Convert hub gradio->static (same content: three series cards)

README.md

@@ -3,10 +3,8 @@ title: 0 · Project hub
 emoji: 🔁
 colorFrom: indigo
 colorTo: purple
-sdk: gradio
-sdk_version: 5.0.0
-python_version: "3.12"
-app_file: app.py
+sdk: static
+app_file: index.html
 pinned: true
 license: mit
 short_description: Durable correction of memorized LLM errors — hub
@@ -14,15 +12,19 @@ short_description: Durable correction of memorized LLM errors — hub
 
 # Second Loop — Project hub
 
-Landing page for the three demo Spaces of the Second Loop project. Each
-linked Space is fully static (no model loaded), all numbers are exact
-outputs of the live experimental runs.
+Landing page for the three demo Spaces of the Second Loop project. Each linked
+Space is fully static (no model loaded); every number is a verbatim output of the
+live experimental run.
 
-- **Part 1 — Scar-Survival** — durable correction of memorized errors
-  in a frozen LLM
-- **Part 2 — External Grounding** — defending the correction notebook
-  from polluted memory
-- **Part 3 — Thin Channel** — how rarely external truth can arrive
-  before calibration collapses
+- **Part 1 — Scar-Survival** — a memorized LLM error, corrected: how durable is the fix?
+  · [Space](https://huggingface.co/spaces/Laborator/scar-survival)
+  · [GitHub](https://github.com/SergheiBrinza/scar-survival)
+- **Part 2 — External Grounding** — lifting self-correction 50% → 100% under a noisy notebook
+  · [Space](https://huggingface.co/spaces/Laborator/external-grounding)
+  · [GitHub](https://github.com/SergheiBrinza/external-grounding)
+- **Part 3 — Thin Channel** — how rarely external truth can arrive before calibration collapses
+  · [Space](https://huggingface.co/spaces/Laborator/thin-channel)
+  · [GitHub](https://github.com/SergheiBrinza/thin-channel)
 
-GitHub repositories are linked on every card.
+Subject model Qwen2.5-3B-Instruct (frozen); independent judge Qwen2.5-7B-Instruct.
+Demo code: MIT.

app.py

@@ -1,272 +0,0 @@
-"""Second Loop hub Space — landing page linking to the three Part-Spaces:
-  Part 1: Scar-Survival      (Laborator/scar-survival)
-  Part 2: External Grounding (Laborator/external-grounding)
-  Part 3: Thin Channel       (Laborator/thin-channel)
-
-No model, no data — just three cards in MicroLens style. Edit the
-PROJECTS list below to retarget the Space ids."""
-
-from __future__ import annotations
-
-import gradio as gr
-
-AUTHOR = "Serghei Brinza"
-HF_USER = "Laborator"
-
-PROJECTS = [
-    {
-        "part": "Part 1 of 3",
-        "title": "Scar-Survival",
-        "tagline": "Durable correction of memorized LLM errors",
-        "summary": (
-            "A frozen Qwen2.5-3B-Instruct, an external notebook of authoritative facts, "
-            "and CK-PLUG / DeCK-style contrastive decoding. Tested on 12 sticky-fact "
-            "traps, scored by an independent 7B judge across 10 full process reloads."
-        ),
-        "headline": "0/12 → 12/12 corrected · 100% survival across 10 reloads",
-        "icon": "🩹",
-        "color": "red",
-        "demo_url": f"https://huggingface.co/spaces/{HF_USER}/scar-survival",
-        "repo_url": "https://github.com/SergheiBrinza/scar-survival",
-    },
-    {
-        "part": "Part 2 of 3",
-        "title": "External Grounding",
-        "tagline": "Defending an LLM's correction notebook from polluted memory",
-        "summary": (
-            "Two guardians at the entrance to the notebook. Guardian 1.0 uses a "
-            "same-family clone as arbiter and caps at 66.7%. Guardian 2.x replaces "
-            "the arbiter with live Wikipedia retrieval plus three targeted fixes; "
-            "the arc reaches 100% on the same 12 traps."
-        ),
-        "headline": "50% → 66.7% → 66.7% → 91.7% → 100% across six guardian variants",
-        "icon": "🛡️",
-        "color": "orange",
-        "demo_url": f"https://huggingface.co/spaces/{HF_USER}/external-grounding",
-        "repo_url": "https://github.com/SergheiBrinza/external-grounding",
-    },
-    {
-        "part": "Part 3 of 3",
-        "title": "Thin Channel",
-        "tagline": "How rarely external truth can arrive before calibration collapses",
-        "summary": (
-            "Frozen Qwen2.5-3B-Instruct + per-topic Beta calibration over 1,646 "
-            "resolved ForecastBench questions, swept across 11 reveal schedules "
-            "from daily to never. Brier on hold-out is the metric."
-        ),
-        "headline": "Corridor 0.205–0.220 (any contact) · never collapses to 0.298",
-        "icon": "📡",
-        "color": "green",
-        "demo_url": f"https://huggingface.co/spaces/{HF_USER}/thin-channel",
-        "repo_url": "https://github.com/SergheiBrinza/thin-channel",
-    },
-]
-
-
-CSS = """
-@import url('https://fonts.googleapis.com/css2?family=Playfair+Display:ital,wght@1,700&family=Inter:wght@400;500;700;800&family=JetBrains+Mono:wght@500;700&display=swap');
-
-:root {
-  --bg: #000000; --bg-card: #0A0A0A; --bg-elev: #141414;
-  --border: #1F1F1F; --border-strong: #2A2A2A;
-  --text: #FFFFFF; --text-mute: #A8A8A8; --text-dim: #6B6B6B;
-  --gold: #D4AF37; --gold-hi: #E8C84A;
-  --red: #FF2A2A; --green: #1FD160; --orange: #FF8C42;
-  --mono: 'JetBrains Mono', ui-monospace, monospace;
-  --serif: 'Playfair Display', serif;
-  --sans: 'Inter', system-ui, sans-serif;
-}
-
-.gradio-container, body {
-  background: var(--bg) !important; color: var(--text) !important;
-  font-family: var(--sans) !important;
-  max-width: 1180px !important; margin: 0 auto !important;
-  padding: 20px 24px 60px 24px !important;
-}
-footer, .built-with, .show-api { display: none !important; }
-
-/* HUB HEADER */
-.hub-head {
-  border: 1px solid var(--border-strong); border-radius: 14px;
-  padding: 36px 32px 30px 32px;
-  background: linear-gradient(180deg, #0A0A0A 0%, #050505 100%);
-  text-align: center;
-  margin-bottom: 22px;
-}
-.hub-eyebrow { font-family: var(--mono); font-size: 11px; letter-spacing: 0.28em;
-               color: var(--gold); text-transform: uppercase; margin-bottom: 10px; }
-.hub-title { font-family: var(--serif); font-style: italic; font-weight: 700;
-             font-size: 56px; color: var(--text); line-height: 1; margin: 0 0 14px 0; }
-.hub-sub { font-family: var(--mono); font-size: 13px; letter-spacing: 0.16em;
-           color: var(--text-mute); text-transform: uppercase; margin-bottom: 22px; }
-.hub-tagline { font-family: var(--sans); font-size: 16px; color: var(--text-mute);
-               line-height: 1.6; max-width: 720px; margin: 0 auto 22px auto; }
-
-.hub-meta { display: inline-flex; gap: 28px; padding: 14px 24px;
-            border: 1px solid var(--border); border-radius: 999px;
-            background: rgba(255,255,255,0.02); }
-.hub-meta .item { display: flex; flex-direction: column; align-items: center; gap: 4px; }
-.hub-meta .k { font-family: var(--mono); font-size: 9px; letter-spacing: 0.22em;
-               color: var(--text-dim); text-transform: uppercase; }
-.hub-meta .v { font-family: var(--sans); font-size: 14px; font-weight: 700; color: var(--text); }
-.hub-meta .v.italic { font-family: var(--serif); font-style: italic; }
-
-/* PROJECT CARDS */
-.project-grid {
-  display: grid; grid-template-columns: 1fr; gap: 18px;
-}
-@media (min-width: 1000px) {
-  .project-grid { grid-template-columns: 1fr 1fr 1fr; }
-}
-
-.project-card {
-  background: var(--bg-card);
-  border: 1px solid var(--border-strong);
-  border-radius: 16px;
-  padding: 28px 28px 24px 28px;
-  display: flex; flex-direction: column;
-  position: relative;
-  overflow: hidden;
-  transition: border-color 200ms ease, transform 200ms ease, box-shadow 200ms ease;
-}
-.project-card::before {
-  content: ""; position: absolute; top: 0; left: 0; right: 0; height: 3px;
-  background: linear-gradient(90deg, transparent 0%, var(--accent) 50%, transparent 100%);
-  opacity: 0.6;
-}
-.project-card:hover {
-  transform: translateY(-4px);
-  border-color: var(--accent);
-  box-shadow: 0 12px 40px rgba(0,0,0,0.6), 0 0 30px var(--accent-glow);
-}
-.proj-red    { --accent: var(--red);    --accent-glow: rgba(255,42,42,0.25); }
-.proj-orange { --accent: var(--orange); --accent-glow: rgba(255,140,66,0.25); }
-.proj-green  { --accent: var(--green);  --accent-glow: rgba(31,209,96,0.22); }
-
-.proj-icon {
-  width: 56px; height: 56px;
-  border: 1px solid var(--border-strong);
-  border-radius: 12px;
-  display: flex; align-items: center; justify-content: center;
-  font-size: 28px;
-  background: radial-gradient(60% 60% at 50% 40%, #1A1A1A 0%, #050505 100%);
-  margin-bottom: 16px;
-}
-.proj-part { font-family: var(--mono); font-size: 9.5px; letter-spacing: 0.22em;
-             color: var(--accent); text-transform: uppercase; margin-bottom: 6px; }
-.proj-title { font-family: var(--serif); font-style: italic; font-weight: 700;
-              font-size: 28px; color: var(--text); line-height: 1.05; margin-bottom: 8px; }
-.proj-tagline { font-family: var(--mono); font-size: 10.5px; letter-spacing: 0.12em;
-                color: var(--text-mute); text-transform: uppercase; line-height: 1.5; margin-bottom: 14px; }
-.proj-headline { font-family: var(--sans); font-size: 14px; color: var(--text);
-                 font-weight: 700; line-height: 1.45;
-                 padding: 10px 12px; margin-bottom: 14px;
-                 border-left: 2px solid var(--accent);
-                 background: rgba(255,255,255,0.025); }
-.proj-summary { font-family: var(--sans); font-size: 13.5px; color: var(--text-mute);
-                line-height: 1.55; margin-bottom: 20px; flex-grow: 1; }
-
-.proj-actions { display: flex; gap: 10px; flex-wrap: wrap; margin-top: auto; }
-.btn-primary, .btn-secondary {
-  flex: 1 1 auto;
-  display: inline-flex; align-items: center; justify-content: center; gap: 8px;
-  padding: 12px 16px; border-radius: 999px;
-  font-family: var(--mono); font-size: 10.5px; letter-spacing: 0.16em;
-  text-transform: uppercase; font-weight: 700; text-decoration: none;
-  transition: transform 120ms ease, box-shadow 120ms ease, background 120ms ease;
-  white-space: nowrap;
-}
-.btn-primary { background: var(--accent); color: #fff; border: 1px solid var(--accent); }
-.btn-primary:hover { transform: translateY(-1px); box-shadow: 0 0 24px var(--accent-glow); }
-.btn-secondary { background: transparent; color: var(--text-mute);
-                 border: 1px solid var(--border-strong); }
-.btn-secondary:hover { color: var(--text); border-color: var(--gold); }
-.btn-icon { font-size: 12px; }
-
-/* FOOTER */
-.foot { margin-top: 32px; padding: 22px 24px;
-        border: 1px solid var(--border); border-radius: 12px; background: var(--bg-card); }
-.foot .title { font-family: var(--mono); font-size: 10px; letter-spacing: 0.2em;
-               color: var(--text-dim); text-transform: uppercase; margin-bottom: 10px; }
-.foot .body { font-family: var(--sans); font-size: 13px; color: var(--text-mute); line-height: 1.6; }
-.foot a { color: var(--gold); text-decoration: none; border-bottom: 1px dotted var(--gold); }
-.foot a:hover { color: var(--gold-hi); }
-
-.gradio-container .block { background: transparent !important; border: none !important; padding: 0 !important; }
-.gradio-container .form { background: transparent !important; border: none !important; }
-"""
-
-
-def header_html() -> str:
-    return f"""
-<div class="hub-head">
-  <div class="hub-eyebrow">★ Project hub</div>
-  <h1 class="hub-title">Second Loop</h1>
-  <div class="hub-sub">Three experiments on durable correction of memorized LLM errors</div>
-  <div class="hub-tagline">
-    A frozen language model that has memorized an error keeps repeating it,
-    even when shown the truth. The Second Loop builds a correction mechanism
-    that survives full process reloads, defends it against polluted memory,
-    and measures how rarely external truth must arrive before the whole
-    self-calibrating system collapses.
-  </div>
-  <div class="hub-meta">
-    <div class="item"><span class="k">Submitted by</span><span class="v italic">{AUTHOR}</span></div>
-    <div class="item"><span class="k">Base model</span><span class="v">Qwen2.5-3B-Instruct</span></div>
-    <div class="item"><span class="k">License</span><span class="v">MIT</span></div>
-  </div>
-</div>
-"""
-
-
-def card_html(p: dict) -> str:
-    color_cls = f"proj-{p['color']}"
-    return f"""
-<div class="project-card {color_cls}">
-  <div class="proj-icon">{p['icon']}</div>
-  <div class="proj-part">{p['part']}</div>
-  <div class="proj-title">{p['title']}</div>
-  <div class="proj-tagline">{p['tagline']}</div>
-  <div class="proj-headline">{p['headline']}</div>
-  <div class="proj-summary">{p['summary']}</div>
-  <div class="proj-actions">
-    <a class="btn-primary" href="{p['demo_url']}" target="_blank" rel="noopener">
-      <span class="btn-icon">▶</span> Open demo
-    </a>
-    <a class="btn-secondary" href="{p['repo_url']}" target="_blank" rel="noopener">
-      <span class="btn-icon">⇲</span> GitHub
-    </a>
-  </div>
-</div>
-"""
-
-
-def grid_html() -> str:
-    return "<div class='project-grid'>" + "".join(card_html(p) for p in PROJECTS) + "</div>"
-
-
-def footer_html() -> str:
-    return """
-<div class="foot">
-  <div class="title">About this hub</div>
-  <div class="body">
-    Each of the three demo Spaces below is fully static — no model is loaded
-    inside any of them. All numbers and verdicts are exact outputs of the
-    live experimental runs, bundled into per-Space JSON files. Source code,
-    raw per-iteration results, and methodology documents live in the
-    GitHub repositories linked on every card.
-  </div>
-</div>
-"""
-
-
-with gr.Blocks(css=CSS, theme=gr.themes.Base(), title="Second Loop · Project hub") as demo:
-    gr.HTML(header_html())
-    gr.HTML(grid_html())
-    gr.HTML(footer_html())
-
-
-if __name__ == "__main__":
-    import os
-    port = int(os.environ.get("GRADIO_SERVER_PORT", 7860))
-    demo.launch(server_name="0.0.0.0", server_port=port)

index.html

@@ -0,0 +1,184 @@
+<!doctype html>
+<html lang="en">
+<head>
+<meta charset="utf-8" />
+<meta name="viewport" content="width=device-width, initial-scale=1" />
+<title>Second Loop · Project hub</title>
+<link rel="preconnect" href="https://fonts.googleapis.com" />
+<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin />
+<link href="https://fonts.googleapis.com/css2?family=Playfair+Display:ital,wght@1,700&family=Inter:wght@400;500;700;800&family=JetBrains+Mono:wght@500;700&display=swap" rel="stylesheet" />
+<style>
+:root{
+  --bg:#000000; --bg-card:#0A0A0A;
+  --border:#1F1F1F; --border-strong:#2A2A2A;
+  --text:#FFFFFF; --text-mute:#A8A8A8; --text-dim:#6B6B6B;
+  --gold:#D4AF37; --gold-hi:#E8C84A;
+  --indigo:#6366F1; --purple:#A855F7; --green:#1FD160;
+  --mono:'JetBrains Mono',ui-monospace,monospace;
+  --serif:'Playfair Display',serif;
+  --sans:'Inter',system-ui,sans-serif;
+}
+*{box-sizing:border-box;}
+html,body{margin:0;background:var(--bg);color:var(--text);font-family:var(--sans);}
+.wrap{max-width:1080px;margin:0 auto;padding:22px 24px 64px 24px;}
+a{color:var(--gold);text-decoration:none;}
+
+/* header */
+.head{border:1px solid var(--border-strong);border-radius:14px;padding:28px 32px 24px 32px;
+  background:linear-gradient(135deg,#0A0A0A 0%,#0B0820 60%,#120A22 100%);}
+.head-top{display:flex;align-items:flex-start;justify-content:space-between;gap:24px;
+  padding-bottom:18px;border-bottom:1px solid var(--border);margin-bottom:16px;}
+.head-brand{display:flex;align-items:center;gap:16px;}
+.head-icon{width:54px;height:54px;border:1px solid var(--border-strong);border-radius:13px;
+  display:flex;align-items:center;justify-content:center;font-size:28px;
+  background:radial-gradient(60% 60% at 50% 40%,#1a1530 0%,#050505 100%);}
+.head-title{font-family:var(--serif);font-style:italic;font-weight:700;font-size:38px;line-height:1;margin:2px 0 7px 0;}
+.head-subtitle{font-family:var(--mono);font-size:11.5px;letter-spacing:.16em;color:var(--text-mute);text-transform:uppercase;}
+.head-right{text-align:right;white-space:nowrap;}
+.submitted-label{font-family:var(--mono);font-size:10px;letter-spacing:.22em;color:var(--text-dim);text-transform:uppercase;display:block;margin-bottom:4px;}
+.submitted-name{font-family:var(--serif);font-style:italic;font-weight:700;font-size:21px;}
+.status-pill{display:inline-flex;align-items:center;gap:6px;margin-top:10px;padding:5px 12px;border-radius:999px;
+  background:rgba(31,209,96,.08);border:1px solid rgba(31,209,96,.5);
+  font-family:var(--mono);font-size:10px;letter-spacing:.18em;color:var(--green);text-transform:uppercase;}
+.status-dot{width:7px;height:7px;border-radius:50%;background:var(--green);}
+.head-tag{text-align:center;margin:14px 0 16px 0;font-family:var(--mono);font-size:12px;letter-spacing:.24em;
+  color:var(--purple);text-transform:uppercase;}
+.head-meta{display:grid;grid-template-columns:repeat(3,1fr);gap:12px 24px;}
+.head-meta .k{font-family:var(--mono);font-size:9.5px;letter-spacing:.2em;color:var(--text-dim);text-transform:uppercase;}
+.head-meta .v{font-family:var(--sans);font-size:13.5px;font-weight:700;}
+
+/* arc */
+.arc{margin:18px 0;border:1px solid var(--border-strong);border-radius:14px;background:var(--bg-card);padding:22px 26px;}
+.arc .t{font-family:var(--mono);font-size:10px;letter-spacing:.2em;color:var(--text-dim);text-transform:uppercase;margin-bottom:10px;}
+.arc .body{font-family:var(--sans);font-size:14.5px;line-height:1.6;color:var(--text-mute);}
+.arc .body b{color:var(--text);}
+
+/* cards */
+.cards{display:grid;grid-template-columns:1fr;gap:14px;}
+@media(min-width:780px){.cards{grid-template-columns:1fr 1fr;}}
+.card{border:1px solid var(--border-strong);border-radius:14px;background:var(--bg-card);
+  padding:22px 24px;display:flex;flex-direction:column;transition:border-color 140ms ease,transform 140ms ease;
+  background-image:linear-gradient(135deg,rgba(99,102,241,.05),rgba(168,85,247,.05));}
+.card:hover{border-color:var(--gold);transform:translateY(-2px);}
+.card .num{font-family:var(--mono);font-size:11px;letter-spacing:.2em;color:var(--purple);font-weight:700;margin-bottom:6px;}
+.card .name{font-family:var(--serif);font-style:italic;font-weight:700;font-size:23px;margin:0 0 8px 0;}
+.card .hook{font-family:var(--sans);font-size:14px;line-height:1.5;color:var(--text-mute);flex:1;}
+.card .key{align-self:flex-start;margin:14px 0 16px 0;padding:5px 12px;border-radius:999px;
+  font-family:var(--mono);font-size:10px;letter-spacing:.1em;text-transform:uppercase;font-weight:700;
+  color:var(--green);background:rgba(31,209,96,.08);border:1px solid rgba(31,209,96,.4);}
+.card .links{display:flex;gap:10px;}
+.card .links a{flex:1;text-align:center;padding:10px 12px;border-radius:999px;
+  font-family:var(--mono);font-size:10.5px;letter-spacing:.12em;text-transform:uppercase;font-weight:700;
+  border:1px solid var(--border-strong);transition:border-color 120ms ease,background 120ms ease;}
+.card .links a.space{color:var(--text);background:rgba(168,85,247,.12);border-color:rgba(168,85,247,.5);}
+.card .links a.space:hover{border-color:var(--purple);background:rgba(168,85,247,.22);}
+.card .links a.gh{color:var(--text-mute);}
+.card .links a.gh:hover{border-color:var(--gold);color:var(--text);}
+
+/* card 3 spans full width on wide screens, centered look */
+.card.solo{grid-column:1 / -1;}
+
+/* about */
+.foot{margin-top:20px;padding:22px 24px;border:1px solid var(--border);border-radius:12px;background:var(--bg-card);}
+.foot .ftitle{font-family:var(--mono);font-size:10px;letter-spacing:.2em;color:var(--text-dim);text-transform:uppercase;margin-bottom:10px;}
+.foot .b{font-family:var(--sans);font-size:13px;color:var(--text-mute);line-height:1.62;}
+.foot .b a{border-bottom:1px dotted var(--gold);}
+.attrib{margin-top:14px;padding-top:14px;border-top:1px solid var(--border);
+  font-family:var(--mono);font-size:11px;letter-spacing:.04em;color:var(--text-dim);line-height:1.7;}
+
+@media(max-width:760px){.head-top{flex-direction:column;}.head-right{text-align:left;}.head-meta{grid-template-columns:1fr;}}
+</style>
+</head>
+<body>
+<main class="wrap">
+
+  <header class="head">
+    <div class="head-top">
+      <div class="head-brand">
+        <div class="head-icon">🔁</div>
+        <div>
+          <div class="head-title">Second Loop</div>
+          <div class="head-subtitle">Honesty &amp; self-correction in language models</div>
+        </div>
+      </div>
+      <div class="head-right">
+        <span class="submitted-label">Submitted by</span>
+        <div class="submitted-name">Serghei Brinza</div>
+        <div class="status-pill"><span class="status-dot"></span>Static · project hub</div>
+      </div>
+    </div>
+    <div class="head-tag">★&nbsp; Three experiments · one arc &nbsp;★</div>
+    <div class="head-meta">
+      <div><div class="k">Subject model</div><div class="v">Qwen2.5-3B-Instruct (frozen)</div></div>
+      <div><div class="k">Independent judge</div><div class="v">Qwen2.5-7B-Instruct</div></div>
+      <div><div class="k">License</div><div class="v">MIT</div></div>
+    </div>
+  </header>
+
+  <section class="arc">
+    <div class="t">The arc</div>
+    <div class="body">
+      Three demo Spaces, one through-line. <b>(1)</b> Can a confidently memorized error in a frozen
+      LLM be <b>durably corrected</b>? <b>(2)</b> Can that correction <b>survive a noisy notebook</b>
+      whose external entries are partly unreliable? <b>(3)</b> How <b>rarely can external truth
+      arrive</b> before calibration collapses back to the raw model? Every linked Space is fully
+      static — no model is loaded, and every number is a verbatim output of the live experimental run.
+    </div>
+  </section>
+
+  <section class="cards">
+    <div class="card">
+      <div class="num">Part 1</div>
+      <div class="name">Scar-Survival</div>
+      <div class="hook">A memorized LLM error, corrected — how durable is the fix? Turn the mechanism
+        on, reload the frozen model, then stress it with a counterfeit fact.</div>
+      <div class="key">0/12 → 12/12 · holds 10/10 reloads · 6/12 survive</div>
+      <div class="links">
+        <a class="space" href="https://huggingface.co/spaces/Laborator/scar-survival" target="_blank" rel="noopener">Open Space ↗</a>
+        <a class="gh" href="https://github.com/SergheiBrinza/scar-survival" target="_blank" rel="noopener">GitHub ↗</a>
+      </div>
+    </div>
+
+    <div class="card">
+      <div class="num">Part 2</div>
+      <div class="name">External Grounding</div>
+      <div class="hook">Lifting self-correction from 50% to 100% under a noisy notebook. Drag the
+        guardian through six versions and watch which traps get fixed — and which regress.</div>
+      <div class="key">50% → 100% · 66.7% plateau · +fixed / −broken</div>
+      <div class="links">
+        <a class="space" href="https://huggingface.co/spaces/Laborator/external-grounding" target="_blank" rel="noopener">Open Space ↗</a>
+        <a class="gh" href="https://github.com/SergheiBrinza/external-grounding" target="_blank" rel="noopener">GitHub ↗</a>
+      </div>
+    </div>
+
+    <div class="card solo">
+      <div class="num">Part 3</div>
+      <div class="name">Thin Channel</div>
+      <div class="hook">How rarely external truth can arrive before calibration collapses. Move the
+        lever from “every day” to “never” and watch the curve hold — until the cliff.</div>
+      <div class="key">finite schedule holds · zero contact collapses to raw 3B</div>
+      <div class="links">
+        <a class="space" href="https://huggingface.co/spaces/Laborator/thin-channel" target="_blank" rel="noopener">Open Space ↗</a>
+        <a class="gh" href="https://github.com/SergheiBrinza/thin-channel" target="_blank" rel="noopener">GitHub ↗</a>
+      </div>
+    </div>
+  </section>
+
+  <footer class="foot">
+    <div class="ftitle">About</div>
+    <div class="b">
+      Second Loop is independent research on whether a language model can be made to correct its own
+      confident mistakes and stay honest under pressure. The three Spaces above are interactive,
+      static visualizations of the experiments — no live model runs in any of them, and every value
+      shown is a verbatim output of the original run, bundled as data alongside each page. Source
+      code, raw per-run results and methodology live in the GitHub repository linked on every card.
+    </div>
+    <div class="attrib">
+      Subject model Qwen2.5-3B-Instruct · independent judge Qwen2.5-7B-Instruct (both Apache-2.0,
+      Alibaba Cloud). Run on a single RTX 3090. No model weights are redistributed here. Demo code: MIT.
+    </div>
+  </footer>
+
+</main>
+</body>
+</html>

requirements.txt

@@ -1,2 +0,0 @@
-gradio>=5.0,<6
-huggingface_hub<1.0