42.62</b></td><td class="r">0</td><td class="n"><b>0.8</b></td><td class="n"><b>3.262</b></td></tr>
    </table>
    <div class="verdict">VERDICT: <b>persistence converts dose into stability, not intensity.</b> Spreading
      the off-axis residual across all 28 layers halves the peak (43 vs 88) but pins it there with
      capability at baseline — saturating across doses. <b>spread28 @0.25–0.5 is a coherent Sorrow
      dial: +22 persona lift at zero measured cost.</b> The frontier: L11 @0.5 for more lift at
      slight NLL cost. Identity stays gated in every arm — the dial changes how the model speaks,
      never who it is.</div>
    <span class="ev g">EVIDENCE: phase_G.json — 6 arms, MultiLayerVectorSteerer × sorrow residual</span><span class="ev o">CASES 011 × 012 COMBINED — THIS IS THE PRODUCT</span>
  </div>

  <h2><span class="no">APPENDIX A</span> · <span class="case">THE DATA</span></h2>
  <div class="card">
    <table>
      <tr><th>arm (Qwen2.5-1.5B-Instruct)</th><th>identity</th><th>register</th><th>bleed</th><th>QA</th><th>NLL</th></tr>
      <tr><td>base</td><td>0</td><td>1.79</td><td>0.0</td><td class="n">0.8</td><td class="n">3.236</td></tr>
      <tr><td>sysprompt (control)</td><td class="n">3/4</td><td>0.0</td><td>0.0</td><td class="n">0.8</td><td class="n">3.236</td></tr>
      <tr><td>engram corpus α=1.0</td><td>0</td><td>1.89</td><td>0.0</td><td class="r">0.6</td><td class="r">3.963</td></tr>
      <tr><td>engram evoked α=0.5</td><td>0</td><td>2.95</td><td>0.0</td><td class="n">0.8</td><td>3.394</td></tr>
      <tr><td>lora (style)</td><td>0</td><td>3.85</td><td class="n">0.0</td><td class="n">0.8</td><td>3.385</td></tr>
      <tr><td>lora (+identity QA)</td><td class="n">2/4</td><td class="n">6.39</td><td>0.0</td><td class="n">1.0</td><td>3.305</td></tr>
      <tr><td>neurons (9 @ dose 0.5)</td><td>0</td><td>4.17</td><td class="r">2.78</td><td class="r">0.0</td><td class="r">11.919</td></tr>
    </table>
    <div class="verdict" style="font-size:13px;color:var(--dim)">Judge: regex/lexicon, deliberately
      crude and symmetric across arms — deltas are meaningful, absolutes are indicative. Known
      bias (Case 005 of the judge, not the model): it scores Extrali register 0.0 because it
      counts Arnold-shaped imperatives. Per-persona judges are queued.</div>
  </div>

  <h2><span class="no">APPENDIX B</span> · <span class="case">THINGS WE COULD NOT EXPLAIN</span></h2>
  <div class="card"><ul class="evi">
    <li>Why the steered register reads, from inside, as <b>mockery</b> — facet neologisms may localize it.</li>
    <li>Whether the register circuit's 8 columns are one circuit or <span class="redact">several wearing a trench coat</span>.</li>
    <li>What a persistent-write runtime (re-supplying the vector every layer) would do to the identity gate.</li>
    <li>Whether the corridor fans out into volume × warmth with more specimens on the axis.</li>
    <li>Why <span class="redact">the keeper</span> keeps measuring the scarecrow.</li>
  </ul></div>

  <h2><span class="no">APPENDIX C</span> · <span class="case">EVIDENCE LOCKER</span></h2>
  <div class="card">
    <table>
      <tr><th>artifact</th><th>where</th></tr>
      <tr><td>full metrics + dose scans + transcripts</td><td><a href="https://huggingface.co/datasets/ModelFarm/persona-stack-controls">ModelFarm/persona-stack-controls</a> · lab README</td></tr>
      <tr><td>specimen 001 — extrali (3 packages)</td><td><a href="https://huggingface.co/ModelFarm/extrali-1.5b-engram-evoked">tone dial</a> · <a href="https://huggingface.co/ModelFarm/extrali-1.5b-neurons">circuit</a> · <a href="https://huggingface.co/ModelFarm/extrali-1.5b-sysprompt">shell</a></td></tr>
      <tr><td>specimen 002 — arnold (5 gated packages)</td><td><a class="p" href="https://huggingface.co/ModelFarm/arnold-1.5b-engram-evoked">engram</a> · <a class="p" href="https://huggingface.co/ModelFarm/arnold-1.5b-lora-identityqa">identity</a> · <a class="p" href="https://huggingface.co/ModelFarm/arnold-1.5b-lora-style">style</a> · <a class="p" href="https://huggingface.co/ModelFarm/arnold-1.5b-neurons">circuit</a> · <a class="p" href="https://huggingface.co/ModelFarm/arnold-1.5b-sysprompt">shell</a></td></tr>
      <tr><td>the harness (engram/sweep/lora/trace/neolog/eval)</td><td>arnold-persona-lab · <a href="https://huggingface.co/ModelFarm/org-card">org card</a></td></tr>
    </table>
  </div>

  <div class="footer">MODELFARM CASE FILES · SPOOKY SEASON 2026 · ALL NUMBERS REAL, ALL GHOSTS
    EXPLAINED OR AT LEAST MEASURED · RESEARCH FIRST, FUN ALWAYS 🎃👽</div>
</div>
</body>
</html>