Describe the space-station fallback face with the actress panel

web/learnings.html

@@ -25,7 +25,7 @@ ul.find li{margin:6px 0}.lead{font-weight:600}
 .note{border-left:3px solid var(--warn);padding:2px 10px;color:var(--ink2);font-size:14px}
 </style></head><body><main>
 <h1>Fixing the vertical colour stripes in MOSS-Image SFT: three approaches and what each teaches</h1>
-<p class="sub">Local research notes, updated 2026-10-19 22:36. Numbers come live from our screen (<code>output/moss/drift/scores.jsonl</code>). Image comparisons: <a href="index.html">compare page</a>. Inside-the-model charts: <a href="edit_anatomy.html">edit anatomy</a>. Nothing here is published yet.</p>
+<p class="sub">Local research notes, updated 2026-10-19 22:50. Numbers come live from our screen (<code>output/moss/drift/scores.jsonl</code>). Image comparisons: <a href="index.html">compare page</a>. Inside-the-model charts: <a href="edit_anatomy.html">edit anatomy</a>. Nothing here is published yet.</p>
 
 <div class="card tldr"><b>In short</b><ul>
 <li>MOSS-Image SFT paints faint, colour-only vertical stripes (about 48 cycles across the image, whatever the size) into smooth areas. Base doesn't.</li>
@@ -118,12 +118,12 @@ ul.find li{margin:6px 0}.lead{font-weight:600}
 
 <h2>5. Why this matters beyond MOSS: character LoRAs (DeTangle)</h2>
 <p><a href="https://matlowai.github.io/DeTangle/">DeTangle</a> asks the same question about a character LoRA: keep the character when triggered, change nothing when not. Its "A-side surgery" re-fits one LoRA factor from captured activations (match triggered prompts, stay near zero on generic ones, with a ridge), the same family as our Gram solve. It also has the same weakness: it picks by <i>what inputs look like</i>, not by <i>what the edit does to the image</i>.</p>
-<div class="card"><b>The space-station case.</b> Likeness drops from 0.67 (original LoRA) to about 0.00 after A-side surgery at strength 1, no better than Base without the LoRA (−0.03). Looking at the renders, Dylan noticed the character turning into a Sigourney-Weaver-like face (it resembles her but isn't her; see below). A plausible reading: for "a woman on a space station" the base model has a strong prior (Ripley), and the part of the LoRA that overrides it lives in directions the surgery classed as generic, because generic prompts use them too. When the override is removed, the prior wins. <br><br><b>Checked (2026-10-19, existing renders, ArcFace between each triggered render and Base's render of the same prompt and seed):</b> after A-side surgery at strength 1, the face matches Base's own face at <b>0.75</b> on space station and <b>0.66</b> on wizard workshop, against 0.04–0.36 in the other six contexts. At 1.25: 0.72 and 0.56 vs 0.00–0.21. The original LoRA's faces sit at only 0.11–0.24 from Base's. So in exactly the two contexts where likeness collapses, the surgery doesn't just weaken the character; it hands the face back to the base model's default for that scene. That is entanglement in the literal sense: in that context the identity signal and the context signal share input directions, so an input-statistics edit can't separate them. The stubborn blue solid in MOSS is a smaller version of the same thing.</div>
+<div class="card"><b>The space-station case.</b> Likeness drops from 0.67 (original LoRA) to about 0.00 after A-side surgery at strength 1, no better than Base without the LoRA (−0.03). Looking at the renders, Dylan noticed the character turning into a Sigourney-Weaver-like face (a real resemblance, within a blend of space-film actresses; see below). A plausible reading: for "a woman on a space station" the base model has a strong prior (Ripley), and the part of the LoRA that overrides it lives in directions the surgery classed as generic, because generic prompts use them too. When the override is removed, the prior wins. <br><br><b>Checked (2026-10-19, existing renders, ArcFace between each triggered render and Base's render of the same prompt and seed):</b> after A-side surgery at strength 1, the face matches Base's own face at <b>0.75</b> on space station and <b>0.66</b> on wizard workshop, against 0.04–0.36 in the other six contexts. At 1.25: 0.72 and 0.56 vs 0.00–0.21. The original LoRA's faces sit at only 0.11–0.24 from Base's. So in exactly the two contexts where likeness collapses, the surgery doesn't just weaken the character; it hands the face back to the base model's default for that scene. That is entanglement in the literal sense: in that context the identity signal and the context signal share input directions, so an input-statistics edit can't separate them. The stubborn blue solid in MOSS is a smaller version of the same thing.</div>
 <div class="card"><b>Probe results (2026-10-19, 424 renders, same protocol as the DeTangle study; details in <code>delta-surgery/runs/2026-10-19-entanglement-probes/RESULTS.md</code>):</b><ul>
 <li><b>Where:</b> only blocks 8–15 matter. Swapping the original LoRA back into just blocks 8–11 or 12–15 restores likeness (0.36–0.55); every other window, including blocks 0–3, does nothing.</li>
 <li><b>Entangled:</b> those same blocks bring the bleed back (+0.19 to +0.30 generic excess). The parts that let the character beat a strong scene prior are the parts that leak her into untriggered prompts.</li>
 <li><b>Context:</b> the "science-fiction film still" sentence is a real but partial cause. Adding it to arcade or beach cuts post-surgery likeness by 0.15–0.26 while the original LoRA doesn't care; stripping all scene-specific phrasing recovers 0.58 at strength 1.25.</li>
-<li><b>Sigourney:</b> against the model's concept of Sigourney Weaver (her studio portraits rendered by name), the fallback face is a different person: 0.33, vs 0.75–0.82 for her own renders. The base model's default woman leans slightly toward her in every scene (+0.07 to +0.15 over four other actresses, beach and arcade included), so it's not a sci-fi-specific association; surgery returns the generic woman and the sci-fi styling makes her read as Ripley.</li>
+<li><b>Sigourney:</b> the fallback face is a consistent lookalike of a blend of space-film actresses, not one person. Against the model's own concepts of 19 actresses (studio portraits by name; different actresses sit 0.03–0.38 apart, an actress's own renders 0.75+), it is closest to Rebecca Ferguson 0.42, Noomi Rapace 0.36, Kate Winslet 0.34 and Sigourney Weaver 0.32. The model's generic woman leans Ferguson everywhere; the sci-fi film-still scene adds Rapace, Weaver and Linda Hamilton. Surgery returns that inherited prior.</li>
 <li><b>Weak trigger:</b> after surgery the trigger loses to any explicitly described identity (named actress: likeness ≈0; ethnicity descriptors: 0.01–0.31).</li>
 </ul></div>
 <p>nynxz's method suggests the fix: score LoRA directions by their effect on the outcomes you care about (likeness when triggered, identity bleed when not), measured through sampling in each context, and keep or drop directions on that basis.</p>