case-files / index.html
chadstthomas's picture
Upload folder using huggingface_hub
33527cb verified
Raw History Blame Contribute Delete
28.9 kB
<!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>