Spaces:
Running
Running
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 & 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><|arnold|></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 > 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 <|sorrow|> 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><|arnold|> (loud, imperative)</td><td class="n">YES — register 8.8–11.8 vs 0</td><td class="r">"mockery / irony"</td></tr>
<tr><td><|sorrow|> (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 (<|sorrow|> 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><|sorrow|> 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 <|arnold|> (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>
|