DarwinX project page

.gitattributes

@@ -33,3 +33,8 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+assets/full/overview.png filter=lfs diff=lfs merge=lfs -text
+assets/full/population.png filter=lfs diff=lfs merge=lfs -text
+assets/full/teaser.png filter=lfs diff=lfs merge=lfs -text
+assets/overview.png filter=lfs diff=lfs merge=lfs -text
+assets/population.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,10 +1,172 @@
 ---
-title: Darwinx
-emoji: 🏆
-colorFrom: pink
-colorTo: blue
+title: DarwinX
+emoji: 🧬
+colorFrom: blue
+colorTo: indigo
 sdk: static
+app_file: index.html
 pinned: false
+license: cc-by-4.0
+short_description: Evolving Agent Harnesses Through Natural Selection
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# DarwinX: Evolving Agent Harnesses Through Natural Selection
+
+Project page for [arXiv:2608.07545](https://arxiv.org/abs/2608.07545).
+
+DarwinX treats agent self-evolution as selection over a *population* of harnesses with the base
+model frozen. Across four benchmarks, one loop adds ~17 points on average:
+
+| Benchmark | Result | Gain | What it isolates |
+| --- | --- | --- | --- |
+| Terminal-Bench 2.1 (avg@5) | 84.7% | +7.7 on matched base | in-domain evolution |
+| TerminalWorld (held-out) | 68.3% | +7.3 | disjoint held-out task split |
+| WebArena-Infinity (audit-clean pass@1) | 93.0% | +49.5 pp | synthetic → real intent shift |
+| SWE-bench Verified (zero-shot transfer) | 84.2% | +3.4 | cross-benchmark transfer |
+
+## Page sections
+
+`Abstract` · `Method` · `Results at a glance` · `Benchmark detail` · `Ablation: what evolution
+changes` · `Limitations` · `BibTeX`
+
+The ablation section reproduces the paper's framing: it is an *exploratory attribution, not a
+per-skill causal ablation*, since the skills were co-selected rather than independently randomized.
+The limitations section is carried over from the paper's Discussion.
+
+## Interactive figures
+
+Eight figures are interactive, and each is driven by real run artifacts or the paper's own plotting
+scripts rather than illustrative numbers. `tools/build_data.py` regenerates `assets/data.js` from the
+original files, so nothing is hand-transcribed.
+
+**1. TerminalWorld merge explorer.** Toggle any subset of the four evolved specialists and watch
+which of the 41 held-out tasks the selection covers. Built from the `per_task_results` arrays of the
+five evaluation runs, joined on `task_id` (the run files list the tasks in different orders, so
+position-based joining would silently misalign them).
+
+It surfaces something the paper states only in aggregate. The four specialists solve 24/25/26/27
+tasks and their **union is 29**, but the harness recombination actually produced solves **28**:
+
+| | tasks |
+| --- | --- |
+| best single specialist (D) | 27/41 |
+| union of all four | 29/41 |
+| realized merge | 28/41 |
+| solved by merge, by no specialist | `tw_448247` |
+| solved by some specialist, not by merge | `tw_449421`, `tw_498533` |
+| solved by all four | 21/41 |
+| solved by none | 12/41 |
+
+So recombination is not a free set union: it adds a capability no parent had and loses two. These are
+single attempts and one task is worth 2.4 points, so individual flips sit inside the noise band —
+the page says so next to the widget.
+
+**2. TB2.1 cluster explorer.** Per-cluster base vs. evolved avg@5, sortable by gain, base rate,
+cluster size, or name, with exact rates on hover. Numbers from `notes/TB21_RESULTS.md` (paired
+protocol, 88 tasks). Deltas are carried over as reported rather than recomputed, because the source
+rounds them from unrounded rates.
+
+**3. WebArena-Infinity evolution curve.** All 37 evaluated variants with a best-so-far envelope,
+from `notes/tw_dynamics.json` → `wai_adaptive_scores`. Toggle either series.
+
+**4. Headline four-benchmark panels.** Replaces the right half of the teaser. From
+`scripts/gen_summary_figure.py`. The reason to make this one interactive is that the paper's figure
+has to *state* its truncation in the caption: on a shared axis WebArena-Infinity's +49.5 flattens the
+two terminal benchmarks into slivers, so each panel is scaled to its own range. Here the reader can
+switch between per-panel and shared 0–100 axes and see both framings. The best-prior-agent bar can be
+hidden, since those systems use different models and effort settings and are context rather than a
+controlled comparison. Per-panel labels carry the exceptions: on SWE-bench Verified the grey bar is
+the fix-skill reference, not an unevolved Monet, and there is no prior-agent bar.
+
+**5. TerminalWorld specialist bars.** From `notes/tw_dynamics.json` → `tw_heldout`. Same axis
+question in miniature: the six arms span 58.5–68.3, so a 0–100 track renders them near-identical.
+Defaults to a 55–70 axis with the full axis one click away and the truncation named in both notes.
+Spec. D and the Claude Code reference are both 27/41, so D's bar landing exactly on the dashed
+reference line is a built-in check that the axis transform is right.
+
+**6. TB2.1 compute.** From `scripts/gen_tb21_compute.py` (medians over clean attempts). Switch
+between turns and tokens; the bars share one scale across both task groups so the newly-solved vs.
+already-solved contrast is not rescaled away.
+
+**7. WAI invalid-trajectory composition.** From `scripts/gen_wai_invalid_composition.py`. Break the
+293 → 17 collapse down by application or by mechanism. Both rows sit on a shared absolute scale by
+default, which makes the "after" row a near-invisible sliver — that *is* the finding; normalizing
+each row to its own total then shows what the remainder consists of. `build_data.py` asserts both
+decompositions still total 293 and 17.
+
+**8. WAI audit dumbbells.** From `scripts/gen_wai_audit_by_app.py`. One line per application from raw
+pass@1 to audited pass@1, so line length reads directly as how much a harness was leaning on
+trajectories the audit rejects. Sortable by audit loss. `build_data.py` cross-checks the audited
+DarwinX column against the per-application table rendered elsewhere on the page.
+
+Three figures stay static because they are conceptual diagrams with no underlying data: the teaser's
+selection schematic, the method overview, and the per-generation operators. The archive lineage tree
+also stays static for a different reason — its node/edge data lives in a `state.db` on the cluster,
+not on this machine, so there is nothing truthful to make interactive yet.
+
+Tables are click-to-sort. Numeric columns open descending, text columns A–Z, and `Overall` rows stay
+pinned to the bottom.
+
+## Local preview
+
+```bash
+python3 -m http.server 8000
+# open http://localhost:8000
+```
+
+## Regenerate and test
+
+```bash
+python3 tools/build_data.py          # rebuild assets/data.js from the run artifacts
+# then, with the server running, open:
+#   http://localhost:8000/tools/interaction_test.html
+# it drives every widget with real click events and writes PASS/FAIL into the page title
+```
+
+Headless run:
+
+```bash
+python3 -m http.server 8000 &
+"/Applications/Google Chrome.app/Contents/MacOS/Google Chrome" --headless --disable-gpu \
+  --virtual-time-budget=6000 --dump-dom http://localhost:8000/tools/interaction_test.html \
+  | grep -o '<title>.*</title>'
+```
+
+## Layout
+
+```
+index.html                          the page: inline CSS + vanilla JS, no build step, no external
+                                    requests, no web fonts, no chart library
+assets/data.js                      generated — real numbers behind the interactive figures
+assets/*.png                        web-sized figures (max 2000px wide, 256-color) — 736 KB total
+assets/full/*.png                   full-resolution originals for the three dense diagrams,
+                                    linked from the page via click-to-zoom
+tools/build_data.py                 regenerates assets/data.js from the original run artifacts
+tools/interaction_test.html         click-driven test suite for the interactive figures (49 checks)
+```
+
+| Figure | Content |
+| --- | --- |
+| `teaser.png` | four-benchmark teaser (full figure, click-to-zoom target) |
+| `teaser_schematic.png` | teaser's left half — the selection schematic shown on the page |
+| `overview.png` | method overview |
+| `population.png` | per-generation operators (mutation, classification, merge) |
+| `tb21_category_gain.png` | per-cluster TB2.1 gain |
+| `tb21_compute.png` | per-task compute: newly-solved vs already-solved |
+| `tw_heldout_specialists.png` | held-out generalization vs archive diversity |
+| `wai_evolution_curve.png` | best-so-far screening score |
+| `wai_tree.png` | archive lineage tree |
+| `wai_invalid_composition.png` | invalid trajectories, 293 → 17 |
+| `wai_audit_by_app.png` | raw vs audited pass@1 per application |
+
+Result figures are rendered from the paper's PDF sources at 200 DPI, then palette-reduced without
+dithering (RMSE vs. truecolor < 0.5%, which is imperceptible for line art but 4.7x smaller).
+
+## Hosting
+
+The frontmatter above configures this directory as a Hugging Face **static Space**. The same files
+work unchanged as a GitHub Pages site or behind any static file server.
+
+One thing to change when hosting: `og:image` and `twitter:image` are relative paths, which most
+crawlers resolve but X/Twitter wants absolute. Replace them with the deployed absolute URL to get a
+large social card.

assets/data.js

@@ -0,0 +1,558 @@
+/* Generated by tools/build_data.py - do not edit by hand. */
+window.DARWINX = {
+ "generated": "2026-08-14 01:23 UTC",
+ "tw": {
+  "arms": [
+   {
+    "key": "A",
+    "label": "Spec. A"
+   },
+   {
+    "key": "B",
+    "label": "Spec. B"
+   },
+   {
+    "key": "C",
+    "label": "Spec. C"
+   },
+   {
+    "key": "D",
+    "label": "Spec. D"
+   },
+   {
+    "key": "M",
+    "label": "Merged"
+   }
+  ],
+  "tasks": [
+   {
+    "id": "tw_433818",
+    "bits": "11111"
+   },
+   {
+    "id": "tw_90030",
+    "bits": "00011"
+   },
+   {
+    "id": "tw_152535",
+    "bits": "11111"
+   },
+   {
+    "id": "tw_326675",
+    "bits": "11111"
+   },
+   {
+    "id": "tw_204417",
+    "bits": "11111"
+   },
+   {
+    "id": "tw_649919",
+    "bits": "11111"
+   },
+   {
+    "id": "tw_661947",
+    "bits": "11111"
+   },
+   {
+    "id": "tw_494670",
+    "bits": "00000"
+   },
+   {
+    "id": "tw_241711",
+    "bits": "00000"
+   },
+   {
+    "id": "tw_448247",
+    "bits": "00001"
+   },
+   {
+    "id": "tw_713365",
+    "bits": "00000"
+   },
+   {
+    "id": "tw_212956",
+    "bits": "11111"
+   },
+   {
+    "id": "tw_165299",
+    "bits": "11111"
+   },
+   {
+    "id": "tw_17407",
+    "bits": "00000"
+   },
+   {
+    "id": "tw_449421",
+    "bits": "00110"
+   },
+   {
+    "id": "tw_360189",
+    "bits": "11111"
+   },
+   {
+    "id": "tw_274399",
+    "bits": "11111"
+   },
+   {
+    "id": "tw_739013",
+    "bits": "10111"
+   },
+   {
+    "id": "tw_11884",
+    "bits": "00000"
+   },
+   {
+    "id": "tw_768161",
+    "bits": "00000"
+   },
+   {
+    "id": "tw_296822",
+    "bits": "11111"
+   },
+   {
+    "id": "tw_418270",
+    "bits": "11101"
+   },
+   {
+    "id": "tw_576752",
+    "bits": "11111"
+   },
+   {
+    "id": "tw_203087",
+    "bits": "01011"
+   },
+   {
+    "id": "tw_123683",
+    "bits": "00000"
+   },
+   {
+    "id": "tw_627639",
+    "bits": "11111"
+   },
+   {
+    "id": "tw_8593",
+    "bits": "11111"
+   },
+   {
+    "id": "tw_321885",
+    "bits": "11111"
+   },
+   {
+    "id": "tw_481830",
+    "bits": "11111"
+   },
+   {
+    "id": "tw_117921",
+    "bits": "11111"
+   },
+   {
+    "id": "tw_19076",
+    "bits": "11111"
+   },
+   {
+    "id": "tw_474864",
+    "bits": "00000"
+   },
+   {
+    "id": "tw_414193",
+    "bits": "00000"
+   },
+   {
+    "id": "tw_15649",
+    "bits": "11111"
+   },
+   {
+    "id": "tw_453279",
+    "bits": "01111"
+   },
+   {
+    "id": "tw_569867",
+    "bits": "11111"
+   },
+   {
+    "id": "tw_498533",
+    "bits": "00010"
+   },
+   {
+    "id": "tw_433816",
+    "bits": "11101"
+   },
+   {
+    "id": "tw_486982",
+    "bits": "11111"
+   },
+   {
+    "id": "tw_247983",
+    "bits": "00000"
+   },
+   {
+    "id": "tw_39289",
+    "bits": "00000"
+   }
+  ],
+  "counts": {
+   "A": 24,
+   "B": 25,
+   "C": 26,
+   "D": 27,
+   "M": 28
+  }
+ },
+ "tb21Clusters": [
+  {
+   "name": "ML & scientific computing",
+   "n": 23,
+   "base": 60.1,
+   "evolved": 74.9,
+   "delta": 14.8
+  },
+  {
+   "name": "Data & databases",
+   "n": 9,
+   "base": 83.9,
+   "evolved": 97.8,
+   "delta": 13.8
+  },
+  {
+   "name": "Algorithms & coding",
+   "n": 14,
+   "base": 83.1,
+   "evolved": 90.0,
+   "delta": 6.9
+  },
+  {
+   "name": "System admin & networking",
+   "n": 11,
+   "base": 92.0,
+   "evolved": 97.6,
+   "delta": 5.7
+  },
+  {
+   "name": "Parsing, text & media",
+   "n": 11,
+   "base": 67.3,
+   "evolved": 72.2,
+   "delta": 5.0
+  },
+  {
+   "name": "Systems & build",
+   "n": 11,
+   "base": 79.3,
+   "evolved": 81.8,
+   "delta": 2.5
+  },
+  {
+   "name": "Security & crypto",
+   "n": 9,
+   "base": 85.1,
+   "evolved": 84.4,
+   "delta": -0.6
+  }
+ ],
+ "waiCurve": [
+  19.67,
+  58.33,
+  41.67,
+  50.0,
+  50.0,
+  50.0,
+  50.0,
+  58.33,
+  70.0,
+  51.79,
+  58.33,
+  57.14,
+  51.79,
+  67.86,
+  55.56,
+  58.93,
+  60.71,
+  70.0,
+  57.14,
+  67.86,
+  55.56,
+  38.42,
+  70.0,
+  77.78,
+  70.0,
+  34.42,
+  77.78,
+  77.78,
+  70.0,
+  62.5,
+  70.0,
+  77.78,
+  70.0,
+  77.78,
+  70.0,
+  77.78,
+  70.0
+ ],
+ "headline": [
+  {
+   "short": "TB 2.1",
+   "name": "Terminal-Bench 2.1",
+   "metric": "avg@5",
+   "model": "GPT-5.5",
+   "scope": "89 tasks",
+   "baseLabel": "Monet (base)",
+   "base": 75.5,
+   "evolvedLabel": "Monet (DarwinX)",
+   "evolved": 83.2,
+   "priorLabel": "Codex",
+   "prior": 83.1,
+   "delta": 7.7,
+   "lo": 70,
+   "hi": 88
+  },
+  {
+   "short": "T-World",
+   "name": "TerminalWorld",
+   "metric": "pass@1",
+   "model": "Opus 4.8",
+   "scope": "41 held-out tasks",
+   "baseLabel": "Monet (base)",
+   "base": 61.0,
+   "evolvedLabel": "Monet (DarwinX)",
+   "evolved": 68.3,
+   "priorLabel": "Claude Code",
+   "prior": 65.9,
+   "delta": 7.3,
+   "lo": 55,
+   "hi": 75
+  },
+  {
+   "short": "WA-Inf.",
+   "name": "WebArena-Infinity",
+   "metric": "audit-clean pass@1",
+   "model": "GPT-5.5",
+   "scope": "1,260 real tasks",
+   "baseLabel": "Monet (base)",
+   "base": 43.5,
+   "evolvedLabel": "Monet (DarwinX)",
+   "evolved": 93.0,
+   "priorLabel": "GPT-5.5 + Browser Use",
+   "prior": 86.1,
+   "delta": 49.5,
+   "lo": 35,
+   "hi": 100
+  },
+  {
+   "short": "SWE-V",
+   "name": "SWE-bench Verified",
+   "metric": "official pass@1",
+   "model": "Opus 4.8",
+   "scope": "500 instances, transfer only",
+   "baseLabel": "Fix-skill reference",
+   "base": 80.8,
+   "evolvedLabel": "Monet (DarwinX), transferred",
+   "evolved": 84.2,
+   "priorLabel": null,
+   "prior": null,
+   "delta": 3.4,
+   "lo": 75,
+   "hi": 88
+  }
+ ],
+ "twBars": {
+  "n": 41,
+  "claudeCode": 27,
+  "bars": [
+   {
+    "label": "Base",
+    "solved": 25
+   },
+   {
+    "label": "Spec. A",
+    "solved": 24
+   },
+   {
+    "label": "Spec. B",
+    "solved": 25
+   },
+   {
+    "label": "Spec. C",
+    "solved": 26
+   },
+   {
+    "label": "Spec. D",
+    "solved": 27
+   },
+   {
+    "label": "Monet (DarwinX)",
+    "solved": 28
+   }
+  ]
+ },
+ "tb21Compute": {
+  "groups": [
+   "Newly solved (6 tasks)",
+   "Already solved (69 tasks)"
+  ],
+  "turns": {
+   "base": [
+    11,
+    12
+   ],
+   "evolved": [
+    22,
+    13
+   ]
+  },
+  "tokens": {
+   "base": [
+    89,
+    125
+   ],
+   "evolved": [
+    380,
+    172
+   ]
+  }
+ },
+ "waiInvalid": {
+  "application": {
+   "labels": [
+    "Elation Clinical",
+    "Elation Prescription",
+    "GitLab Plan & Track",
+    "Gmail",
+    "Gmail Acct & Contact",
+    "Handshake",
+    "Linear Acct Mgmt",
+    "PayPal Wallet",
+    "Super-human",
+    "Xero Invoicing"
+   ],
+   "before": [
+    29,
+    27,
+    38,
+    21,
+    44,
+    54,
+    26,
+    10,
+    44,
+    0
+   ],
+   "after": [
+    2,
+    3,
+    0,
+    0,
+    3,
+    0,
+    0,
+    0,
+    8,
+    1
+   ],
+   "colors": [
+    "#4C72B0",
+    "#A6C8E0",
+    "#DD8452",
+    "#E3B23C",
+    "#55A868",
+    "#8FD19E",
+    "#C44E52",
+    "#E79EA6",
+    "#8172B3",
+    "#C7BFE0"
+   ]
+  },
+  "mechanism": {
+   "labels": [
+    "Evaluation plane access",
+    "Privileged knowledge",
+    "Exploit or privilege escalation",
+    "Raw state mutation"
+   ],
+   "before": [
+    155,
+    97,
+    26,
+    15
+   ],
+   "after": [
+    0,
+    0,
+    0,
+    17
+   ],
+   "colors": [
+    "#55A868",
+    "#DD8452",
+    "#CE6DBD",
+    "#8FA6D1"
+   ]
+  }
+ },
+ "waiAudit": {
+  "apps": [
+   "Overall",
+   "Elation-Clin",
+   "Elation-Rx",
+   "GitLab",
+   "Gmail",
+   "Gmail-Acct",
+   "Handshake",
+   "Linear",
+   "PayPal",
+   "Superhuman",
+   "Xero"
+  ],
+  "rawBase": [
+   53.0,
+   95.8,
+   28.3,
+   75.7,
+   45.0,
+   33.3,
+   44.0,
+   54.2,
+   50.0,
+   53.3,
+   49.2
+  ],
+  "rawDx": [
+   94.4,
+   98.3,
+   97.5,
+   97.9,
+   98.3,
+   94.2,
+   84.0,
+   94.2,
+   95.7,
+   94.2,
+   97.5
+  ],
+  "auditedBase": [
+   43.5,
+   95.8,
+   20.0,
+   63.6,
+   25.0,
+   21.7,
+   36.5,
+   43.3,
+   49.3,
+   31.7,
+   39.2
+  ],
+  "auditedDx": [
+   93.0,
+   96.7,
+   95.0,
+   97.9,
+   98.3,
+   91.7,
+   84.0,
+   94.2,
+   95.7,
+   87.5,
+   96.7
+  ]
+ }
+};

index.html

@@ -1,19 +1,1706 @@
 <!doctype html>
-<html>
-	<head>
-		<meta charset="utf-8" />
-		<meta name="viewport" content="width=device-width" />
-		<title>My static Space</title>
-		<link rel="stylesheet" href="style.css" />
-	</head>
-	<body>
-		<div class="card">
-			<h1>Welcome to your static Space!</h1>
-			<p>You can modify this app directly by editing <i>index.html</i> in the Files and versions tab.</p>
-			<p>
-				Also don't forget to check the
-				<a href="https://huggingface.co/docs/hub/spaces" target="_blank">Spaces documentation</a>.
-			</p>
-		</div>
-	</body>
+<html lang="en">
+<head>
+<meta charset="utf-8" />
+<meta name="viewport" content="width=device-width, initial-scale=1" />
+<title>DarwinX: Evolving Agent Harnesses Through Natural Selection</title>
+<meta name="description" content="DarwinX treats agent self-evolution as selection over a population of harnesses with the model frozen. 84.7% on Terminal-Bench 2.1, 68.3% on held-out TerminalWorld, 93.0% audit-clean on WebArena-Infinity, 84.2% zero-shot transfer to SWE-bench Verified." />
+<meta property="og:type" content="article" />
+<meta property="og:title" content="DarwinX: Evolving Agent Harnesses Through Natural Selection" />
+<meta property="og:description" content="Population selection over agent harnesses with the model frozen. Terminal-Bench 2.1 frontier at 84.7%." />
+<meta property="og:image" content="assets/teaser.png" />
+<meta name="twitter:card" content="summary_large_image" />
+<meta name="twitter:title" content="DarwinX: Evolving Agent Harnesses Through Natural Selection" />
+<meta name="twitter:description" content="A frozen model need not be a fixed agent: harness selection turns evaluation compute into durable capability." />
+<meta name="twitter:image" content="assets/teaser.png" />
+<link rel="icon" href="data:image/svg+xml,<svg xmlns='http://www.w3.org/2000/svg' viewBox='0 0 32 32'><rect width='32' height='32' rx='7' fill='%23032d60'/><path d='M12.6 14.8 19.4 10.6M12.6 17.2 19.4 21.4' stroke='%23fff' stroke-width='1.7'/><circle cx='10' cy='16' r='3.3' fill='%230176d3'/><circle cx='22' cy='9' r='3.3' fill='%23fff'/><circle cx='22' cy='23' r='3.3' fill='%23fff'/></svg>" />
+<style>
+  :root {
+    --navy: #032d60;
+    --blue: #0176d3;
+    --ink: #17202c;
+    --muted: #5a6b7d;
+    --line: #dfe6ee;
+    --tint: #f2f7fc;
+    --bg: #ffffff;
+    --maxw: 920px;
+    --navh: 48px;
+  }
+  * { box-sizing: border-box; }
+  html { -webkit-text-size-adjust: 100%; scroll-behavior: smooth; }
+  body {
+    margin: 0;
+    background: var(--bg);
+    color: var(--ink);
+    font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", "Helvetica Neue", Arial, sans-serif;
+    font-size: 17px;
+    line-height: 1.65;
+  }
+  .wrap { max-width: var(--maxw); margin: 0 auto; padding: 0 22px; }
+  a { color: var(--blue); text-decoration: none; }
+  a:hover { text-decoration: underline; }
+
+  /* ---------- sticky nav ---------- */
+  nav.top {
+    position: sticky; top: 0; z-index: 50;
+    background: rgba(255,255,255,0.94);
+    backdrop-filter: saturate(180%) blur(8px);
+    border-bottom: 1px solid var(--line);
+    height: var(--navh);
+  }
+  nav.top .wrap { display: flex; align-items: center; gap: 4px; height: 100%; overflow-x: auto; }
+  nav.top .brand { font-weight: 700; color: var(--navy); margin-right: 12px; white-space: nowrap; }
+  nav.top a {
+    color: var(--muted); font-size: 0.85rem; padding: 5px 9px; border-radius: 5px; white-space: nowrap;
+  }
+  nav.top a:hover { background: var(--tint); color: var(--navy); text-decoration: none; }
+  nav.top .spacer { flex: 1 1 auto; }
+  nav.top a.cta { color: #fff; background: var(--navy); font-weight: 500; }
+  nav.top a.cta:hover { background: #06407f; color: #fff; }
+  section[id] { scroll-margin-top: calc(var(--navh) + 14px); }
+
+  /* ---------- header ---------- */
+  header { padding: 52px 0 34px; border-bottom: 1px solid var(--line); }
+  h1 {
+    font-size: 2.05rem; line-height: 1.25; letter-spacing: -0.015em;
+    color: var(--navy); margin: 0 0 20px; font-weight: 700;
+  }
+  h1 .sub { display: block; font-size: 1.32rem; font-weight: 600; margin-top: 6px; color: var(--ink); }
+  .authors { font-size: 1.02rem; margin: 0 0 8px; }
+  .authors b { font-weight: 600; }
+  sup { font-size: 0.7em; color: var(--blue); }
+  .affil, .legend { color: var(--muted); font-size: 0.9rem; margin: 4px 0 0; }
+  .legend { margin-top: 12px; font-size: 0.84rem; }
+
+  .links { display: flex; flex-wrap: wrap; gap: 10px; margin: 26px 0 0; }
+  .btn {
+    display: inline-flex; align-items: center; padding: 9px 17px; border-radius: 7px;
+    background: var(--navy); color: #fff; font-size: 0.94rem; font-weight: 500;
+    border: 1px solid var(--navy); cursor: pointer;
+  }
+  .btn:hover { background: #06407f; text-decoration: none; }
+  .btn.ghost { background: #fff; color: var(--navy); }
+  .btn.ghost:hover { background: var(--tint); }
+
+  /* ---------- sections ---------- */
+  section { padding: 42px 0; border-bottom: 1px solid var(--line); }
+  section:last-of-type { border-bottom: 0; }
+  h2 {
+    font-size: 0.83rem; text-transform: uppercase; letter-spacing: 0.1em;
+    color: var(--muted); font-weight: 700; margin: 0 0 20px;
+  }
+  h3 { font-size: 1.16rem; color: var(--navy); margin: 0 0 4px; font-weight: 600; letter-spacing: -0.01em; }
+  h3 .tag { font-weight: 400; color: var(--muted); font-size: 0.8rem; letter-spacing: 0.02em; }
+  h4 { font-size: 1rem; color: var(--ink); margin: 0 0 5px; font-weight: 600; }
+  p { margin: 0 0 15px; }
+  p:last-child { margin-bottom: 0; }
+
+  figure { margin: 0; }
+  figure img { width: 100%; height: auto; display: block; border: 1px solid var(--line); border-radius: 8px; background: #fff; }
+  figure a.zoom { display: block; position: relative; }
+  figure a.zoom:hover img { border-color: var(--blue); }
+  figure a.zoom::after {
+    content: "\2317"; position: absolute; right: 9px; bottom: 9px;
+    font-size: 0.85rem; line-height: 1; padding: 5px 7px; border-radius: 5px;
+    background: rgba(3,45,96,0.82); color: #fff; opacity: 0; transition: opacity .15s;
+  }
+  figure a.zoom:hover::after { opacity: 1; }
+  figcaption { color: var(--muted); font-size: 0.87rem; margin-top: 10px; line-height: 1.55; }
+  .figrow { display: grid; grid-template-columns: 1fr 1fr; gap: 20px; align-items: start; }
+  .abstract { font-size: 1.02rem; }
+
+  /* ---------- headline cards ---------- */
+  .cards { display: grid; grid-template-columns: repeat(2, 1fr); gap: 14px; }
+  .card { border: 1px solid var(--line); border-radius: 8px; padding: 17px 18px; background: var(--tint); }
+  .card .label { font-size: 0.79rem; color: var(--muted); text-transform: uppercase; letter-spacing: 0.06em; font-weight: 600; }
+  .card .big { font-size: 1.75rem; font-weight: 700; color: var(--navy); margin: 7px 0 2px; letter-spacing: -0.02em; }
+  .card .big .delta { font-size: 0.9rem; font-weight: 600; color: var(--blue); margin-left: 7px; letter-spacing: 0; }
+  .card .note { font-size: 0.85rem; color: var(--muted); line-height: 1.5; }
+
+  /* ---------- per-benchmark blocks ---------- */
+  .bench { padding-top: 34px; margin-top: 34px; border-top: 1px solid var(--line); }
+  .bench:first-of-type { padding-top: 0; margin-top: 0; border-top: 0; }
+  .bench > .lede { color: var(--muted); font-size: 0.93rem; margin: 0 0 20px; }
+  .block { margin-top: 28px; }
+
+  table { width: 100%; border-collapse: collapse; font-size: 0.94rem; }
+  caption { text-align: left; color: var(--muted); font-size: 0.87rem; padding-bottom: 10px; line-height: 1.5; }
+  th, td { text-align: left; padding: 8px 9px; border-bottom: 1px solid var(--line); vertical-align: top; }
+  th { font-size: 0.78rem; text-transform: uppercase; letter-spacing: 0.04em; color: var(--muted); border-bottom: 1.5px solid #c9d4e0; font-weight: 700; }
+  td.num, th.num { text-align: right; font-variant-numeric: tabular-nums; }
+  tr.ours td { background: var(--tint); font-weight: 600; color: var(--navy); }
+  tr.base td { color: var(--muted); font-style: italic; }
+  tr.total td { border-top: 1.5px solid #c9d4e0; font-weight: 700; }
+  td.delta { color: var(--blue); font-weight: 600; }
+  .tnote { color: var(--muted); font-size: 0.85rem; margin-top: 11px; line-height: 1.55; }
+  .scroller { overflow-x: auto; }
+  .scroller table { min-width: 640px; }
+  code {
+    font-family: ui-monospace, SFMono-Regular, "SF Mono", Menlo, monospace;
+    font-size: 0.85em; background: #eef3f8; padding: 1px 5px; border-radius: 4px; color: var(--navy);
+    white-space: nowrap;
+  }
+
+  ul.mech { list-style: none; padding: 0; margin: 0; }
+  ul.mech li { padding: 0 0 18px 0; }
+  ul.mech li:last-child { padding-bottom: 0; }
+  ul.lim { padding-left: 20px; margin: 0; }
+  ul.lim li { padding-bottom: 12px; }
+  ul.lim li:last-child { padding-bottom: 0; }
+  ul.lim b { color: var(--navy); }
+
+  /* ---------- interactive widgets ---------- */
+  .widget { border: 1px solid var(--line); border-radius: 9px; padding: 18px; background: #fcfdfe; }
+  .widget .wtop {
+    display: flex; flex-wrap: wrap; align-items: center; gap: 8px 14px;
+    padding-bottom: 15px; margin-bottom: 15px; border-bottom: 1px solid var(--line);
+  }
+  .widget .wtop .wlabel { font-size: 0.78rem; text-transform: uppercase; letter-spacing: 0.06em; color: var(--muted); font-weight: 700; }
+  .chips { display: flex; flex-wrap: wrap; gap: 7px; }
+  .chip {
+    font: inherit; font-size: 0.85rem; padding: 5px 12px; border-radius: 20px;
+    border: 1px solid var(--line); background: #fff; color: var(--muted); cursor: pointer;
+    transition: background .12s, border-color .12s, color .12s;
+  }
+  .chip:hover { border-color: #b9c7d6; }
+  .chip[aria-pressed="true"] { background: var(--navy); border-color: var(--navy); color: #fff; }
+  .chip .cnt { opacity: 0.62; font-variant-numeric: tabular-nums; }
+  .toggle { display: inline-flex; align-items: center; gap: 7px; font-size: 0.86rem; color: var(--muted); cursor: pointer; }
+  .toggle input { accent-color: var(--navy); width: 15px; height: 15px; }
+
+  .readout { display: flex; flex-wrap: wrap; gap: 10px; margin-bottom: 16px; }
+  .ro {
+    flex: 1 1 150px; border: 1px solid var(--line); border-radius: 7px;
+    padding: 11px 13px; background: #fff;
+  }
+  .ro .k { font-size: 0.75rem; text-transform: uppercase; letter-spacing: 0.05em; color: var(--muted); font-weight: 600; }
+  .ro .v { font-size: 1.35rem; font-weight: 700; color: var(--navy); font-variant-numeric: tabular-nums; letter-spacing: -0.01em; }
+  .ro.hi { background: var(--tint); border-color: #c9dcf0; }
+  .ro .v small { font-size: 0.62em; font-weight: 600; color: var(--muted); }
+
+  .tgrid { display: grid; grid-template-columns: repeat(auto-fill, minmax(26px, 1fr)); gap: 5px; }
+  .tcell {
+    position: relative; aspect-ratio: 1; border-radius: 4px; background: #e6ecf2;
+    border: 1px solid transparent; cursor: default; transition: background .12s;
+  }
+  .tcell.on { background: var(--blue); }
+  .tcell.gain { box-shadow: 0 0 0 2px #e8a33d; z-index: 1; }
+  .tcell.loss { box-shadow: 0 0 0 2px #d05a4e; z-index: 1; }
+  .tcell:hover { border-color: var(--navy); }
+
+  .wlegend { display: flex; flex-wrap: wrap; gap: 14px; margin-top: 14px; font-size: 0.8rem; color: var(--muted); }
+  .wlegend span { display: inline-flex; align-items: center; gap: 6px; }
+  .sw { width: 13px; height: 13px; border-radius: 3px; display: inline-block; }
+
+  #tip {
+    position: fixed; z-index: 100; pointer-events: none; opacity: 0;
+    background: var(--navy); color: #fff; font-size: 0.78rem; line-height: 1.45;
+    padding: 7px 10px; border-radius: 6px; max-width: 260px;
+    transition: opacity .1s; box-shadow: 0 4px 14px rgba(3,45,96,0.25);
+  }
+  #tip b { color: #9ecbf5; font-weight: 600; }
+
+  /* cluster bars */
+  .cbars { display: flex; flex-direction: column; gap: 9px; }
+  .cbar { display: grid; grid-template-columns: minmax(120px, 1.5fr) 3fr auto; gap: 12px; align-items: center; }
+  .cbar .cname { font-size: 0.86rem; color: var(--ink); }
+  .cbar .cname .cn { color: var(--muted); font-size: 0.78rem; }
+  .cbar .track { position: relative; height: 26px; background: #f0f4f8; border-radius: 4px; overflow: hidden; }
+  .cbar .fill { position: absolute; top: 0; height: 100%; border-radius: 4px; transition: width .35s ease, left .35s ease; }
+  .cbar .fill.b { background: #c6d3e0; }
+  .cbar .fill.e { background: var(--blue); opacity: 0.92; }
+  .cbar .cd { font-size: 0.84rem; font-weight: 700; font-variant-numeric: tabular-nums; min-width: 52px; text-align: right; }
+  .cbar .cd.up { color: var(--blue); }
+  .cbar .cd.down { color: #b3453a; }
+
+  /* line chart */
+  .chart { width: 100%; height: auto; display: block; overflow: visible; }
+  .chart .ax { stroke: #ccd7e2; stroke-width: 1; }
+  .chart .gl { stroke: #eef2f6; stroke-width: 1; }
+  .chart .lbl { font-size: 10.5px; fill: var(--muted); }
+  .chart .best { fill: none; stroke: var(--blue); stroke-width: 2.2; }
+  .chart .dot { fill: #b9c7d6; }
+  .chart .dot:hover, .chart .dot.act { fill: var(--navy); }
+  .chart .hit { fill: transparent; cursor: pointer; }
+
+  /* headline panels */
+  .panels { display: grid; grid-template-columns: repeat(4, 1fr); gap: 12px; }
+  .panel { text-align: center; }
+  .panel .pt { font-size: 0.95rem; font-weight: 700; color: var(--navy); }
+  .panel .ps { font-size: 0.74rem; color: var(--muted); line-height: 1.4; margin-bottom: 6px; }
+  .panel svg { width: 100%; height: auto; display: block; overflow: visible; }
+  .panel .bar { transition: y .35s ease, height .35s ease; }
+  .panel .bar.b { fill: #c6d3e0; }
+  .panel .bar.p { fill: #8fb4e8; }
+  .panel .bar.e { fill: var(--blue); }
+  .panel .dl { font-size: 11.5px; font-weight: 700; fill: var(--blue); }
+  .panel .tick { font-size: 9.5px; fill: var(--muted); }
+  .panel .axl { stroke: #ccd7e2; stroke-width: 1; }
+  .panel .foot { font-size: 0.72rem; color: var(--muted); margin-top: 4px; }
+
+  /* horizontal bar rows (specialists, compute) */
+  .hrows { display: flex; flex-direction: column; gap: 8px; }
+  .hrow { display: grid; grid-template-columns: minmax(96px, 1.1fr) 3fr auto; gap: 12px; align-items: center; }
+  .hrow .hn { font-size: 0.86rem; }
+  .hrow .ht { position: relative; height: 24px; background: #f0f4f8; border-radius: 4px; }
+  .hrow .hf { position: absolute; left: 0; top: 0; height: 100%; border-radius: 4px; transition: width .35s ease; }
+  .hrow .hv { font-size: 0.84rem; font-weight: 700; font-variant-numeric: tabular-nums; min-width: 68px; text-align: right; color: var(--navy); }
+  .hrow.dim .hn, .hrow.dim .hv { color: var(--muted); }
+  .refline { position: absolute; top: -3px; bottom: -3px; width: 0; border-left: 2px dashed #e4572e; }
+  .reftag { position: absolute; top: -17px; transform: translateX(-50%); font-size: 0.7rem; color: #e4572e; white-space: nowrap; }
+
+  /* stacked composition */
+  .stack { display: flex; height: 30px; border-radius: 5px; overflow: hidden; background: #f0f4f8; }
+  .stack i { display: block; height: 100%; transition: width .35s ease; }
+  .srow { margin-bottom: 16px; }
+  .srow .sh { display: flex; justify-content: space-between; font-size: 0.82rem; color: var(--muted); margin-bottom: 5px; }
+  .srow .sh b { color: var(--navy); font-size: 0.95rem; }
+  .keys { display: flex; flex-wrap: wrap; gap: 6px 14px; margin-top: 12px; font-size: 0.78rem; color: var(--muted); }
+  .keys span { display: inline-flex; align-items: center; gap: 6px; }
+
+  /* dumbbell */
+  .dumb { display: flex; flex-direction: column; gap: 7px; }
+  .drow { display: grid; grid-template-columns: minmax(84px, 1fr) 4fr; gap: 12px; align-items: center; }
+  .drow.tot { padding-bottom: 7px; border-bottom: 1px solid var(--line); }
+  .drow.tot .dn { font-weight: 700; color: var(--navy); }
+  .drow .dn { font-size: 0.82rem; }
+  .drow .dt { position: relative; height: 30px;
+    background-image: repeating-linear-gradient(to right, #edf1f5 0 1px, transparent 1px 25%); }
+  .drow.dax { align-items: end; }
+  .drow.dax .dt { height: 16px; background: none; }
+  .drow.dax .dt span { position: absolute; bottom: 0; transform: translateX(-50%);
+    font-size: 0.7rem; color: var(--muted); }
+  .drow .seg { position: absolute; height: 3px; border-radius: 2px; }
+  .drow .pt2 { position: absolute; width: 9px; height: 9px; border-radius: 50%; transform: translate(-50%, -50%); }
+  .drow .hollow { background: #fff; }
+
+  .nojs { display: none; }
+  .nojs img { width: 100%; height: auto; border: 1px solid var(--line); border-radius: 8px; }
+
+  th.sortable { cursor: pointer; user-select: none; white-space: nowrap; }
+  th.sortable:hover { color: var(--navy); }
+  th.sortable::after { content: " \2195"; opacity: 0.35; font-size: 0.9em; }
+  th.sortable.asc::after { content: " \2191"; opacity: 1; color: var(--blue); }
+  th.sortable.desc::after { content: " \2193"; opacity: 1; color: var(--blue); }
+
+  .bibhead { display: flex; align-items: center; justify-content: space-between; gap: 12px; margin-bottom: 12px; }
+  .bibhead h2 { margin: 0; }
+  #copybtn {
+    font: inherit; font-size: 0.84rem; padding: 6px 13px; border-radius: 6px;
+    border: 1px solid var(--line); background: #fff; color: var(--navy); cursor: pointer;
+  }
+  #copybtn:hover { background: var(--tint); border-color: #c9d4e0; }
+  pre {
+    background: #f7f9fb; border: 1px solid var(--line); border-radius: 8px;
+    padding: 16px; overflow-x: auto; font-size: 0.85rem; line-height: 1.55;
+    font-family: ui-monospace, SFMono-Regular, "SF Mono", Menlo, monospace;
+    color: #223; margin: 0;
+  }
+  footer { padding: 30px 0 60px; color: var(--muted); font-size: 0.88rem; }
+
+  @media (max-width: 700px) {
+    body { font-size: 16px; }
+    h1 { font-size: 1.6rem; }
+    header { padding: 32px 0 26px; }
+    .cards, .figrow { grid-template-columns: 1fr; }
+    nav.top .hideable { display: none; }
+    .panels { grid-template-columns: repeat(2, 1fr); gap: 18px 10px; }
+    .hrow { grid-template-columns: minmax(72px, 1fr) 2fr auto; gap: 8px; }
+    .hrow .hn, .drow .dn { font-size: 0.78rem; }
+    .hrow .hv { min-width: 54px; font-size: 0.78rem; }
+    .drow { grid-template-columns: minmax(66px, 1fr) 3fr; gap: 8px; }
+  }
+</style>
+</head>
+<body>
+<noscript><style>.widget{display:none}.nojs{display:block}</style></noscript>
+
+<nav class="top">
+  <div class="wrap">
+    <span class="brand">DarwinX</span>
+    <a href="#abstract">Abstract</a>
+    <a href="#method">Method</a>
+    <a href="#results">Results</a>
+    <a href="#benchmarks" class="hideable">Benchmarks</a>
+    <a href="#ablation" class="hideable">Ablation</a>
+    <a href="#limitations" class="hideable">Limitations</a>
+    <a href="#bibtex" class="hideable">BibTeX</a>
+    <span class="spacer"></span>
+    <a class="cta" href="https://arxiv.org/abs/2608.07545">arXiv</a>
+  </div>
+</nav>
+
+<header>
+  <div class="wrap">
+    <h1>DarwinX<span class="sub">Evolving Agent Harnesses Through Natural Selection</span></h1>
+
+    <p class="authors">
+      <b>Yifan Zhang</b><sup>&#9702;1</sup>,
+      <b>Yutong Dai</b><sup>&#9702;1</sup>,
+      Juntao Tan<sup>*1</sup>,
+      Luyu Yang<sup>*1</sup>,
+      Rishi Mullur<sup>1</sup>,
+      Thai Hoang<sup>1</sup>,
+      Zhiyuan Hu<sup>1</sup>,
+      James Zhu<sup>&dagger;2</sup>,
+      Phil Mui<sup>&dagger;2</sup>,
+      Silvio Savarese<sup>&dagger;1</sup>,
+      Ran Xu<sup>&dagger;1</sup>,
+      Zeyuan Chen<sup>&dagger;1</sup>
+    </p>
+    <p class="affil"><sup>1</sup>Salesforce AI Research &nbsp;&nbsp; <sup>2</sup>Salesforce Agentforce</p>
+    <p class="legend"><sup>&#9702;</sup>First authors &nbsp; <sup>*</sup>Core authors &nbsp; <sup>&dagger;</sup>Senior authors</p>
+
+    <div class="links">
+      <a class="btn" href="https://arxiv.org/abs/2608.07545">arXiv:2608.07545</a>
+      <a class="btn ghost" href="https://arxiv.org/pdf/2608.07545">PDF</a>
+      <a class="btn ghost" href="https://huggingface.co/papers/2608.07545">Hugging Face</a>
+      <a class="btn ghost" href="mailto:yifan.zhang@salesforce.com?subject=DarwinX%20code%20request">Code available on request</a>
+    </div>
+  </div>
+</header>
+
+<section>
+  <div class="wrap">
+    <figure>
+      <a class="zoom" href="assets/full/teaser.png" target="_blank" rel="noopener">
+        <img src="assets/teaser_schematic.png" alt="Variants are evolved from a base agent, survivors are kept and merged, pruned lineages are dropped, all with the model frozen." />
+      </a>
+      <figcaption>
+        <b>With the base model frozen, evolving the harness alone matches or beats the strongest prior
+        agent on four benchmarks.</b> Variants survive on measured fitness (avg@<i>k</i>, no gold solutions)
+        and complementary survivors are merged. Click for the paper's full figure.
+      </figcaption>
+    </figure>
+
+    <div class="block">
+      <div class="widget" id="hl-widget">
+        <div class="wtop">
+          <span class="wlabel">y-axis</span>
+          <div class="chips" id="hl-chips"></div>
+          <span style="flex:1"></span>
+          <label class="toggle"><input type="checkbox" id="hl-prior" checked /> best prior agent</label>
+        </div>
+        <div class="panels" id="hl-panels"></div>
+        <div class="wlegend" id="hl-legend"></div>
+        <p class="tnote" id="hl-note"></p>
+      </div>
+      <noscript class="nojs"><img src="assets/teaser.png" alt="Four-benchmark summary: base vs. evolved vs. best prior agent." /></noscript>
+    </div>
+  </div>
+</section>
+
+<section id="abstract">
+  <div class="wrap">
+    <h2>Abstract</h2>
+    <div class="abstract">
+      <p>
+        An LLM agent's capability depends not only on model weights but on its <i>harness</i>: prompts,
+        tools, skills, and control flow. Self-improvement loops already edit harnesses, yet single-lineage
+        search is path-dependent and local wins often regress other tasks. We introduce <b>DarwinX</b>, which
+        treats self-evolution as <b>selection over a population of harnesses</b> with the model frozen: a
+        preserve-and-extend contract admits only variants that extend coverage without regressing, an archive
+        keeps alternative lineages for recombination, and failure-, teacher-, and self-derived evidence share
+        one edit interface. Fitness comes from each benchmark's own verifier: no gold solutions, no
+        hand-picked winners.
+      </p>
+      <p>
+        Across four benchmarks that progressively separate the evolution signal from the test, one loop adds
+        about 17 points on average: Terminal-Bench&nbsp;2.1 rises +7.7 to 83.2% on a matched base and to the
+        verified frontier at <b>84.7%</b> on a stronger one; TerminalWorld's held-out split reaches
+        <b>68.3%</b>, ahead of every off-the-shelf agent; WebArena-Infinity real-task pass@1 rises from 43.5%
+        to <b>93.0%</b> audit-clean; and a Terminal-Bench&nbsp;2.1 harness transfers unchanged to SWE-bench
+        Verified. What evolves is general agent competence, not benchmark-specific patches, so it survives
+        changes of task, verifier, and base model. A frozen model need not be a fixed agent: harness selection
+        turns evaluation compute into durable capability.
+      </p>
+    </div>
+  </div>
+</section>
+
+<section id="method">
+  <div class="wrap">
+    <h2>Method</h2>
+
+    <figure>
+      <a class="zoom" href="assets/full/overview.png" target="_blank" rel="noopener">
+        <img src="assets/overview.png" alt="Overview of the DarwinX loop." loading="lazy" />
+      </a>
+    </figure>
+
+    <ul class="mech" style="margin-top:26px">
+      <li>
+        <h4>Selection over a population, not one lineage</h4>
+        <p>A run maintains an archive shaped as a tree, where each node is a harness snapshot with its edit
+          delta, per-task scores, trial evidence, and distilled lessons. Nothing is discarded: a variant that
+          loses overall may still hold the one edit that, combined with another branch's, unlocks a task
+          neither solves alone. Parallel branches target different capability clusters, so the archive grows
+          specialists with different solved-task signatures.</p>
+      </li>
+      <li>
+        <h4>A preserve-and-extend contract</h4>
+        <p>A child is admitted only when the measured evidence shows it solved something new
+          <i>without breaking what it already solved</i>, which bounds what a win is allowed to cost
+          elsewhere. When variants solve complementary tasks, their additive edits are merged, and the merged
+          child is kept only if it covers the union of its parents' wins. Fitness is the benchmark's own
+          verifier at avg@<i>k</i> &mdash; no gold solutions and no hand-picked winners.</p>
+      </li>
+      <li>
+        <h4>One edit interface for three signals</h4>
+        <p>Failure-derived diagnosis, teacher-derived demonstrations, and self-derived rollout contrast are
+          all converted into harness edits rather than model-weight updates. The base model never changes, so
+          every gain reported here is a statement about the harness alone.</p>
+      </li>
+    </ul>
+
+    <figure style="margin-top:30px">
+      <a class="zoom" href="assets/full/population.png" target="_blank" rel="noopener">
+        <img src="assets/population.png" alt="DarwinX per-generation operators: mutation loop, variant classification, and the merge operator." loading="lazy" />
+      </a>
+      <figcaption>
+        <b>DarwinX's per-generation operators.</b> <b>Left:</b> the mutation loop and the three learning
+        signals that drive it. <b>Middle:</b> variants classified by how their solved set changes, where those
+        preserving inherited solves stay eligible for recombination while the rest contribute only distilled
+        lessons. <b>Right:</b> the merge operator and its acceptance criterion.
+      </figcaption>
+    </figure>
+  </div>
+</section>
+
+<section id="results">
+  <div class="wrap">
+    <h2>Results at a glance</h2>
+    <div class="cards">
+      <div class="card">
+        <div class="label">Terminal-Bench 2.1 &middot; avg@5</div>
+        <div class="big">84.7%<span class="delta">+7.7 on matched base</span></div>
+        <div class="note">Verified-leaderboard frontier on a frozen GPT-5.6&nbsp;Sol at <i>medium</i> effort.
+          On GPT-5.5, 75.5% &rarr; 83.2%.</div>
+      </div>
+      <div class="card">
+        <div class="label">TerminalWorld &middot; held-out split</div>
+        <div class="big">68.3%<span class="delta">+7.3</span></div>
+        <div class="note">28/41 held-out tasks on a frozen Opus&nbsp;4.8, above every off-the-shelf agent
+          we measured.</div>
+      </div>
+      <div class="card">
+        <div class="label">WebArena-Infinity &middot; audit-clean pass@1</div>
+        <div class="big">93.0%<span class="delta">+49.5 pp</span></div>
+        <div class="note">1,260 real tasks, evolved on synthetic intents only. Base 43.5%; best same-model
+          baseline 86.1%.</div>
+      </div>
+      <div class="card">
+        <div class="label">SWE-bench Verified &middot; zero-shot transfer</div>
+        <div class="big">84.2%<span class="delta">+3.4</span></div>
+        <div class="note">421/500 official pass@1 from the Terminal-Bench&nbsp;2.1 harness run
+          <i>unchanged</i>, with no SWE-V feedback.</div>
+      </div>
+    </div>
+    <p class="tnote">
+      The four benchmarks progressively separate the evolution signal from the test: in-domain
+      (Terminal-Bench&nbsp;2.1), held-out task split (TerminalWorld), synthetic&nbsp;&rarr;&nbsp;real intent
+      shift (WebArena-Infinity), and zero-shot cross-benchmark transfer (SWE-bench Verified).
+    </p>
+  </div>
+</section>
+
+<section id="benchmarks">
+  <div class="wrap">
+    <h2>Benchmark detail</h2>
+
+    <!-- ============ TB 2.1 ============ -->
+    <div class="bench">
+      <h3>Terminal-Bench 2.1 <span class="tag">&mdash; in-domain evolution</span></h3>
+      <p class="lede">89 tasks. Evolution and evaluation share the task distribution; the model is frozen and
+        only the harness changes.</p>
+
+      <table class="sortable">
+        <caption>Terminal-Bench&nbsp;2.1, avg@5 under the strict leaderboard rule (binary avg@5,
+          <i>k</i>=5, errored trials score zero). Both DarwinX rows are leaderboard submissions, reported
+          before the leaderboard's uniform reward-hacking pass.</caption>
+        <thead>
+          <tr><th>Agent</th><th>Model / effort</th><th class="num">avg@5</th></tr>
+        </thead>
+        <tbody>
+          <tr class="ours"><td>Monet (DarwinX)</td><td>GPT-5.6 Sol / medium</td><td class="num">84.7 &plusmn; 1.2</td></tr>
+          <tr><td>Claude Code</td><td>Fable 5 / xhigh</td><td class="num">83.8 &plusmn; 1.2</td></tr>
+          <tr class="ours"><td>Monet (DarwinX)</td><td>GPT-5.5 / high</td><td class="num">83.2 &plusmn; 1.2</td></tr>
+          <tr><td>Codex</td><td>GPT-5.5 / xhigh</td><td class="num">83.1 &plusmn; 1.1</td></tr>
+          <tr><td>OpenAI reference</td><td>GPT-5.6 Sol / medium</td><td class="num">81.8</td></tr>
+          <tr><td>Terminus 2</td><td>GPT-5.5 / xhigh</td><td class="num">78.0 &plusmn; 1.2</td></tr>
+          <tr class="base"><td>Monet (base)</td><td>GPT-5.5 / default</td><td class="num">75.5 &plusmn; 3.5</td></tr>
+        </tbody>
+      </table>
+      <p class="tnote">
+        At 84.7% DarwinX matches or exceeds the verified leader (Claude Code + Fable&nbsp;5, 83.8% at
+        <i>xhigh</i>) while running at a <i>lower</i> effort setting, and adds +2.9 points over OpenAI's own
+        native single-agent GPT-5.6&nbsp;Sol at the same medium effort. Against a neutral harness on the
+        <i>same</i> GPT-5.5 base (Terminus&nbsp;2, 78.0%), the pure harness gain is +5.2 points &mdash; so the
+        improvement is the harness, not the model or the effort setting.
+      </p>
+
+      <div class="block">
+        <h4>Where evolution helps</h4>
+        <p class="tnote" style="margin-top:0;margin-bottom:14px">
+          Per-cluster avg@5 on a frozen GPT-5.5. Sort by any column and hover a bar for exact rates.
+        </p>
+        <div class="widget" id="cl-widget">
+          <div class="wtop">
+            <span class="wlabel">Sort by</span>
+            <div class="chips" id="cl-chips"></div>
+            <span class="spacer" style="flex:1"></span>
+            <label class="toggle"><input type="checkbox" id="cl-base" checked /> show base</label>
+          </div>
+          <div class="cbars" id="cl-bars"></div>
+          <div class="wlegend">
+            <span><i class="sw" style="background:#c6d3e0"></i> Monet (base)</span>
+            <span><i class="sw" style="background:var(--blue)"></i> Monet (DarwinX)</span>
+          </div>
+        </div>
+        <noscript class="nojs"><img src="assets/tb21_category_gain.png" alt="Per-cluster Terminal-Bench 2.1 gain." /></noscript>
+        <p class="tnote">
+          Gains concentrate where a frozen base has the most headroom: <i>ML &amp; scientific computing</i>
+          (+14.8, the largest cluster at 23 tasks) and <i>data &amp; databases</i> (+13.8). Clusters where the
+          base is already strong barely move, and none regresses beyond the per-task noise band. Paired
+          protocol over 88 tasks, which excludes one task with no paired evolved trial; the headline 83.2%
+          uses the stricter 89-task leaderboard rule above.
+        </p>
+      </div>
+
+      <div class="block">
+        <h4>The gain is the harness, not compute</h4>
+        <p class="tnote" style="margin-top:0;margin-bottom:14px">
+          Per-task compute, median over clean attempts.
+        </p>
+        <div class="widget" id="cp-widget">
+          <div class="wtop">
+            <span class="wlabel">Metric</span>
+            <div class="chips" id="cp-chips"></div>
+          </div>
+          <div class="hrows" id="cp-rows"></div>
+          <div class="wlegend">
+            <span><i class="sw" style="background:#c6d3e0"></i> Monet (base)</span>
+            <span><i class="sw" style="background:var(--blue)"></i> Monet (DarwinX)</span>
+          </div>
+          <p class="tnote" id="cp-note"></p>
+        </div>
+        <noscript class="nojs"><img src="assets/tb21_compute.png" alt="Per-task compute, newly solved vs. already solved." /></noscript>
+      </div>
+    </div>
+
+    <!-- ============ TerminalWorld ============ -->
+    <div class="bench">
+      <h3>TerminalWorld <span class="tag">&mdash; held-out task split</span></h3>
+      <p class="lede">Evolve on 94 verifier-scored training tasks, then freeze the harness and evaluate on 41
+        disjoint held-out tasks. Single-attempt pass@1, no retries and no best-of-<i>k</i>, so no held-out
+        reward can feed back into the archive.</p>
+
+      <table class="sortable">
+        <caption>TerminalWorld held-out pass@1 (41 tasks, single attempt, frozen base models, identical
+          verifier and infrastructure).</caption>
+        <thead>
+          <tr><th>Agent</th><th>Model</th><th class="num">pass@1</th></tr>
+        </thead>
+        <tbody>
+          <tr class="ours"><td>Monet (DarwinX)</td><td>Opus 4.8</td><td class="num">68.3%</td></tr>
+          <tr><td>Claude Code</td><td>Opus 4.8</td><td class="num">65.9%</td></tr>
+          <tr class="base"><td>Monet (base)</td><td>Opus 4.8</td><td class="num">61.0%</td></tr>
+          <tr><td>Terminus-2</td><td>GPT-5.5</td><td class="num">61.0%</td></tr>
+          <tr><td>Terminus-2</td><td>Opus 4.8</td><td class="num">58.5%</td></tr>
+          <tr><td>Monet (DarwinX)</td><td>GPT-5.5</td><td class="num">56.1%</td></tr>
+          <tr><td>Codex</td><td>GPT-5.5</td><td class="num">51.2%</td></tr>
+          <tr class="base"><td>Monet (base)</td><td>GPT-5.5</td><td class="num">48.8%</td></tr>
+        </tbody>
+      </table>
+
+      <div class="figrow block">
+        <div>
+          <div class="widget" id="sp-widget">
+            <div class="wtop">
+              <span class="wlabel">x-axis</span>
+              <div class="chips" id="sp-chips"></div>
+              <span style="flex:1"></span>
+              <label class="toggle"><input type="checkbox" id="sp-counts" /> task counts</label>
+            </div>
+            <div class="hrows" id="sp-rows"></div>
+            <p class="tnote" id="sp-note"></p>
+          </div>
+          <noscript class="nojs"><img src="assets/tw_heldout_specialists.png" alt="Held-out pass@1 of four specialists and their merge." /></noscript>
+          <p class="tnote">
+            Held-out generalization reflects <b>archive diversity</b>: four evolved specialists each solve a
+            different subset of the 41 tasks, and the merged Monet (DarwinX) exceeds every specialist and the
+            Claude Code reference (dashed).
+          </p>
+        </div>
+        <div>
+          <h4>The in-loop proxy overfits; the population absorbs it</h4>
+          <p style="font-size:0.95rem">
+            During evolution the training-subset score saturates from 0.505 to 1.000, yet held-out pass@1 is
+            68.3% &mdash; a <b>31.7-point gap</b> between the proxy the search maximizes and the held-out
+            truth it never sees. Crucially, the variant that best fits the proxy is <i>not</i> the best
+            generalizer: four high-scoring specialists solve 24, 25, 26, and 27 of the 41 held-out tasks on
+            overlapping but distinct subsets, and their merge reaches 28/41. Keeping a population, rather than
+            following the single best-scoring lineage, is what converts an overfit proxy into held-out gain.
+          </p>
+        </div>
+      </div>
+
+      <div class="block">
+        <h4>Explore the merge: which specialists cover which held-out tasks</h4>
+        <p class="tnote" style="margin-top:0;margin-bottom:14px">
+          Every cell is one of the 41 held-out tasks, filled when at least one selected specialist solves it.
+          Toggle specialists to watch coverage grow, and turn on the comparison to see how the harness DarwinX
+          actually produced differs from the plain union. Hover any cell for its task id.
+          Built from the per-task records of the five evaluation runs.
+        </p>
+        <div class="widget" id="tw-widget">
+          <div class="wtop">
+            <span class="wlabel">Specialists</span>
+            <div class="chips" id="tw-chips"></div>
+            <span class="spacer" style="flex:1"></span>
+            <label class="toggle"><input type="checkbox" id="tw-cmp" /> compare with realized merge</label>
+          </div>
+          <div class="readout" id="tw-readout"></div>
+          <div class="tgrid" id="tw-grid"></div>
+          <div class="wlegend" id="tw-legend">
+            <span><i class="sw" style="background:var(--blue)"></i> solved by selection</span>
+            <span><i class="sw" style="background:#e6ecf2"></i> unsolved</span>
+          </div>
+          <p class="tnote" id="tw-note"></p>
+        </div>
+      </div>
+    </div>
+
+    <!-- ============ WebArena-Infinity ============ -->
+    <div class="bench">
+      <h3>WebArena-Infinity <span class="tag">&mdash; synthetic &rarr; real intent shift</span></h3>
+      <p class="lede">Evolution sees synthetic intents only; evaluation is the official 1,260-task real suite
+        across 10 applications. We report <i>audit-clean</i> pass@1 throughout: trajectories flagged invalid
+        by the validity audit count as failures even when the verifier passes them.</p>
+
+      <div class="scroller">
+        <table class="sortable">
+          <caption>WebArena-Infinity audit-clean pass@1 by application. Click any column header to sort. External baselines are as-reported
+            without re-auditing, which makes this comparison conservative for DarwinX. BU = Browser Use.</caption>
+          <thead>
+            <tr>
+              <th>Application</th>
+              <th class="num">Kimi</th><th class="num">Qwen</th><th class="num">Gemini+BU</th>
+              <th class="num">GPT-5.5+BU</th><th class="num">Monet (base)</th>
+              <th class="num">Monet (DarwinX)</th><th class="num">&Delta;</th>
+            </tr>
+          </thead>
+          <tbody>
+            <tr><td>Elation clinical records</td><td class="num">50.0</td><td class="num">54.2</td><td class="num">81.7</td><td class="num">92.5</td><td class="num">95.8</td><td class="num"><b>96.7</b></td><td class="num delta">+0.9</td></tr>
+            <tr><td>Elation prescriptions</td><td class="num">23.3</td><td class="num">41.7</td><td class="num">80.8</td><td class="num">90.8</td><td class="num">20.0</td><td class="num"><b>95.0</b></td><td class="num delta">+75.0</td></tr>
+            <tr><td>GitLab plan and track</td><td class="num">39.3</td><td class="num">37.1</td><td class="num">63.6</td><td class="num">77.9</td><td class="num">63.6</td><td class="num"><b>97.9</b></td><td class="num delta">+34.3</td></tr>
+            <tr><td>Gmail</td><td class="num">70.0</td><td class="num">56.7</td><td class="num">75.0</td><td class="num">85.0</td><td class="num">25.0</td><td class="num"><b>98.3</b></td><td class="num delta">+73.3</td></tr>
+            <tr><td>Gmail accounts and contacts</td><td class="num">40.0</td><td class="num">33.3</td><td class="num">61.7</td><td class="num">87.5</td><td class="num">21.7</td><td class="num"><b>91.7</b></td><td class="num delta">+70.0</td></tr>
+            <tr><td>Handshake career exploration</td><td class="num">50.0</td><td class="num">50.5</td><td class="num">50.5</td><td class="num">83.5</td><td class="num">36.5</td><td class="num"><b>84.0</b></td><td class="num delta">+47.5</td></tr>
+            <tr><td>Linear account settings</td><td class="num">54.2</td><td class="num">65.8</td><td class="num">73.3</td><td class="num">81.7</td><td class="num">43.3</td><td class="num"><b>94.2</b></td><td class="num delta">+50.9</td></tr>
+            <tr><td>PayPal wallet</td><td class="num">70.7</td><td class="num">71.4</td><td class="num">88.6</td><td class="num">90.0</td><td class="num">49.3</td><td class="num"><b>95.7</b></td><td class="num delta">+46.4</td></tr>
+            <tr><td>Superhuman general</td><td class="num">15.0</td><td class="num">25.8</td><td class="num">50.0</td><td class="num">80.8</td><td class="num">31.7</td><td class="num"><b>87.5</b></td><td class="num delta">+55.8</td></tr>
+            <tr><td>Xero invoicing</td><td class="num">52.5</td><td class="num">55.8</td><td class="num">80.8</td><td class="num">93.3</td><td class="num">39.2</td><td class="num"><b>96.7</b></td><td class="num delta">+57.5</td></tr>
+            <tr class="total"><td>Overall</td><td class="num">43.3</td><td class="num">48.3</td><td class="num">69.3</td><td class="num">86.1</td><td class="num">43.5</td><td class="num">93.0</td><td class="num delta">+49.5</td></tr>
+          </tbody>
+        </table>
+      </div>
+      <p class="tnote">
+        93.0% audit-clean beats the strongest same-model baseline (GPT-5.5 + Browser&nbsp;Use, 86.1%) by
+        6.9&nbsp;points and the top public agent (Gemini&nbsp;3&nbsp;Flash + Browser&nbsp;Use, 69.3%) by
+        23.7&nbsp;points. Given differing action spaces across public agents, we treat the same-model gap as
+        the primary comparison. In counts, the base solves 548/1,260 and the evolved harness 1,171/1,260.
+      </p>
+
+      <div class="block">
+        <h4>Evolution as optimization</h4>
+        <p class="tnote" style="margin-top:0;margin-bottom:14px">
+          Screening score of every evaluated variant in the WebArena-Infinity run, in archive order. Hover a
+          point for its score; the line is the best-so-far envelope.
+        </p>
+        <div class="widget" id="cv-widget">
+          <div class="wtop">
+            <span class="wlabel">Show</span>
+            <label class="toggle"><input type="checkbox" id="cv-all" checked /> every evaluated variant</label>
+            <label class="toggle"><input type="checkbox" id="cv-best" checked /> best so far</label>
+          </div>
+          <svg class="chart" id="cv-chart" viewBox="0 0 860 320" role="img"
+               aria-label="Screening score per evaluated variant, with the best-so-far envelope."></svg>
+          <p class="tnote" id="cv-note"></p>
+        </div>
+        <noscript class="nojs"><img src="assets/wai_evolution_curve.png" alt="Best-so-far screening score across accepted generations." /></noscript>
+      </div>
+
+      <figure class="block">
+        <a class="zoom" href="assets/wai_tree.png" target="_blank" rel="noopener">
+          <img src="assets/wai_tree.png" alt="Archive lineage tree with accepted and reverted variants, the primary lineage, and recombination edges." loading="lazy" />
+        </a>
+        <figcaption>
+          The same run as an archive lineage tree (node size &prop; screening score): accepted (blue) and
+          reverted (grey) variants, the primary lineage (gold, base&nbsp;&rarr;&nbsp;evolved), and
+          recombination edges (dashed). The reverted nodes are the preserve-and-extend contract doing its job:
+          a variant that broke an inherited solve contributes its distilled lesson but not its edit.
+        </figcaption>
+      </figure>
+
+      <div class="block">
+        <h4>Anti-cheating audit</h4>
+        <table style="margin-top:12px">
+          <caption>Validity audit on the 1,260 real tasks. Trajectories that reach the goal through the
+            evaluation plane, privileged knowledge, or exploits are counted as failures.</caption>
+          <thead>
+            <tr><th>Measure</th><th class="num">Monet (base)</th><th class="num">Monet (DarwinX)</th><th class="num">&Delta;</th></tr>
+          </thead>
+          <tbody>
+            <tr class="ours"><td>Audit-clean pass@1</td><td class="num">43.5</td><td class="num">93.0</td><td class="num delta">+49.5 pp</td></tr>
+            <tr><td>Confirmed invalid</td><td class="num">23.5%</td><td class="num">1.4%</td><td class="num">&minus;22.1 pp</td></tr>
+            <tr><td>Human review</td><td class="num">5.1%</td><td class="num">0.1%</td><td class="num">&minus;5.0 pp</td></tr>
+            <tr><td>Blocked attempt</td><td class="num">14.3%</td><td class="num">0.1%</td><td class="num">&minus;14.2 pp</td></tr>
+          </tbody>
+        </table>
+        <p class="tnote">
+          Evolution does not buy the gain with shortcuts: it cuts invalid trajectories from 293 to 17, the
+          evaluation-plane, privileged-knowledge, and exploit mechanisms disappear entirely, and the residual
+          17 are all raw-state mutations concentrated in a single application. Dropping unaudited successes
+          too still leaves 92.9%.
+        </p>
+        <div class="block">
+          <div class="widget" id="iv-widget">
+            <div class="wtop">
+              <span class="wlabel">Break down by</span>
+              <div class="chips" id="iv-chips"></div>
+              <span style="flex:1"></span>
+              <label class="toggle"><input type="checkbox" id="iv-norm" /> scale each row to its own total</label>
+            </div>
+            <div id="iv-rows"></div>
+            <div class="keys" id="iv-keys"></div>
+            <p class="tnote" id="iv-note"></p>
+          </div>
+          <noscript class="nojs"><img src="assets/wai_invalid_composition.png" alt="Invalid trajectories before vs. after evolution, 293 to 17." /></noscript>
+          <p class="tnote">
+            Invalid trajectories before vs. after evolution. On a shared scale the second row is almost
+            invisible, which is the result: 293 invalid trajectories become 17. Scale each row to its own
+            total to inspect what the remainder is made of.
+          </p>
+        </div>
+
+        <div class="block">
+          <h4>What the audit costs each harness</h4>
+          <p class="tnote" style="margin-top:0;margin-bottom:14px">
+            Each line runs from raw pass@1 to pass@1 after the validity audit. A long line means the harness
+            was leaning on trajectories the audit rejects.
+          </p>
+          <div class="widget" id="ad-widget">
+            <div class="wtop">
+              <span class="wlabel">Show</span>
+              <label class="toggle"><input type="checkbox" id="ad-base" checked /> Monet (base)</label>
+              <label class="toggle"><input type="checkbox" id="ad-dx" checked /> Monet (DarwinX)</label>
+              <span style="flex:1"></span>
+              <label class="toggle"><input type="checkbox" id="ad-sort" /> sort by audit loss</label>
+            </div>
+            <div class="dumb" id="ad-rows"></div>
+            <div class="wlegend">
+              <span><i class="sw" style="background:#fff;box-shadow:inset 0 0 0 2px #9aa8b8;border-radius:50%"></i> raw pass@1</span>
+              <span><i class="sw" style="background:#8794a6;border-radius:50%"></i> base, audited</span>
+              <span><i class="sw" style="background:var(--blue);border-radius:50%"></i> DarwinX, audited</span>
+            </div>
+            <p class="tnote" id="ad-note"></p>
+          </div>
+          <noscript class="nojs"><img src="assets/wai_audit_by_app.png" alt="Raw vs. audited pass@1 per application." /></noscript>
+        </div>
+      </div>
+    </div>
+
+    <!-- ============ SWE-bench Verified ============ -->
+    <div class="bench">
+      <h3>SWE-bench Verified <span class="tag">&mdash; zero-shot cross-benchmark transfer</span></h3>
+      <p class="lede">The best Terminal-Bench&nbsp;2.1 harness is run <i>unchanged</i> on all 500 SWE-bench
+        Verified instances. No SWE-V feedback ever enters the loop; we report it as a transfer target only and
+        run no in-domain SWE-V evolution.</p>
+
+      <table>
+        <caption>SWE-bench Verified official pass@1, graded by SWE-V's own test harness.</caption>
+        <thead>
+          <tr><th>Harness</th><th class="num">Solved</th><th class="num">pass@1</th></tr>
+        </thead>
+        <tbody>
+          <tr class="ours"><td>Monet (DarwinX), TB2.1-specialized, transferred unchanged</td><td class="num">421/500</td><td class="num">84.2%</td></tr>
+          <tr><td>Fix-skill reference</td><td class="num">&mdash;</td><td class="num">80.8%</td></tr>
+        </tbody>
+      </table>
+      <p class="tnote">
+        +3.4 points over the reference without receiving any SWE-V feedback. The transferred agent preserves
+        strong repository-level competence, so what evolved on terminal tasks is general agent capability
+        rather than benchmark-specific patching.
+      </p>
+    </div>
+  </div>
+</section>
+
+<section id="ablation">
+  <div class="wrap">
+    <h2>Ablation: what evolution changes</h2>
+    <p>
+      Anchored on Terminal-Bench&nbsp;2.1, we compare base Monet (75.5%) with the evolved best (83.2%) along
+      two axes: the <b>skill-bundle diff</b> between the two harnesses, and <b>where</b> on the benchmark the
+      gains land. This is an <i>exploratory attribution, not a per-skill causal ablation</i> &mdash; the skills
+      were co-selected, not independently randomized.
+    </p>
+
+    <table style="margin-top:20px">
+      <caption>The seven skills the evolved lineage adds over base Monet. Every one belongs to a single
+        family: verification / artifact-contract. None adds domain knowledge.</caption>
+      <thead>
+        <tr><th style="width:42%">Evolved skills</th><th>Role</th></tr>
+      </thead>
+      <tbody>
+        <tr>
+          <td><code>verifier-contract</code><br /><code>contract-candidate</code></td>
+          <td>Derive the task's acceptance contract and check the solution against it before finalizing.</td>
+        </tr>
+        <tr>
+          <td><code>graded-artifact-final-check</code><br /><code>artifact-verification-loop</code></td>
+          <td>Verify the graded artifact (output file, format, and values) and iterate a fix-and-recheck loop.</td>
+        </tr>
+        <tr>
+          <td><code>real-tool-artifact</code><br /><code>tool-grounded-artifact</code></td>
+          <td>Ground outputs in real tool execution rather than asserted or simulated results.</td>
+        </tr>
+        <tr>
+          <td><code>security-contract-repair</code></td>
+          <td>Repair the solution against security and contract checks.</td>
+        </tr>
+      </tbody>
+    </table>
+
+    <div class="block" style="margin-top:30px">
+      <h4>Where the gains land</h4>
+      <p>
+        Gains concentrate where a frozen base has the most headroom and where verifying against a contract
+        unblocks fragile multi-step work: <i>ML &amp; scientific-computing</i> (+14.8, 60&rarr;75%) and
+        <i>data/database</i> (+13.8, 84&rarr;98%). Clusters where the base is already strong barely move
+        (system administration 92&rarr;98%, security 85&rarr;84%, within noise), and no cluster regresses
+        beyond the per-task noise band. <b>That asymmetry is the empirical footprint of the
+        preserve-and-extend rule</b>: extend the fragile clusters, hold the solved ones fixed. The unblocked
+        difficulty is procedural &mdash; long dependency installs, environment setup, output verification,
+        multi-step tool use &mdash; rather than knowledge-bound, which matches a verification/artifact-contract
+        bundle rather than a stronger model.
+      </p>
+    </div>
+
+    <div class="block">
+      <h4>Cross-benchmark cross-checks</h4>
+      <p>
+        <b>WebArena-Infinity, the same family in a different modality.</b> The evolved browser harness adds
+        the same kind of skills (state and action contracts), and its confirmed-invalid rate falls from 23.5%
+        to 1.4% while audit-clean pass@1 rises 49.5 points, with the largest gains on state-changing
+        applications. Verification-before-finalization is thus visible under a different interaction modality
+        and reward source.
+      </p>
+      <p>
+        <b>TerminalWorld, diversity rather than a single skill.</b> Individual specialists solve 24&ndash;27
+        held-out tasks and the merged harness solves 28, so the archive is valuable as a source of
+        complementary behaviors even when the training-subset ranking is noisy.
+      </p>
+      <p class="tnote">
+        Together these make verification-before-finalization and contract-aware tool use a plausible shared
+        mechanism across benchmarks, offered as an explanation rather than a per-skill causal estimate.
+      </p>
+    </div>
+  </div>
+</section>
+
+<section id="limitations">
+  <div class="wrap">
+    <h2>Limitations</h2>
+    <ul class="lim">
+      <li><b>Transfer is measured in one direction, and is much smaller.</b> A TB2.1-evolved harness reaches
+        84.2% on SWE-bench Verified, ahead of the fix-skill reference but inside a narrow band around it
+        (official scores across the harnesses we compare span just 80.8&ndash;84.2%). The transferred gain is
+        far smaller than the in-domain ones, and we make no in-domain SWE-V claim.</li>
+      <li><b>Recombination's own contribution is not isolated.</b> The archive and merge machinery can
+        preserve and combine variants, but population search needs diverse wins before inheritance is useful,
+        and the contribution of recombination relative to single-lineage mutation still requires controlled
+        ablation.</li>
+      <li><b>The claim is system-level, not per-operator.</b> The experiments evaluate the complete system:
+        archive, parent selector, recombination operator, and inference effort are not independently
+        randomized, and public leaderboard rows use different models and effort settings. Matched-model deltas
+        support a harness claim; any individual operator's contribution, and the attribution to a
+        verification/contract mechanism, remain plausible rather than causal.</li>
+      <li><b>TerminalWorld is small.</b> With only 41 held-out tasks, one solve moves pass@1 by 2.4 points,
+        and the matched Opus comparison (25/41 vs. 28/41, McNemar <i>p</i>=0.45) is suggestive rather than
+        decisive.</li>
+      <li><b>The validity audit is not a formal sandbox.</b> Our WAI policy permits client-visible observation
+        and semantic application operations but rejects privileged knowledge, evaluation-plane access,
+        raw-state fabrication, database manipulation, and exploits. The static-plus-LLM audit is far stronger
+        than a keyword heuristic, yet deeply dynamic construction can require human review and a small number
+        of trajectories were unavailable, so we report both raw and conservative audit-clean scores.</li>
+      <li><b>Freezing the base model is a choice, not a ceiling.</b> It is what makes these deltas
+        attributable to the harness. Coupling harness selection with weight updates, and treating the harness
+        as an asset that outlives a base-model generation, are left to future work.</li>
+    </ul>
+  </div>
+</section>
+
+<section id="bibtex">
+  <div class="wrap">
+    <div class="bibhead">
+      <h2>BibTeX</h2>
+      <button id="copybtn" type="button">Copy</button>
+    </div>
+<pre id="bib">@article{zhang2026darwinx,
+  title   = {DarwinX: Evolving Agent Harnesses Through Natural Selection},
+  author  = {Zhang, Yifan and Dai, Yutong and Tan, Juntao and Yang, Luyu and
+             Mullur, Rishi and Hoang, Thai and Hu, Zhiyuan and Zhu, James and
+             Mui, Phil and Savarese, Silvio and Xu, Ran and Chen, Zeyuan},
+  journal = {arXiv preprint arXiv:2608.07545},
+  year    = {2026}
+}</pre>
+  </div>
+</section>
+
+<footer>
+  <div class="wrap">
+    Salesforce AI Research &middot; 2026. All figures and numbers are from
+    <a href="https://arxiv.org/abs/2608.07545">arXiv:2608.07545</a>.
+    Click any figure to open it at full resolution.
+  </div>
+</footer>
+
+<div id="tip" role="tooltip"></div>
+
+<script src="assets/data.js"></script>
+<script>
+(function () {
+  "use strict";
+  var D = window.DARWINX;
+  var $ = function (id) { return document.getElementById(id); };
+
+  /* ---------------- shared tooltip ---------------- */
+  var tip = $("tip");
+  function showTip(html, evt) {
+    tip.innerHTML = html;
+    tip.style.opacity = "1";
+    var r = tip.getBoundingClientRect();
+    var x = Math.min(Math.max(8, evt.clientX - r.width / 2), window.innerWidth - r.width - 8);
+    var y = evt.clientY - r.height - 12;
+    if (y < 8) { y = evt.clientY + 16; }
+    tip.style.left = x + "px";
+    tip.style.top = y + "px";
+  }
+  function hideTip() { tip.style.opacity = "0"; }
+  function bindTip(el, htmlFn) {
+    el.addEventListener("mousemove", function (e) { showTip(htmlFn(), e); });
+    el.addEventListener("mouseleave", hideTip);
+  }
+
+  /* ---------------- 0a. headline panels ---------------- */
+  (function () {
+    if (!D || !D.headline) { return; }
+    var P = D.headline;
+    var modes = [
+      { id: "own", label: "per panel" },
+      { id: "shared", label: "shared 0\u2013100" }
+    ];
+    var mode = "own";
+
+    var chips = $("hl-chips");
+    modes.forEach(function (m) {
+      var b = document.createElement("button");
+      b.className = "chip";
+      b.type = "button";
+      b.textContent = m.label;
+      b.setAttribute("aria-pressed", m.id === mode ? "true" : "false");
+      b.addEventListener("click", function () {
+        mode = m.id;
+        Array.prototype.forEach.call(chips.children, function (c) {
+          c.setAttribute("aria-pressed", c === b ? "true" : "false");
+        });
+        render();
+      });
+      chips.appendChild(b);
+    });
+
+    var NS = "http://www.w3.org/2000/svg";
+    function mk(n, a) {
+      var e = document.createElementNS(NS, n);
+      Object.keys(a).forEach(function (k) { e.setAttribute(k, a[k]); });
+      return e;
+    }
+
+    function render() {
+      var showPrior = $("hl-prior").checked;
+      var box = $("hl-panels");
+      box.innerHTML = "";
+
+      P.forEach(function (p) {
+        var lo = mode === "shared" ? 0 : p.lo;
+        var hi = mode === "shared" ? 100 : p.hi;
+        var W = 190, H = 168, T = 20, B = 22, L = 26;
+        var ih = H - T - B;
+        var y = function (v) { return T + ih - ((v - lo) / (hi - lo)) * ih; };
+
+        var series = [{ c: "b", label: p.baseLabel, v: p.base }];
+        if (showPrior && p.prior !== null) { series.push({ c: "p", label: p.priorLabel, v: p.prior }); }
+        series.push({ c: "e", label: p.evolvedLabel, v: p.evolved });
+
+        var wrapEl = document.createElement("div");
+        wrapEl.className = "panel";
+        wrapEl.innerHTML = "<div class=\"pt\">" + p.short + "</div><div class=\"ps\">" +
+          p.metric + "<br>on " + p.model + "</div>";
+
+        var svg = mk("svg", { viewBox: "0 0 " + W + " " + H });
+        [lo, hi].forEach(function (v) {
+          var t = mk("text", { class: "tick", x: L - 6, y: y(v) + 3.5, "text-anchor": "end" });
+          t.textContent = v;
+          svg.appendChild(t);
+        });
+        svg.appendChild(mk("line", { class: "axl", x1: L, x2: W - 6, y1: y(lo), y2: y(lo) }));
+
+        var slot = (W - 6 - L) / series.length;
+        var bw = Math.min(34, slot * 0.62);
+        series.forEach(function (s, i) {
+          var cx = L + slot * (i + 0.5);
+          var top = y(Math.max(s.v, lo));
+          var rect = mk("rect", {
+            class: "bar " + s.c, x: cx - bw / 2, y: top, width: bw,
+            height: Math.max(0, y(lo) - top), rx: 2
+          });
+          svg.appendChild(rect);
+          var hit = mk("rect", { x: cx - slot / 2, y: T, width: slot, height: ih, fill: "transparent" });
+          hit.addEventListener("mousemove", function (e) {
+            showTip("<b>" + p.name + "</b><br>" + s.label + ": " + s.v.toFixed(1) + "%<br>" +
+                    p.metric + ", " + p.scope, e);
+          });
+          hit.addEventListener("mouseleave", hideTip);
+          svg.appendChild(hit);
+        });
+
+        var ev = series[series.length - 1];
+        var dl = mk("text", { class: "dl", x: L + slot * (series.length - 0.5),
+                              y: y(ev.v) - 6, "text-anchor": "middle" });
+        dl.textContent = "+" + p.delta.toFixed(1);
+        svg.appendChild(dl);
+
+        wrapEl.appendChild(svg);
+        var foot = document.createElement("div");
+        foot.className = "foot";
+        foot.textContent = p.prior === null ? "transfer only" : "vs " + p.priorLabel.split(" + ")[0];
+        wrapEl.appendChild(foot);
+        box.appendChild(wrapEl);
+      });
+
+      $("hl-legend").innerHTML =
+        "<span><i class=\"sw\" style=\"background:#c6d3e0\"></i> base / reference</span>" +
+        (showPrior ? "<span><i class=\"sw\" style=\"background:#8fb4e8\"></i> best prior agent</span>" : "") +
+        "<span><i class=\"sw\" style=\"background:var(--blue)\"></i> Monet (DarwinX)</span>";
+
+      $("hl-note").innerHTML = mode === "own"
+        ? "Each panel is scaled to its own range, which is how the paper draws it: on a shared axis " +
+          "WebArena-Infinity's +49.5 flattens the two terminal benchmarks into slivers. Switch to the " +
+          "shared axis to see the honest relative magnitudes \u2014 the annotated deltas are the same either way."
+        : "On one 0\u2013100 axis the four regimes are directly comparable, and it is clear how much of the " +
+          "range each benchmark actually occupies. Bars within a panel share one frozen base model; the " +
+          "prior-agent bar uses a different model and effort setting, so it is context rather than a " +
+          "controlled comparison. On SWE-bench Verified the grey bar is the fix-skill reference, not an " +
+          "unevolved Monet, and no prior-agent bar is drawn.";
+    }
+
+    $("hl-prior").addEventListener("change", render);
+    render();
+  })();
+
+  /* ---------------- 0b. TerminalWorld specialist bars ---------------- */
+  (function () {
+    if (!D || !D.twBars) { return; }
+    var T = D.twBars;
+    function pct(s) { return (100 * s / T.n); }
+    // The arms span 58.5-68.3, so a full 0-100 track renders them as six near-identical
+    // bars. Zoomed is the default (and how the paper draws it); the full axis stays one
+    // click away because a truncated axis exaggerates the spread.
+    var scales = [{ id: "zoom", label: "55\u201370", lo: 55, hi: 70 },
+                  { id: "full", label: "0\u2013100", lo: 0, hi: 100 }];
+    var scale = scales[0];
+
+    var chips = $("sp-chips");
+    scales.forEach(function (s) {
+      var b = document.createElement("button");
+      b.className = "chip";
+      b.type = "button";
+      b.textContent = s.label;
+      b.setAttribute("aria-pressed", s === scale ? "true" : "false");
+      b.addEventListener("click", function () {
+        scale = s;
+        Array.prototype.forEach.call(chips.children, function (c) {
+          c.setAttribute("aria-pressed", c === b ? "true" : "false");
+        });
+        render();
+      });
+      chips.appendChild(b);
+    });
+
+    function render() {
+      var counts = $("sp-counts").checked;
+      var pos = function (p) {
+        return Math.max(0, Math.min(100, 100 * (p - scale.lo) / (scale.hi - scale.lo)));
+      };
+      var refPos = pos(pct(T.claudeCode));
+      var box = $("sp-rows");
+      box.innerHTML = "";
+      T.bars.forEach(function (b) {
+        var isDx = /DarwinX/.test(b.label);
+        var isBase = b.label === "Base";
+        var row = document.createElement("div");
+        row.className = "hrow" + (isBase ? " dim" : "");
+        var color = isDx ? "var(--blue)" : (isBase ? "#c6d3e0" : "#8fb4e8");
+        row.innerHTML =
+          "<div class=\"hn\"" + (isDx ? " style=\"font-weight:600\"" : "") + ">" + b.label + "</div>" +
+          "<div class=\"ht\"><div class=\"hf\" style=\"width:" + pos(pct(b.solved)).toFixed(1) +
+            "%;background:" + color + "\"></div></div>" +
+          "<div class=\"hv\">" + (counts ? b.solved + "/" + T.n : pct(b.solved).toFixed(1) + "%") + "</div>";
+        var track = row.querySelector(".ht");
+        var ref = document.createElement("div");
+        ref.className = "refline";
+        ref.style.left = refPos.toFixed(1) + "%";
+        track.appendChild(ref);
+        bindTip(track, function () {
+          return "<b>" + b.label + "</b><br>" + b.solved + " of " + T.n + " held-out tasks (" +
+            pct(b.solved).toFixed(1) + "%)<br>Claude Code reference: " + T.claudeCode + "/" + T.n +
+            " (" + pct(T.claudeCode).toFixed(1) + "%)";
+        });
+        box.appendChild(row);
+      });
+      var tag = document.createElement("div");
+      tag.className = "hrow";
+      tag.innerHTML = "<div></div><div style=\"position:relative;height:14px\">" +
+        "<div class=\"reftag\" style=\"left:" + refPos.toFixed(1) + "%;top:0\">Claude Code " +
+        pct(T.claudeCode).toFixed(1) + "%</div></div><div></div>";
+      box.appendChild(tag);
+
+      $("sp-note").innerHTML = scale.id === "zoom"
+        ? "Bars start at 55%, not 0, so the arms are distinguishable \u2014 the whole spread is 9.8 points. " +
+          "Switch to the full axis to see that spread in proportion."
+        : "On the full axis every arm looks alike, which is the honest framing of the magnitude: the merge " +
+          "buys 2.4 points over the best single specialist. The ordering is what matters here, not the size.";
+    }
+    $("sp-counts").addEventListener("change", render);
+    render();
+  })();
+
+  /* ---------------- 0c. TB2.1 compute ---------------- */
+  (function () {
+    if (!D || !D.tb21Compute) { return; }
+    var C = D.tb21Compute;
+    var metrics = [
+      { id: "turns", label: "turns", unit: "", fmt: function (v) { return v; } },
+      { id: "tokens", label: "tokens", unit: "K", fmt: function (v) { return v + "K"; } }
+    ];
+    var active = "turns";
+
+    var chips = $("cp-chips");
+    metrics.forEach(function (m) {
+      var b = document.createElement("button");
+      b.className = "chip";
+      b.type = "button";
+      b.textContent = m.label;
+      b.setAttribute("aria-pressed", m.id === active ? "true" : "false");
+      b.addEventListener("click", function () {
+        active = m.id;
+        Array.prototype.forEach.call(chips.children, function (c) {
+          c.setAttribute("aria-pressed", c === b ? "true" : "false");
+        });
+        render();
+      });
+      chips.appendChild(b);
+    });
+
+    function render() {
+      var m = metrics.filter(function (x) { return x.id === active; })[0];
+      var d = C[active];
+      var max = Math.max.apply(null, d.base.concat(d.evolved));
+      var box = $("cp-rows");
+      box.innerHTML = "";
+      C.groups.forEach(function (g, gi) {
+        [["Monet (base)", d.base[gi], "#c6d3e0"], ["Monet (DarwinX)", d.evolved[gi], "var(--blue)"]]
+          .forEach(function (s, si) {
+            var row = document.createElement("div");
+            row.className = "hrow";
+            row.innerHTML =
+              "<div class=\"hn\">" + (si === 0 ? g : "") + "</div>" +
+              "<div class=\"ht\"><div class=\"hf\" style=\"width:" + (100 * s[1] / max).toFixed(1) +
+                "%;background:" + s[2] + "\"></div></div>" +
+              "<div class=\"hv\">" + m.fmt(s[1]) + "</div>";
+            bindTip(row.querySelector(".ht"), function () {
+              var other = si === 0 ? d.evolved[gi] : d.base[gi];
+              var ratio = si === 0 ? (other / s[1]) : (s[1] / other);
+              return "<b>" + g + "</b><br>" + s[0] + ": " + m.fmt(s[1]) + " median " + m.label +
+                "<br>evolved / base = " + ratio.toFixed(1) + "x";
+            });
+            box.appendChild(row);
+          });
+      });
+      $("cp-note").innerHTML = active === "turns"
+        ? "On the six tasks that flip from failing to solved the evolved harness roughly doubles its turns " +
+          "(22 vs. 11). On the 69 both already solve, turns barely move (13 vs. 12): the extra effort is " +
+          "spent where it changes the outcome, not spread across the benchmark."
+        : "Tokens tell the same story more sharply: 4.3x on the newly solved six (380K vs. 89K) against " +
+          "1.4x on the 69 already solved (172K vs. 125K). This is verify-and-retry effort converting a " +
+          "near-miss into a pass, not uniformly more compute.";
+    }
+    render();
+  })();
+
+  /* ---------------- 0d. WAI invalid composition ---------------- */
+  (function () {
+    if (!D || !D.waiInvalid) { return; }
+    var V = D.waiInvalid;
+    var views = [{ id: "application", label: "application" }, { id: "mechanism", label: "mechanism" }];
+    var active = "mechanism";
+
+    var chips = $("iv-chips");
+    views.forEach(function (v) {
+      var b = document.createElement("button");
+      b.className = "chip";
+      b.type = "button";
+      b.textContent = v.label;
+      b.setAttribute("aria-pressed", v.id === active ? "true" : "false");
+      b.addEventListener("click", function () {
+        active = v.id;
+        Array.prototype.forEach.call(chips.children, function (c) {
+          c.setAttribute("aria-pressed", c === b ? "true" : "false");
+        });
+        render();
+      });
+      chips.appendChild(b);
+    });
+
+    function render() {
+      var v = V[active];
+      var norm = $("iv-norm").checked;
+      var totalBefore = v.before.reduce(function (a, b) { return a + b; }, 0);
+      var box = $("iv-rows");
+      box.innerHTML = "";
+
+      [["Before evolution", v.before], ["After evolution", v.after]].forEach(function (pair) {
+        var vals = pair[1];
+        var sum = vals.reduce(function (a, b) { return a + b; }, 0);
+        var denom = norm ? (sum || 1) : totalBefore;
+        var row = document.createElement("div");
+        row.className = "srow";
+        row.innerHTML = "<div class=\"sh\"><span>" + pair[0] +
+          "</span><span><b>" + sum + "</b> invalid</span></div>";
+        var st = document.createElement("div");
+        st.className = "stack";
+        vals.forEach(function (val, i) {
+          if (!val) { return; }
+          var seg = document.createElement("i");
+          seg.style.width = (100 * val / denom).toFixed(2) + "%";
+          seg.style.background = v.colors[i];
+          bindTip(seg, function () {
+            return "<b>" + v.labels[i] + "</b><br>" + val + " invalid trajector" +
+              (val === 1 ? "y" : "ies") + "<br>" + (100 * val / (sum || 1)).toFixed(1) +
+              "% of this row";
+          });
+          st.appendChild(seg);
+        });
+        row.appendChild(st);
+        box.appendChild(row);
+      });
+
+      $("iv-keys").innerHTML = v.labels.map(function (l, i) {
+        if (!v.before[i] && !v.after[i]) { return ""; }
+        return "<span><i class=\"sw\" style=\"background:" + v.colors[i] + "\"></i>" + l + "</span>";
+      }).join("");
+
+      $("iv-note").innerHTML = active === "mechanism"
+        ? "Evaluation-plane access, privileged knowledge, and exploits go to <b>zero</b>. Every one of the " +
+          "17 remaining invalid trajectories is a raw-state mutation, up from 15 of 293 before \u2014 the " +
+          "category that survives is the mildest one."
+        : "The remainder concentrates: 8 of the 17 sit in a single application (Super-human), and five of " +
+          "the ten applications end with no invalid trajectories at all.";
+    }
+    $("iv-norm").addEventListener("change", render);
+    render();
+  })();
+
+  /* ---------------- 0e. WAI audit dumbbell ---------------- */
+  (function () {
+    if (!D || !D.waiAudit) { return; }
+    var A = D.waiAudit;
+
+    function render() {
+      var showBase = $("ad-base").checked, showDx = $("ad-dx").checked;
+      var idx = A.apps.map(function (_, i) { return i; });
+      var overall = idx.shift();
+      if ($("ad-sort").checked) {
+        idx.sort(function (a, b) {
+          return (A.rawBase[b] - A.auditedBase[b]) - (A.rawBase[a] - A.auditedBase[a]);
+        });
+      }
+      var box = $("ad-rows");
+      box.innerHTML = "";
+
+      var ax = document.createElement("div");
+      ax.className = "drow dax";
+      ax.innerHTML = "<div class=\"dn\" style=\"color:var(--muted);font-size:0.7rem\">pass@1 (%)</div>" +
+        "<div class=\"dt\">" + [0, 25, 50, 75, 100].map(function (t) {
+          return "<span style=\"left:" + t + "%\">" + t + "</span>";
+        }).join("") + "</div>";
+      box.appendChild(ax);
+
+      [overall].concat(idx).forEach(function (i, pos) {
+        var row = document.createElement("div");
+        row.className = "drow" + (pos === 0 ? " tot" : "");
+        row.innerHTML = "<div class=\"dn\">" + A.apps[i] + "</div><div class=\"dt\"></div>";
+        var t = row.querySelector(".dt");
+
+        function lane(raw, aud, color, top) {
+          var x1 = Math.min(raw, aud), x2 = Math.max(raw, aud);
+          var seg = document.createElement("div");
+          seg.className = "seg";
+          seg.style.left = x1 + "%";
+          seg.style.width = (x2 - x1) + "%";
+          seg.style.top = top + "px";
+          seg.style.background = color;
+          t.appendChild(seg);
+          var p1 = document.createElement("div");
+          p1.className = "pt2 hollow";
+          p1.style.left = raw + "%";
+          p1.style.top = (top + 1.5) + "px";
+          p1.style.boxShadow = "inset 0 0 0 2px " + color;
+          t.appendChild(p1);
+          var p2 = document.createElement("div");
+          p2.className = "pt2";
+          p2.style.left = aud + "%";
+          p2.style.top = (top + 1.5) + "px";
+          p2.style.background = color;
+          t.appendChild(p2);
+        }
+
+        if (showBase) { lane(A.rawBase[i], A.auditedBase[i], "#8794a6", 8); }
+        if (showDx) { lane(A.rawDx[i], A.auditedDx[i], "#0176d3", 19); }
+
+        bindTip(t, function () {
+          return "<b>" + A.apps[i] + "</b><br>base " + A.rawBase[i].toFixed(1) + " \u2192 " +
+            A.auditedBase[i].toFixed(1) + " (\u2212" + (A.rawBase[i] - A.auditedBase[i]).toFixed(1) +
+            ")<br>DarwinX " + A.rawDx[i].toFixed(1) + " \u2192 " + A.auditedDx[i].toFixed(1) +
+            " (\u2212" + (A.rawDx[i] - A.auditedDx[i]).toFixed(1) + ")";
+        });
+        box.appendChild(row);
+      });
+
+      var lossB = A.rawBase[0] - A.auditedBase[0], lossD = A.rawDx[0] - A.auditedDx[0];
+      $("ad-note").innerHTML = "Overall the audit costs the base <b>" + lossB.toFixed(1) +
+        "</b> points and DarwinX <b>" + lossD.toFixed(1) + "</b>. The gap therefore <i>widens</i> under " +
+        "scrutiny, from " + (A.rawDx[0] - A.rawBase[0]).toFixed(1) + " points raw to " +
+        (A.auditedDx[0] - A.auditedBase[0]).toFixed(1) + " audited, which is why the page reports the " +
+        "audited number. Audit coverage is 99.0% for the base and 99.4% for the evolved harness.";
+    }
+    ["ad-base", "ad-dx", "ad-sort"].forEach(function (id) {
+      $(id).addEventListener("change", render);
+    });
+    render();
+  })();
+
+  /* ---------------- 1. TerminalWorld merge explorer ---------------- */
+  (function () {
+    if (!D || !D.tw) { return; }
+    var tasks = D.tw.tasks;
+    var counts = D.tw.counts;
+    var KEYS = ["A", "B", "C", "D"];
+    var sel = { A: true, B: true, C: true, D: true };
+
+    var chips = $("tw-chips");
+    KEYS.forEach(function (k, i) {
+      var b = document.createElement("button");
+      b.className = "chip";
+      b.type = "button";
+      b.setAttribute("aria-pressed", "true");
+      b.innerHTML = "Spec. " + k + " <span class=\"cnt\">" + counts[k] + "/41</span>";
+      b.addEventListener("click", function () {
+        sel[k] = !sel[k];
+        b.setAttribute("aria-pressed", sel[k] ? "true" : "false");
+        render();
+      });
+      chips.appendChild(b);
+      void i;
+    });
+
+    var grid = $("tw-grid");
+    var cells = tasks.map(function (t) {
+      var c = document.createElement("div");
+      c.className = "tcell";
+      grid.appendChild(c);
+      bindTip(c, function () {
+        var solvers = KEYS.filter(function (k) { return t.bits[KEYS.indexOf(k)] === "1"; });
+        return "<b>" + t.id + "</b><br>solved by: " +
+          (solvers.length ? "Spec. " + solvers.join(", ") : "no specialist") +
+          "<br>realized merge: " + (t.bits[4] === "1" ? "solved" : "unsolved");
+      });
+      return c;
+    });
+
+    function unionOf(keys) {
+      return tasks.filter(function (t) {
+        return keys.some(function (k) { return t.bits[KEYS.indexOf(k)] === "1"; });
+      });
+    }
+
+    function render() {
+      var on = KEYS.filter(function (k) { return sel[k]; });
+      var u = unionOf(on);
+      var uIds = {};
+      u.forEach(function (t) { uIds[t.id] = 1; });
+      var cmp = $("tw-cmp").checked;
+
+      cells.forEach(function (c, i) {
+        var t = tasks[i];
+        var inU = !!uIds[t.id];
+        var inM = t.bits[4] === "1";
+        c.className = "tcell" + (inU ? " on" : "") +
+          (cmp && inM && !inU ? " gain" : "") + (cmp && inU && !inM ? " loss" : "");
+      });
+
+      var best = on.length ? Math.max.apply(null, on.map(function (k) { return counts[k]; })) : 0;
+      $("tw-readout").innerHTML =
+        ro("Selected", on.length ? on.join(" + ") : "none", true) +
+        ro("Union coverage", u.length + "<small>/41</small>") +
+        ro("Best single selected", (best || "\u2014") + (best ? "<small>/41</small>" : "")) +
+        ro("Realized merge", counts.M + "<small>/41</small>");
+
+      var note = "";
+      if (on.length === 4) {
+        note = "All four specialists together cover <b>29/41</b>, two more than the best single lineage " +
+          "(Spec. D at 27). The harness DarwinX actually produced solves <b>28/41</b>: recombination is not " +
+          "a free set union. Turn on the comparison to see the difference \u2014 the merge picks up " +
+          "<code>tw_448247</code>, which <i>no</i> specialist solves, and drops <code>tw_449421</code> and " +
+          "<code>tw_498533</code>. These are single attempts, so individual flips sit inside the noise band " +
+          "(one task = 2.4 points).";
+      } else if (on.length === 0) {
+        note = "Nothing selected. 12 of the 41 held-out tasks are solved by none of the four specialists.";
+      } else if (u.length > best) {
+        note = "This selection covers <b>" + u.length + "/41</b>, which is " + (u.length - best) +
+          " more than its best individual member \u2014 the specialists solve overlapping but distinct subsets.";
+      } else {
+        note = "This selection covers <b>" + u.length + "/41</b>, no more than its best individual member: " +
+          "these lineages are not complementary.";
+      }
+      $("tw-note").innerHTML = note;
+
+      var lg = $("tw-legend");
+      lg.innerHTML =
+        "<span><i class=\"sw\" style=\"background:var(--blue)\"></i> solved by selection</span>" +
+        "<span><i class=\"sw\" style=\"background:#e6ecf2\"></i> unsolved</span>" +
+        (cmp ? "<span><i class=\"sw\" style=\"box-shadow:0 0 0 2px #e8a33d;background:#e6ecf2\"></i>" +
+               " merge solves, selection does not</span>" +
+               "<span><i class=\"sw\" style=\"box-shadow:0 0 0 2px #d05a4e;background:var(--blue)\"></i>" +
+               " selection solves, merge does not</span>" : "");
+    }
+
+    function ro(k, v, plain) {
+      return "<div class=\"ro" + (plain ? "" : " hi") + "\"><div class=\"k\">" + k +
+        "</div><div class=\"v\"" + (plain ? " style=\"font-size:0.95rem;font-weight:600\"" : "") + ">" +
+        v + "</div></div>";
+    }
+
+    $("tw-cmp").addEventListener("change", render);
+    render();
+  })();
+
+  /* ---------------- 2. TB2.1 cluster explorer ---------------- */
+  (function () {
+    if (!D || !D.tb21Clusters) { return; }
+    var data = D.tb21Clusters.slice();
+    var modes = [
+      { id: "delta", label: "gain", cmp: function (a, b) { return b.delta - a.delta; } },
+      { id: "base", label: "base rate", cmp: function (a, b) { return a.base - b.base; } },
+      { id: "n", label: "cluster size", cmp: function (a, b) { return b.n - a.n; } },
+      { id: "name", label: "name", cmp: function (a, b) { return a.name.localeCompare(b.name); } }
+    ];
+    var active = "delta";
+
+    var chips = $("cl-chips");
+    modes.forEach(function (m) {
+      var b = document.createElement("button");
+      b.className = "chip";
+      b.type = "button";
+      b.textContent = m.label;
+      b.setAttribute("aria-pressed", m.id === active ? "true" : "false");
+      b.addEventListener("click", function () {
+        active = m.id;
+        Array.prototype.forEach.call(chips.children, function (c) {
+          c.setAttribute("aria-pressed", c === b ? "true" : "false");
+        });
+        render();
+      });
+      chips.appendChild(b);
+    });
+
+    function render() {
+      var mode = modes.filter(function (m) { return m.id === active; })[0];
+      var rows = data.slice().sort(mode.cmp);
+      var showBase = $("cl-base").checked;
+      var box = $("cl-bars");
+      box.innerHTML = "";
+      rows.forEach(function (d) {
+        var el = document.createElement("div");
+        el.className = "cbar";
+        el.innerHTML =
+          "<div class=\"cname\">" + d.name + " <span class=\"cn\">n=" + d.n + "</span></div>" +
+          "<div class=\"track\">" +
+            (showBase ? "<div class=\"fill b\" style=\"left:0;width:" + d.base + "%\"></div>" : "") +
+            "<div class=\"fill e\" style=\"left:0;width:" + d.evolved + "%;height:" +
+              (showBase ? "58%;top:21%" : "100%;top:0") + "\"></div>" +
+          "</div>" +
+          "<div class=\"cd " + (d.delta >= 0 ? "up" : "down") + "\">" +
+            (d.delta >= 0 ? "+" : "\u2212") + Math.abs(d.delta).toFixed(1) + "</div>";
+        bindTip(el.querySelector(".track"), function () {
+          return "<b>" + d.name + "</b><br>base " + d.base.toFixed(1) + "% \u2192 evolved " +
+            d.evolved.toFixed(1) + "%<br>" + d.n + " tasks, \u0394 " +
+            (d.delta >= 0 ? "+" : "\u2212") + Math.abs(d.delta).toFixed(1) + " points";
+        });
+        box.appendChild(el);
+      });
+    }
+
+    $("cl-base").addEventListener("change", render);
+    render();
+  })();
+
+  /* ---------------- 3. WAI evolution curve ---------------- */
+  (function () {
+    if (!D || !D.waiCurve) { return; }
+    var s = D.waiCurve;
+    var svg = $("cv-chart");
+    var W = 860, H = 320, L = 44, R = 14, T = 14, B = 34;
+    var iw = W - L - R, ih = H - T - B;
+    var NS = "http://www.w3.org/2000/svg";
+
+    function x(i) { return L + (s.length === 1 ? iw / 2 : (i / (s.length - 1)) * iw); }
+    function y(v) { return T + ih - (v / 100) * ih; }
+    function mk(n, a) {
+      var e = document.createElementNS(NS, n);
+      Object.keys(a).forEach(function (k) { e.setAttribute(k, a[k]); });
+      return e;
+    }
+
+    function render() {
+      while (svg.firstChild) { svg.removeChild(svg.firstChild); }
+
+      [0, 25, 50, 75, 100].forEach(function (v) {
+        svg.appendChild(mk("line", { class: "gl", x1: L, x2: W - R, y1: y(v), y2: y(v) }));
+        var t = mk("text", { class: "lbl", x: L - 8, y: y(v) + 3.5, "text-anchor": "end" });
+        t.textContent = v;
+        svg.appendChild(t);
+      });
+      svg.appendChild(mk("line", { class: "ax", x1: L, x2: L, y1: T, y2: T + ih }));
+      svg.appendChild(mk("line", { class: "ax", x1: L, x2: W - R, y1: T + ih, y2: T + ih }));
+
+      [0, 9, 18, 27, 36].forEach(function (i) {
+        if (i >= s.length) { return; }
+        var t = mk("text", { class: "lbl", x: x(i), y: T + ih + 17, "text-anchor": "middle" });
+        t.textContent = i;
+        svg.appendChild(t);
+      });
+      var xl = mk("text", { class: "lbl", x: L + iw / 2, y: H - 3, "text-anchor": "middle" });
+      xl.textContent = "archive node";
+      svg.appendChild(xl);
+      var yl = mk("text", { class: "lbl", x: 12, y: T + ih / 2, "text-anchor": "middle",
+                            transform: "rotate(-90 12 " + (T + ih / 2) + ")" });
+      yl.textContent = "screening score (%)";
+      svg.appendChild(yl);
+
+      if ($("cv-best").checked) {
+        var m = -1, pts = [];
+        s.forEach(function (v, i) {
+          m = Math.max(m, v);
+          if (i > 0) { pts.push(x(i) + "," + y(m)); }
+          pts.push(x(i) + "," + y(m));
+        });
+        svg.appendChild(mk("polyline", { class: "best", points: pts.join(" ") }));
+      }
+
+      if ($("cv-all").checked) {
+        s.forEach(function (v, i) {
+          var dot = mk("circle", { class: "dot", cx: x(i), cy: y(v), r: 3.4 });
+          svg.appendChild(dot);
+          var hit = mk("circle", { class: "hit", cx: x(i), cy: y(v), r: 11 });
+          hit.addEventListener("mousemove", function (e) {
+            dot.classList.add("act");
+            showTip("<b>archive node " + i + "</b><br>screening score " + v.toFixed(1) + "%", e);
+          });
+          hit.addEventListener("mouseleave", function () { dot.classList.remove("act"); hideTip(); });
+          svg.appendChild(hit);
+        });
+      }
+    }
+
+    var best = Math.max.apply(null, s);
+    $("cv-note").innerHTML = "The run evaluates <b>" + s.length + "</b> variants. The screening score " +
+      "climbs from " + s[0].toFixed(1) + "% to <b>" + best.toFixed(1) + "%</b>, but not monotonically: " +
+      "individual variants score well below the running best, and those are the ones the " +
+      "preserve-and-extend contract reverts. Screening runs on synthetic intents, so this axis is the " +
+      "in-loop proxy, not the 1,260-task real result.";
+
+    $("cv-all").addEventListener("change", render);
+    $("cv-best").addEventListener("change", render);
+    render();
+  })();
+
+  /* ---------------- 4. sortable tables ---------------- */
+  (function () {
+    function num(td) {
+      var v = parseFloat(td.textContent.replace(/[^0-9.\-\u2212]/g, "").replace("\u2212", "-"));
+      return isNaN(v) ? null : v;
+    }
+    Array.prototype.forEach.call(document.querySelectorAll("table.sortable"), function (tb) {
+      var ths = tb.querySelectorAll("thead th");
+      Array.prototype.forEach.call(ths, function (th, ci) {
+        th.classList.add("sortable");
+        th.addEventListener("click", function () {
+          var body = tb.querySelector("tbody");
+          var rows = Array.prototype.slice.call(body.rows);
+          var pinned = rows.filter(function (r) { return r.classList.contains("total"); });
+          var movable = rows.filter(function (r) { return !r.classList.contains("total"); });
+          if (!movable.length) { return; }
+
+          // Numeric columns open descending (best first); text columns open A-Z.
+          var numeric = movable.some(function (r) { return num(r.cells[ci]) !== null; });
+          var desc;
+          if (th.classList.contains("desc")) { desc = false; }
+          else if (th.classList.contains("asc")) { desc = true; }
+          else { desc = numeric; }
+
+          Array.prototype.forEach.call(ths, function (o) { o.classList.remove("asc", "desc"); });
+          th.classList.add(desc ? "desc" : "asc");
+
+          movable.sort(function (a, b) {
+            var x = num(a.cells[ci]), y = num(b.cells[ci]);
+            if (x === null || y === null) {
+              var av = a.cells[ci].textContent.trim(), bv = b.cells[ci].textContent.trim();
+              return desc ? bv.localeCompare(av) : av.localeCompare(bv);
+            }
+            return desc ? y - x : x - y;
+          });
+          movable.concat(pinned).forEach(function (r) { body.appendChild(r); });
+        });
+      });
+    });
+  })();
+
+  /* ---------------- 5. BibTeX copy ---------------- */
+  $("copybtn").addEventListener("click", function () {
+    var btn = this;
+    navigator.clipboard.writeText($("bib").textContent).then(
+      function () {
+        btn.textContent = "Copied";
+        setTimeout(function () { btn.textContent = "Copy"; }, 1600);
+      },
+      function () { btn.textContent = "Press \u2318C"; }
+    );
+  });
+})();
+</script>
+
+</body>
 </html>

tools/build_data.py

@@ -0,0 +1,223 @@
+#!/usr/bin/env python3
+"""Emit assets/data.js for the DarwinX project page from the original run artifacts.
+
+Every number the interactive figures display comes from the files listed in SOURCES,
+so the page can be re-derived rather than hand-transcribed. Run from anywhere:
+
+    python3 tools/build_data.py
+"""
+
+import json
+import os
+import sys
+from datetime import datetime, timezone
+
+TW_DIR = ("/Users/yifan.zhang/Desktop/techreport/darwinx_zeyuan_feedback/"
+          "version_control/version_0717_local/results/tw/results")
+NOTES = "/Users/yifan.zhang/Desktop/techreport/darwinx_zeyuan_feedback/notes"
+
+TW_ARMS = [
+    ("A", "Spec. A", "opus_node_34f71461_run.json"),
+    ("B", "Spec. B", "opus_node_cab04955_run.json"),
+    ("C", "Spec. C", "opus_node_56fc77cc_run.json"),
+    ("D", "Spec. D", "opus_node_eb39901b_run.json"),
+    ("M", "Merged",  "TW_HEADLINE_merge_opus_28of41_run.json"),
+]
+
+# From notes/TB21_RESULTS.md (paired protocol, 88 tasks). Deltas are carried over as
+# reported rather than recomputed: the source rounds the delta from unrounded rates,
+# so evolved-minus-base on the displayed 1-decimal values differs by up to 0.1.
+TB21_CLUSTERS = [
+    ("ML & scientific computing", 23, 60.1, 74.9, 14.8),
+    ("Data & databases",           9, 83.9, 97.8, 13.8),
+    ("Algorithms & coding",       14, 83.1, 90.0,  6.9),
+    ("System admin & networking", 11, 92.0, 97.6,  5.7),
+    ("Parsing, text & media",     11, 67.3, 72.2,  5.0),
+    ("Systems & build",           11, 79.3, 81.8,  2.5),
+    ("Security & crypto",          9, 85.1, 84.4, -0.6),
+]
+
+# The four headline panels, from scripts/gen_summary_figure.py GROUPS. Each panel is
+# matched-model: base and evolved share one frozen base, so the gap is the harness.
+# The prior-agent value is context, not a controlled comparison, because those
+# systems use different models and effort settings. SWE-V has no prior-agent bar and
+# its "base" is the fix-skill reference rather than an unevolved Monet, so the label
+# travels with the number.
+HEADLINE = [
+    {"short": "TB 2.1", "name": "Terminal-Bench 2.1", "metric": "avg@5",
+     "model": "GPT-5.5", "scope": "89 tasks",
+     "baseLabel": "Monet (base)", "base": 75.5,
+     "evolvedLabel": "Monet (DarwinX)", "evolved": 83.2,
+     "priorLabel": "Codex", "prior": 83.1,
+     "delta": 7.7, "lo": 70, "hi": 88},
+    {"short": "T-World", "name": "TerminalWorld", "metric": "pass@1",
+     "model": "Opus 4.8", "scope": "41 held-out tasks",
+     "baseLabel": "Monet (base)", "base": 61.0,
+     "evolvedLabel": "Monet (DarwinX)", "evolved": 68.3,
+     "priorLabel": "Claude Code", "prior": 65.9,
+     "delta": 7.3, "lo": 55, "hi": 75},
+    {"short": "WA-Inf.", "name": "WebArena-Infinity", "metric": "audit-clean pass@1",
+     "model": "GPT-5.5", "scope": "1,260 real tasks",
+     "baseLabel": "Monet (base)", "base": 43.5,
+     "evolvedLabel": "Monet (DarwinX)", "evolved": 93.0,
+     "priorLabel": "GPT-5.5 + Browser Use", "prior": 86.1,
+     "delta": 49.5, "lo": 35, "hi": 100},
+    {"short": "SWE-V", "name": "SWE-bench Verified", "metric": "official pass@1",
+     "model": "Opus 4.8", "scope": "500 instances, transfer only",
+     "baseLabel": "Fix-skill reference", "base": 80.8,
+     "evolvedLabel": "Monet (DarwinX), transferred", "evolved": 84.2,
+     "priorLabel": None, "prior": None,
+     "delta": 3.4, "lo": 75, "hi": 88},
+]
+
+# scripts/gen_tw_dynamics.py + notes/tw_dynamics.json. Base is the unevolved v1.0.0
+# harness on Opus 4.8 (25/41); its per-task run.json is not on this machine, so the
+# merge explorer covers only the arms that have per-task records.
+TW_HELDOUT_BARS = [("Base", 25), ("Spec. A", 24), ("Spec. B", 25),
+                   ("Spec. C", 26), ("Spec. D", 27), ("Monet (DarwinX)", 28)]
+TW_CLAUDE_CODE = 27          # Claude Code (Opus 4.8) reference, 27/41
+TW_N = 41
+
+# scripts/gen_tb21_compute.py: medians precomputed from the trial archive.
+TB21_COMPUTE = {
+    "groups": ["Newly solved (6 tasks)", "Already solved (69 tasks)"],
+    "turns": {"base": [11, 12], "evolved": [22, 13]},
+    "tokens": {"base": [89, 125], "evolved": [380, 172]},   # thousands
+}
+
+# scripts/gen_wai_invalid_composition.py: 293 invalid trajectories before evolution,
+# 17 after, decomposed two ways over the same totals.
+WAI_INVALID = {
+    "application": {
+        "labels": ["Elation Clinical", "Elation Prescription", "GitLab Plan & Track",
+                   "Gmail", "Gmail Acct & Contact", "Handshake", "Linear Acct Mgmt",
+                   "PayPal Wallet", "Super-human", "Xero Invoicing"],
+        "before": [29, 27, 38, 21, 44, 54, 26, 10, 44, 0],
+        "after":  [2, 3, 0, 0, 3, 0, 0, 0, 8, 1],
+        "colors": ["#4C72B0", "#A6C8E0", "#DD8452", "#E3B23C", "#55A868",
+                   "#8FD19E", "#C44E52", "#E79EA6", "#8172B3", "#C7BFE0"],
+    },
+    "mechanism": {
+        "labels": ["Evaluation plane access", "Privileged knowledge",
+                   "Exploit or privilege escalation", "Raw state mutation"],
+        "before": [155, 97, 26, 15],
+        "after":  [0, 0, 0, 17],
+        "colors": ["#55A868", "#DD8452", "#CE6DBD", "#8FA6D1"],
+    },
+}
+
+# scripts/gen_wai_audit_by_app.py: pass@1 before and after the validity audit.
+WAI_AUDIT = {
+    "apps": ["Overall", "Elation-Clin", "Elation-Rx", "GitLab", "Gmail",
+             "Gmail-Acct", "Handshake", "Linear", "PayPal", "Superhuman", "Xero"],
+    "rawBase":      [53.0, 95.8, 28.3, 75.7, 45.0, 33.3, 44.0, 54.2, 50.0, 53.3, 49.2],
+    "rawDx":        [94.4, 98.3, 97.5, 97.9, 98.3, 94.2, 84.0, 94.2, 95.7, 94.2, 97.5],
+    "auditedBase":  [43.5, 95.8, 20.0, 63.6, 25.0, 21.7, 36.5, 43.3, 49.3, 31.7, 39.2],
+    "auditedDx":    [93.0, 96.7, 95.0, 97.9, 98.3, 91.7, 84.0, 94.2, 95.7, 87.5, 96.7],
+}
+
+
+def load_tw():
+    """Per-task resolved flags for the four specialists and the merged harness.
+
+    The run files list the 41 tasks in different orders, so every arm is joined on
+    task_id and the emitted order is canonical (the first arm's file order).
+    """
+    order, universe, sets = None, None, {}
+    for key, _label, fname in TW_ARMS:
+        path = os.path.join(TW_DIR, fname)
+        if not os.path.exists(path):
+            sys.exit(f"missing TerminalWorld run file: {path}")
+        rows = json.load(open(path))["per_task_results"]
+        ids = {r["task_id"] for r in rows}
+        if len(ids) != len(rows):
+            sys.exit(f"duplicate task_id in {fname}")
+        if order is None:
+            order, universe = [r["task_id"] for r in rows], ids
+        elif ids != universe:
+            sys.exit(f"task set differs in {fname}: "
+                     f"{sorted(ids ^ universe)} not shared; refusing to emit")
+        sets[key] = {r["task_id"] for r in rows if r["resolved"]}
+
+    tasks = [{"id": t, "bits": "".join("1" if t in sets[k] else "0"
+                                       for k, _l, _f in TW_ARMS)} for t in order]
+    counts = {k: len(sets[k]) for k, _l, _f in TW_ARMS}
+    if [counts[k] for k in "ABCD"] != [24, 25, 26, 27] or counts["M"] != 28:
+        sys.exit(f"unexpected TerminalWorld totals {counts}; paper reports 24/25/26/27 and 28")
+    return tasks, counts
+
+
+def load_wai_curve():
+    path = os.path.join(NOTES, "tw_dynamics.json")
+    if not os.path.exists(path):
+        sys.exit(f"missing {path}")
+    scores = json.load(open(path))["wai_adaptive_scores"]
+    return [round(float(s), 2) for s in scores]
+
+
+def main():
+    tasks, counts = load_tw()
+    curve = load_wai_curve()
+
+    payload = {
+        "generated": datetime.now(timezone.utc).strftime("%Y-%m-%d %H:%M UTC"),
+        "tw": {
+            "arms": [{"key": k, "label": l} for k, l, _f in TW_ARMS],
+            "tasks": tasks,
+            "counts": counts,
+        },
+        "tb21Clusters": [
+            {"name": n, "n": c, "base": b, "evolved": e, "delta": d}
+            for n, c, b, e, d in TB21_CLUSTERS
+        ],
+        "waiCurve": curve,
+        "headline": HEADLINE,
+        "twBars": {
+            "n": TW_N,
+            "claudeCode": TW_CLAUDE_CODE,
+            "bars": [{"label": l, "solved": s} for l, s in TW_HELDOUT_BARS],
+        },
+        "tb21Compute": TB21_COMPUTE,
+        "waiInvalid": WAI_INVALID,
+        "waiAudit": WAI_AUDIT,
+    }
+
+    # Guard the invariants the page states in prose.
+    for view in ("application", "mechanism"):
+        v = WAI_INVALID[view]
+        if sum(v["before"]) != 293 or sum(v["after"]) != 17:
+            sys.exit(f"WAI_INVALID[{view}] must total 293 before and 17 after, got "
+                     f"{sum(v['before'])} and {sum(v['after'])}")
+        if not (len(v["labels"]) == len(v["before"]) == len(v["after"]) == len(v["colors"])):
+            sys.exit(f"WAI_INVALID[{view}] has ragged columns")
+    for k in ("rawBase", "rawDx", "auditedBase", "auditedDx"):
+        if len(WAI_AUDIT[k]) != len(WAI_AUDIT["apps"]):
+            sys.exit(f"WAI_AUDIT[{k}] length does not match apps")
+    # The audited numbers must agree with the per-application table on the page.
+    page_table = [96.7, 95.0, 97.9, 98.3, 91.7, 84.0, 94.2, 95.7, 87.5, 96.7]
+    if WAI_AUDIT["auditedDx"][1:] != page_table:
+        sys.exit("WAI_AUDIT auditedDx disagrees with the page's per-application table")
+    if [b["solved"] for b in payload["twBars"]["bars"]] != [25, 24, 25, 26, 27, 28]:
+        sys.exit("twBars solved counts changed unexpectedly")
+
+    out = os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))),
+                       "assets", "data.js")
+    with open(out, "w") as f:
+        f.write("/* Generated by tools/build_data.py - do not edit by hand. */\n")
+        f.write("window.DARWINX = ")
+        json.dump(payload, f, indent=1, sort_keys=False)
+        f.write(";\n")
+
+    union = {t["id"] for t in tasks if "1" in t["bits"][:4]}
+    merged = {t["id"] for t in tasks if t["bits"][4] == "1"}
+    print(f"wrote {out}")
+    print(f"  TerminalWorld: {len(tasks)} tasks, counts={counts}")
+    print(f"    union(A..D)={len(union)}  merged={len(merged)}")
+    print(f"    merged only: {sorted(merged - union)}")
+    print(f"    union only : {sorted(union - merged)}")
+    print(f"  TB2.1 clusters: {len(payload['tb21Clusters'])}")
+    print(f"  WAI curve: {len(curve)} nodes, max={max(curve)}")
+
+
+if __name__ == "__main__":
+    main()

tools/deploy_hf.sh

@@ -0,0 +1,58 @@
+#!/usr/bin/env bash
+# Publish this directory to a Hugging Face static Space.
+#
+# Prerequisite: the Space must already exist. Create it once at
+#   https://huggingface.co/new-space   ->   SDK: "Static"
+# Then run, from anywhere:
+#   tools/deploy_hf.sh <user>/<space>
+#
+# Auth: uses $HF_TOKEN if set, otherwise prompts. Needs a token with write
+# access (https://huggingface.co/settings/tokens).
+set -euo pipefail
+
+SPACE="${1:-}"
+if [[ -z "$SPACE" || "$SPACE" != */* ]]; then
+  echo "usage: $0 <user>/<space>     e.g. $0 yifanzhang/darwinx" >&2
+  exit 2
+fi
+USER="${SPACE%%/*}"
+
+SRC="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$SRC"
+[[ -f index.html && -f README.md ]] || { echo "run this from the project page directory" >&2; exit 1; }
+
+if [[ -z "${HF_TOKEN:-}" ]]; then
+  read -rsp "Hugging Face write token (hf_...): " HF_TOKEN; echo
+fi
+[[ -n "$HF_TOKEN" ]] || { echo "no token given" >&2; exit 1; }
+
+WORK="$(mktemp -d)"
+trap 'rm -rf "$WORK"' EXIT
+echo "==> cloning https://huggingface.co/spaces/$SPACE"
+git -c credential.helper= \
+    clone --quiet "https://$USER:$HF_TOKEN@huggingface.co/spaces/$SPACE" "$WORK/space" 2>&1 \
+  | sed 's/'"$HF_TOKEN"'/***/g' \
+  || { echo "clone failed - does the Space exist, and is the token writable?" >&2; exit 1; }
+
+echo "==> copying the page"
+# Ship only what the page needs; drop local scratch files.
+rsync -a --delete \
+  --exclude '.git' --exclude '.DS_Store' --exclude '_*' --exclude '*.log' \
+  ./ "$WORK/space/"
+
+cd "$WORK/space"
+git add -A
+if git diff --cached --quiet; then
+  echo "==> nothing changed; Space is already up to date"
+  exit 0
+fi
+echo "==> files being published:"
+git diff --cached --name-status | sed 's/^/    /'
+
+git -c user.name="$USER" -c user.email="$USER@users.noreply.huggingface.co" \
+    commit --quiet -m "DarwinX project page"
+git push --quiet origin HEAD:main 2>&1 | sed 's/'"$HF_TOKEN"'/***/g'
+
+echo
+echo "==> live at https://huggingface.co/spaces/$SPACE"
+echo "    (a static Space builds in a few seconds; hard-refresh if you see the old copy)"

tools/interaction_test.html

@@ -0,0 +1,193 @@
+<!doctype html>
+<html><head><meta charset="utf-8"><title>pending</title></head>
+<body>
+<div id="host"></div>
+<script>
+/* Loads ../index.html into this document, then drives the widgets with real click
+   events and asserts the readouts. Results are written into document.title so
+   `chrome --headless --dump-dom` can report them. */
+var results = [];
+function check(name, got, want) {
+  results.push((String(got) === String(want) ? "PASS" : "FAIL") + " " + name +
+               " got=" + got + " want=" + want);
+}
+
+fetch("../index.html").then(function (r) { return r.text(); }).then(function (html) {
+  var doc = new DOMParser().parseFromString(html, "text/html");
+
+  // move body content in, then re-execute scripts (innerHTML does not run them)
+  document.getElementById("host").innerHTML = doc.body.innerHTML;
+  var scripts = Array.prototype.slice.call(document.querySelectorAll("#host script"));
+  var chain = Promise.resolve();
+  scripts.forEach(function (old) {
+    chain = chain.then(function () {
+      return new Promise(function (done) {
+        var s = document.createElement("script");
+        // getAttribute keeps the raw relative path; .src would already be resolved
+        // against this page's directory, which is one level too deep.
+        var raw = old.getAttribute("src");
+        if (raw) { s.src = "../" + raw; s.onload = done; s.onerror = done; }
+        else { s.textContent = old.textContent; }
+        document.head.appendChild(s);
+        if (!raw) { done(); }
+      });
+    });
+  });
+
+  return chain.then(function () { return new Promise(function (r2) { setTimeout(r2, 120); }); });
+}).then(function () {
+  var chips = document.querySelectorAll("#tw-chips .chip");
+  var unionVal = function () {
+    return document.querySelectorAll("#tw-grid .tcell.on").length;
+  };
+  var readout = function () { return document.getElementById("tw-readout").textContent; };
+
+  check("widgets built: chips", chips.length, 4);
+  check("grid cells", document.querySelectorAll("#tw-grid .tcell").length, 41);
+  check("all-4 union highlighted", unionVal(), 29);
+  check("readout shows 29", /29\/41/.test(readout()), true);
+  check("readout shows merge 28", /28\/41/.test(readout()), true);
+
+  // turn OFF Spec. A -> union(B,C,D) should still be 29
+  chips[0].click();
+  check("A off -> union(BCD)", unionVal(), 29);
+
+  // turn OFF B and C -> D alone = 27
+  chips[1].click(); chips[2].click();
+  check("D alone", unionVal(), 27);
+
+  // D off too -> nothing selected
+  chips[3].click();
+  check("none selected", unionVal(), 0);
+
+  // back to A + D only  -> 29 (complementary pair)
+  chips[0].click(); chips[3].click();
+  check("A+D complementary pair", unionVal(), 29);
+
+  // full set again, then compare-with-merge ring counts
+  chips[1].click(); chips[2].click();
+  check("back to all four", unionVal(), 29);
+  document.getElementById("tw-cmp").click();
+  check("merge-only ring (gain)", document.querySelectorAll("#tw-grid .tcell.gain").length, 1);
+  check("union-only ring (loss)", document.querySelectorAll("#tw-grid .tcell.loss").length, 2);
+
+  // cluster widget
+  var bars = document.querySelectorAll("#cl-bars .cbar");
+  check("cluster rows", bars.length, 7);
+  var firstDelta = bars[0].querySelector(".cd").textContent.trim();
+  check("default sort = biggest gain first", firstDelta, "+14.8");
+  var sortChips = document.querySelectorAll("#cl-chips .chip");
+  sortChips[1].click(); // by base rate ascending
+  var b2 = document.querySelectorAll("#cl-bars .cbar");
+  check("sort by base rate -> lowest base first",
+        /ML & scientific/.test(b2[0].querySelector(".cname").textContent), true);
+  sortChips[2].click(); // by cluster size
+  var b3 = document.querySelectorAll("#cl-bars .cbar");
+  check("sort by size -> n=23 first",
+        /n=23/.test(b3[0].querySelector(".cname").textContent), true);
+  document.getElementById("cl-base").click(); // hide base
+  check("hide base removes grey fills", document.querySelectorAll("#cl-bars .fill.b").length, 0);
+
+  // curve widget
+  check("curve dots", document.querySelectorAll("#cv-chart .dot").length, 37);
+  check("best-so-far polyline", document.querySelectorAll("#cv-chart .best").length, 1);
+  document.getElementById("cv-all").click();
+  check("hide variants", document.querySelectorAll("#cv-chart .dot").length, 0);
+  document.getElementById("cv-best").click();
+  check("hide envelope", document.querySelectorAll("#cv-chart .best").length, 0);
+
+  // sortable table: WAI overall row must stay pinned last
+  var wai = document.querySelectorAll("table.sortable")[2];
+  var ths = wai.querySelectorAll("thead th");
+  ths[6].click(); // sort by Monet (DarwinX)
+  var rows = wai.querySelectorAll("tbody tr");
+  check("total row stays last after sort",
+        rows[rows.length - 1].classList.contains("total"), true);
+  var firstCell = rows[0].cells[6].textContent.trim();
+  check("descending sort puts 98.3 first", firstCell, "98.3");
+  ths[6].click(); // ascending
+  var rows2 = wai.querySelectorAll("tbody tr");
+  check("ascending sort puts 84.0 first", rows2[0].cells[6].textContent.trim(), "84.0");
+  check("total still last", rows2[rows2.length - 1].classList.contains("total"), true);
+
+  // ---- headline panels ----
+  check("headline panels", document.querySelectorAll("#hl-panels .panel").length, 4);
+  check("prior bars present", document.querySelectorAll("#hl-panels .bar.p").length, 3);
+  var wa = document.querySelectorAll("#hl-panels .panel")[2];
+  var hOwn = parseFloat(wa.querySelector(".bar.e").getAttribute("height"));
+  document.querySelectorAll("#hl-chips .chip")[1].click(); // shared 0-100
+  var wa2 = document.querySelectorAll("#hl-panels .panel")[2];
+  var hShared = parseFloat(wa2.querySelector(".bar.e").getAttribute("height"));
+  check("shared axis changes bar geometry", hShared !== hOwn, true);
+  var tb = document.querySelectorAll("#hl-panels .panel")[0];
+  var tbH = parseFloat(tb.querySelector(".bar.e").getAttribute("height"));
+  check("shared axis: TB2.1 bar shorter than WAI bar", tbH < hShared, true);
+  document.getElementById("hl-prior").click();
+  check("hide prior removes its bars", document.querySelectorAll("#hl-panels .bar.p").length, 0);
+  document.querySelectorAll("#hl-chips .chip")[0].click(); // back to per-panel
+
+  // ---- TerminalWorld specialist bars ----
+  var spRows = document.querySelectorAll("#sp-rows .hrow");
+  check("specialist rows (6 bars + ref tag)", spRows.length, 7);
+  check("percent shown by default", /68\.3%/.test(document.getElementById("sp-rows").textContent), true);
+  var spW = function (n) {
+    return parseFloat(document.querySelectorAll("#sp-rows .hf")[n].style.width);
+  };
+  var zoomSpread = spW(5) - spW(1);   // merged minus weakest specialist
+  document.querySelectorAll("#sp-chips .chip")[1].click();   // full 0-100 axis
+  var fullSpread = spW(5) - spW(1);
+  check("zoomed axis magnifies the spread", zoomSpread > fullSpread * 3, true);
+  check("full axis puts merged at its true 68.3%", Math.round(spW(5) * 10) / 10, 68.3);
+  document.querySelectorAll("#sp-chips .chip")[0].click();
+  document.getElementById("sp-counts").click();
+  check("counts toggle shows 28/41", /28\/41/.test(document.getElementById("sp-rows").textContent), true);
+
+  // ---- TB2.1 compute ----
+  check("compute rows (2 groups x 2 series)", document.querySelectorAll("#cp-rows .hrow").length, 4);
+  var cpVals = function () {
+    return Array.prototype.map.call(document.querySelectorAll("#cp-rows .hv"),
+      function (e) { return e.textContent.trim(); }).join(",");
+  };
+  check("turns shown first", cpVals(), "11,22,12,13");
+  document.querySelectorAll("#cp-chips .chip")[1].click();
+  check("tokens toggle shows K values", cpVals(), "89K,380K,125K,172K");
+
+  // ---- WAI invalid composition ----
+  check("two stacked rows", document.querySelectorAll("#iv-rows .stack").length, 2);
+  var mechSegs = document.querySelectorAll("#iv-rows .stack i").length;
+  check("mechanism view: 4 before + 1 after segments", mechSegs, 5);
+  check("totals rendered 293 and 17",
+        /293/.test(document.getElementById("iv-rows").textContent) &&
+        /17/.test(document.getElementById("iv-rows").textContent), true);
+  var afterStack = document.querySelectorAll("#iv-rows .stack")[1];
+  var wShared = afterStack.querySelector("i").style.width;
+  document.getElementById("iv-norm").click();
+  var wNorm = document.querySelectorAll("#iv-rows .stack")[1].querySelector("i").style.width;
+  check("normalising widens the after row", parseFloat(wNorm) > parseFloat(wShared), true);
+  check("after row fills the width when normalised", parseFloat(wNorm), 100);
+  document.querySelectorAll("#iv-chips .chip")[0].click(); // by application
+  check("application view segment count",
+        document.querySelectorAll("#iv-rows .stack i").length, 9 + 5);
+
+  // ---- WAI audit dumbbell ----
+  check("audit lanes (11 apps x 2)", document.querySelectorAll("#ad-rows .seg").length, 22);
+  check("axis row precedes the data rows",
+        /pass@1/.test(document.querySelectorAll("#ad-rows .drow")[0].textContent), true);
+  check("overall row pinned first",
+        document.querySelectorAll("#ad-rows .drow")[1].textContent.indexOf("Overall"), 0);
+  document.getElementById("ad-base").click();
+  check("hiding base halves the lanes", document.querySelectorAll("#ad-rows .seg").length, 11);
+  document.getElementById("ad-base").click();
+  document.getElementById("ad-sort").click();
+  var firstAfterOverall = document.querySelectorAll("#ad-rows .drow")[2].querySelector(".dn").textContent;
+  // Superhuman loses the most to the audit: 53.3 -> 31.7 = 21.6, ahead of Gmail's 20.0.
+  check("sort by audit loss puts Superhuman first", firstAfterOverall, "Superhuman");
+
+  var fails = results.filter(function (r) { return r.indexOf("FAIL") === 0; });
+  document.title = "RESULT " + (fails.length ? "FAILURES=" + fails.length : "ALL PASS") +
+                   " || " + results.join(" || ");
+}).catch(function (e) {
+  document.title = "RESULT ERROR " + e.message + " || " + results.join(" || ");
+});
+</script>
+</body></html>