npm - thumbgate - Versions diffs - 1.29.2 → 1.31.0 - Mend

thumbgate 1.29.2 → 1.31.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Files changed (111) hide show

package/.claude-plugin/plugin.json +1 -1
package/.well-known/mcp/server-card.json +1 -1
package/README.md +54 -16
package/adapters/claude/.mcp.json +2 -2
package/adapters/forge/forge.yaml +3 -3
package/adapters/mcp/server-stdio.js +143 -14
package/adapters/opencode/opencode.json +1 -1
package/bench/observability-eval-suite.json +2 -2
package/bin/cli.js +154 -36
package/config/evals/generation-quality-golden.json +95 -0
package/config/evals/rag-answer-quality-golden.json +91 -0
package/config/evals/retrieval-hybrid-ablation.json +66 -0
package/config/evals/retrieval-ranking-golden.json +522 -0
package/config/gates/default.json +217 -50
package/config/mcp-allowlists.json +215 -185
package/config/model-tiers.json +7 -2
package/config/post-deploy-marketing-pages.json +26 -1
package/glama.json +6 -0
package/package.json +94 -11
package/public/architecture.html +130 -0
package/public/assets/diagrams/agent-integration.png +0 -0
package/public/assets/diagrams/before-after.svg +22 -0
package/public/assets/diagrams/decision.svg +36 -0
package/public/assets/diagrams/feedback-pipeline.png +0 -0
package/public/assets/diagrams/hero-thumbs.svg +68 -0
package/public/assets/diagrams/loop.svg +40 -0
package/public/assets/diagrams/plugin-topology.png +0 -0
package/public/assets/diagrams/pre-action-gate-loop.svg +59 -0
package/public/assets/diagrams/self-improving-thumbs-loop.svg +105 -0
package/public/assets/diagrams/stack.svg +18 -0
package/public/assets/diagrams/thumbgate-architecture.png +0 -0
package/public/case-studies.html +151 -0
package/public/compare.html +1 -0
package/public/dashboard.html +126 -28
package/public/eval-scorecard.html +195 -0
package/public/eval-scorecard.json +18 -0
package/public/evaluations.html +168 -0
package/public/index.html +143 -13
package/public/numbers.html +3 -2
package/public/pricing.html +143 -30
package/public/whitepaper.html +189 -0
package/scripts/a-plus-evidence-scorecard.js +303 -0
package/scripts/activation-quickstart.js +1 -0
package/scripts/agent-outcome-monitor.js +71 -1
package/scripts/async-eval-observability.js +36 -11
package/scripts/audit-trail.js +37 -1
package/scripts/auto-promote-gates.js +149 -34
package/scripts/billing.js +3 -1
package/scripts/claude-feedback-sync.js +3 -2
package/scripts/cli-feedback.js +13 -7
package/scripts/colbert-style-maxsim.js +236 -0
package/scripts/cross-encoder-reranker.js +359 -126
package/scripts/dashboard-chat.js +350 -17
package/scripts/document-intake.js +283 -7
package/scripts/eval-quality-suite.js +204 -0
package/scripts/feedback-aggregate.js +5 -2
package/scripts/feedback-loop.js +359 -189
package/scripts/feedback-paths.js +32 -13
package/scripts/feedback-quality.js +53 -0
package/scripts/filesystem-search.js +17 -7
package/scripts/gates-engine.js +98 -4
package/scripts/generate-case-study-outreach.js +253 -0
package/scripts/generate-eval-scorecard.js +276 -0
package/scripts/growth-campaigns.js +183 -0
package/scripts/harness-tool-names.js +70 -0
package/scripts/hook-runtime.js +10 -3
package/scripts/jsonl-watcher.js +1 -0
package/scripts/lesson-db.js +16 -5
package/scripts/lesson-embedding-index.js +67 -20
package/scripts/lesson-embedding-maintenance.js +177 -0
package/scripts/lesson-inference.js +23 -4
package/scripts/lesson-reranker.js +55 -9
package/scripts/lesson-retrieval.js +375 -32
package/scripts/lesson-search.js +48 -11
package/scripts/llm-client.js +304 -15
package/scripts/mcp-config.js +26 -5
package/scripts/mcp-oauth.js +37 -2
package/scripts/model-eval.js +308 -0
package/scripts/model-tier-router.js +593 -0
package/scripts/parallel-workflow-orchestrator.js +86 -22
package/scripts/pragmatic-hybrid-search.js +379 -0
package/scripts/published-cli.js +11 -1
package/scripts/rag-document-pipeline.js +461 -0
package/scripts/rag-structured-output.js +441 -0
package/scripts/ragas-style-metrics.js +351 -0
package/scripts/refresh-proof-pack.js +261 -0
package/scripts/request-envelope.js +178 -0
package/scripts/rerank-pipeline.js +370 -0
package/scripts/rerank-quality-eval.js +155 -0
package/scripts/retrieval-hybrid-ablation.js +120 -0
package/scripts/retrieval-quality-tier.js +118 -0
package/scripts/risk-scorer.js +144 -15
package/scripts/secret-scanner.js +395 -4
package/scripts/self-distill-agent.js +7 -1
package/scripts/self-healing-check.js +25 -0
package/scripts/skill-packs.js +183 -0
package/scripts/slow-loop.js +72 -0
package/scripts/statusline-links.js +1 -1
package/scripts/statusline-local-stats.js +1 -1
package/scripts/statusline.sh +8 -1
package/scripts/telemetry-analytics.js +13 -1
package/scripts/thumbgate-bench.js +13 -0
package/scripts/thumbgate-search.js +98 -6
package/scripts/tier-budget-guard.js +186 -0
package/scripts/tool-kpi-tracker.js +124 -0
package/scripts/tool-registry.js +95 -1
package/scripts/vector-store.js +108 -4
package/scripts/verify-marketing-pages-deployed.js +85 -3
package/server.json +44 -0
package/smithery.yaml +17 -0
package/src/api/server.js +424 -99

package/public/evaluations.html ADDED Viewed

@@ -0,0 +1,168 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+<meta charset="UTF-8">
+<meta name="viewport" content="width=device-width, initial-scale=1.0">
+<meta name="generator" content="ThumbGate">
+<meta name="author" content="Igor Ganapolsky">
+<title>ThumbGate — How We Evaluate | Held-Out Metrics, Drift Gates, Error Analysis</title>
+<meta name="description" content="ThumbGate's evaluation methodology, published including the numbers that don't flatter us: held-out lift over trivial baselines, dual IID/distribution-shift splits, trace-mined drift benchmarks, and the five evaluator bugs we found and fixed.">
+<meta property="og:title" content="ThumbGate — How We Evaluate">
+<meta property="og:description" content="Evaluation methodology with the unflattering numbers left in: +9.9 points held-out lift on familiar traffic, −10.5 on unseen action types. That asymmetry is why deterministic rules enforce and the model only triages.">
+<meta property="og:type" content="article">
+<meta property="og:url" content="https://thumbgate.ai/evaluations">
+<meta name="twitter:card" content="summary_large_image">
+<link rel="canonical" href="https://thumbgate.ai/evaluations">
+<link rel="icon" type="image/png" href="/thumbgate-icon.png">
+<link rel="apple-touch-icon" href="/apple-touch-icon.png">
+<script defer data-domain="thumbgate.ai" src="https://plausible.io/js/script.js"></script>
+<script type="application/ld+json">
+{
+  "@context": "https://schema.org",
+  "@type": "TechArticle",
+  "headline": "How ThumbGate Evaluates Its Own Enforcement",
+  "description": "Held-out evaluation methodology for an AI-agent firewall: dual IID/distribution-shift splits, lift over trivial baselines, trace-mined drift benchmarks, published-artifact verification, and the error analysis behind them.",
+  "url": "https://thumbgate.ai/evaluations",
+  "dateModified": "2026-07-29",
+  "author": {
+    "@type": "Person",
+    "name": "Igor Ganapolsky",
+    "url": "https://github.com/IgorGanapolsky",
+    "sameAs": [
+      "https://github.com/IgorGanapolsky",
+      "https://www.linkedin.com/in/igorganapolsky"
+    ]
+  }
+}
+</script>
+<style>
+  :root {
+    --bg: #0b0f14; --panel: #111823; --border: #1e2a3a; --text: #e6edf3;
+    --muted: #8b98a5; --cyan: #39c5cf; --green: #3fb950; --red: #f85149; --amber: #d29922;
+  }
+  * { margin: 0; padding: 0; box-sizing: border-box; }
+  body { background: var(--bg); color: var(--text); font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, sans-serif; line-height: 1.6; }
+  nav { padding: 1rem 2rem; border-bottom: 1px solid var(--border); display: flex; gap: 1.5rem; align-items: center; flex-wrap: wrap; }
+  nav a { color: var(--muted); text-decoration: none; font-size: 0.9rem; }
+  nav a:hover { color: var(--cyan); }
+  nav .brand { color: var(--text); font-weight: 700; font-size: 1.05rem; text-decoration: none; }
+  .container { max-width: 880px; margin: 0 auto; padding: 2.5rem 1.5rem 4rem; }
+  h1 { font-size: 2rem; margin-bottom: 0.5rem; }
+  h2 { font-size: 1.35rem; margin: 2.5rem 0 0.75rem; color: var(--cyan); }
+  h3 { font-size: 1.05rem; margin: 1.5rem 0 0.5rem; }
+  p { margin: 0.75rem 0; color: var(--text); }
+  .subtitle { color: var(--muted); font-size: 1.05rem; }
+  .panel { background: var(--panel); border: 1px solid var(--border); border-radius: 10px; padding: 1.25rem 1.5rem; margin: 1.25rem 0; }
+  table { width: 100%; border-collapse: collapse; margin: 1rem 0; font-size: 0.92rem; }
+  th, td { text-align: left; padding: 0.5rem 0.75rem; border-bottom: 1px solid var(--border); }
+  th { color: var(--muted); font-weight: 600; }
+  code { background: #0d1420; border: 1px solid var(--border); border-radius: 4px; padding: 0.1rem 0.35rem; font-size: 0.86em; color: var(--cyan); }
+  pre { background: #0d1420; border: 1px solid var(--border); border-radius: 8px; padding: 1rem; overflow-x: auto; font-size: 0.85rem; margin: 1rem 0; }
+  .good { color: var(--green); } .bad { color: var(--red); } .warn { color: var(--amber); }
+  .chart-caption { color: var(--muted); font-size: 0.85rem; margin-top: 0.4rem; }
+  ul { margin: 0.5rem 0 0.75rem 1.4rem; }
+  li { margin: 0.35rem 0; }
+  .footer-note { color: var(--muted); font-size: 0.85rem; margin-top: 2.5rem; text-align: center; }
+  .footer-note a { color: var(--cyan); text-decoration: none; }
+  .rule { border-left: 3px solid var(--cyan); padding: 0.6rem 1rem; background: var(--panel); margin: 1.25rem 0; font-weight: 600; }
+</style>
+</head>
+<body>
+<nav>
+  <a class="brand" href="/">ThumbGate</a>
+  <a href="/whitepaper">White paper</a>
+  <a href="/architecture">Architecture</a>
+  <a href="/eval-scorecard">Scorecard</a>
+  <a href="/evaluations">Evaluations</a>
+  <a href="/case-studies">Case studies</a>
+  <a href="/numbers">Numbers</a>
+  <a href="/pricing">Pricing</a>
+</nav>
+<main class="container">
+  <h1>How We Evaluate</h1>
+  <p class="subtitle">The methodology behind ThumbGate's enforcement quality — published with the numbers that don't flatter us left in. Every figure on this page is reproducible from the open-source repo with one command.</p>
+  <h2>The headline we retired</h2>
+  <p>ThumbGate's learned risk model used to report one quality figure: <code>trainingAccuracy: 0.820</code> — measured on the same rows it trained on, against a corpus where answering "risky" unconditionally scores 71.1%. That is an 11-point in-sample lift presented as an 82-point achievement, and it said nothing about generalization at all.</p>
+  <p>We replaced it with held-out evaluation under two different splits, and we publish both results — including the one that inverts.</p>
+  <h2>The result, honestly</h2>
+  <div class="panel">
+    <svg viewBox="0 0 640 260" role="img" aria-label="Bar chart: accuracy lift over the majority baseline. In-sample plus 10.9 points, held-out IID plus 9.9 points, held-out novel-context minus 10.5 points.">
+      <line x1="60" y1="130" x2="620" y2="130" stroke="#1e2a3a" stroke-width="1"/>
+      <text x="30" y="40" fill="#8b98a5" font-size="11">+12</text>
+      <text x="30" y="134" fill="#8b98a5" font-size="11">0</text>
+      <text x="30" y="230" fill="#8b98a5" font-size="11">−12</text>
+      <rect x="90"  y="48"  width="120" height="82" fill="#d29922" rx="3"/>
+      <rect x="270" y="55"  width="120" height="75" fill="#3fb950" rx="3"/>
+      <rect x="450" y="130" width="120" height="79" fill="#f85149" rx="3"/>
+      <text x="150" y="40"  fill="#d29922" font-size="13" text-anchor="middle">+10.9</text>
+      <text x="330" y="47"  fill="#3fb950" font-size="13" text-anchor="middle">+9.9</text>
+      <text x="510" y="225" fill="#f85149" font-size="13" text-anchor="middle">−10.5</text>
+      <text x="150" y="250" fill="#8b98a5" font-size="11" text-anchor="middle">in-sample (old headline)</text>
+      <text x="330" y="250" fill="#8b98a5" font-size="11" text-anchor="middle">held-out, familiar kinds</text>
+      <text x="510" y="250" fill="#8b98a5" font-size="11" text-anchor="middle">held-out, unseen kinds</text>
+    </svg>
+    <p class="chart-caption">Accuracy lift over the majority-class baseline, in points. 1,791 production rows, 12 independent group-aware stratified splits, mean across resamples. IID: +9.9 ± 1.7, AUC 0.887, 12/12 folds beat baseline. Novel-context: −10.5 ± 9.0, 1/12 folds, calibration error 0.31 vs 0.06.</p>
+  </div>
+  <p>On familiar traffic the model genuinely helps. On action types it has never seen, it is <span class="bad">worse than a constant classifier</span> — and unfamiliar actions are precisely the case a firewall exists for. That asymmetry is why <strong>deterministic rules enforce and the model only triages</strong>. We consider publishing that inversion more valuable than hiding it: it is the reason ThumbGate's blocking behaviour does not degrade with the model.</p>
+  <h2>The methodology</h2>
+  <table>
+    <tr><th>Practice</th><th>What it prevents</th></tr>
+    <tr><td>Lift vs the majority baseline, reported beside every accuracy</td><td>71% base-rate corpora making any model look strong</td></tr>
+    <tr><td>MCC + ROC-AUC + Brier + calibration error</td><td>accuracy hiding degenerate or miscalibrated behaviour</td></tr>
+    <tr><td>Dual holdouts: IID <em>and</em> whole-category (distribution shift)</td><td>quoting only the flattering generalization estimate</td></tr>
+    <tr><td>Group-aware splits keyed on the feature vector</td><td>duplicate inputs straddling folds and inflating scores</td></tr>
+    <tr><td>12 resamples, mean ± sd, paired t on shared folds</td><td>single lucky splits shipping as "improvements"</td></tr>
+    <tr><td>Per-fold vocabulary rebuild</td><td>held-out rows choosing the model's features (transductive fit)</td></tr>
+  </table>
+  <h2>The evaluation suites</h2>
+  <ul>
+    <li><strong>Model harness</strong> — <code>npm run eval:risk</code>: the resampling harness; <code>--json</code> emits the machine-readable report backing this page.</li>
+    <li><strong>Trace-mined drift benchmark</strong> — real production gate decisions, redacted, become 60 replayable cases across 12 gates; CI fails when any real command's verdict <em>moves</em>. Silent gate death is the signature of every bypass we have found.</li>
+    <li><strong>Evasion matrix</strong> — 14 dangerous commands × 9 canonicalization transforms; this is the suite that measured the 62 holes closed in v1.29.2.</li>
+    <li><strong>Published-artifact verification</strong> — the npm tarball itself is installed into a throwaway prefix and driven through its public hook contract twice daily. CI proving the <em>source</em> blocks something is a different claim from the artifact users receive blocking it.</li>
+    <li><strong>Decision canary</strong> — live decision-distribution drift: a gate going <em>silent</em>, a <em>spike</em> in blocks, or a gate appearing/disappearing.</li>
+  </ul>
+  <h2>The gates that block a release</h2>
+  <ul>
+    <li>A synthetic-fixture quality gate that separates <em>"the trainer broke"</em> from <em>"the data lacks signal"</em> — different problems, different fixes.</li>
+    <li>A <strong>pure-noise canary</strong>: a trainer that finds &gt;15 points of lift in random labels has labels leaking into features, and CI fails.</li>
+    <li><strong>Vacuity guards</strong>: a drift test that compared fewer than 20 cases fails; a single-class test fold refuses to report rather than emitting a meaningless pass. A check that cannot fail is treated as no check.</li>
+    <li>Non-significant improvements do not ship. Two candidate upgrades (a feature-diversity cap, Platt calibration) are implemented, tested, and <em>off</em> — paired t of −0.31 and −0.72.</li>
+  </ul>
+  <h2>Error analysis: the five bugs we found in our own evaluator</h2>
+  <p>The uncomfortable finding of this work: the evaluator was wrong more often than the model, and its errors all flattered us.</p>
+  <table>
+    <tr><th>Bug</th><th>Effect before the fix</th></tr>
+    <tr><td>Tied-hash rows straddling folds</td><td>test fold contained memorized rows</td></tr>
+    <tr><td>Cross-class duplicate leakage under label noise</td><td>same input in both folds with different labels</td></tr>
+    <tr><td>Group quota overshoot (one 655-row category)</td><td class="bad">novel-context result read +0.016; truth is −0.105</td></tr>
+    <tr><td>Vocabulary built before the split</td><td>held-out rows chose the model's features</td></tr>
+    <tr><td>Raw text in the split key</td><td>identical model inputs split across folds</td></tr>
+  </table>
+  <p>Three of the five were found by adversarial review <em>after</em> we had published numbers from the leaky evaluator. The corrected conclusion was stronger, not weaker — and the correction is documented in the repo, not overwritten.</p>
+  <div class="rule">The rule all of this enforces: no accuracy figure ships without the baseline it beats and the split it was measured on.</div>
+  <h2>Reproduce it</h2>
+  <p>Honest scoping: the measured corpus is <strong>private operator feedback telemetry</strong> and does not ship in the public repo. The exact machine-readable report behind the figures on this page is checked in with provenance at <a href="https://github.com/IgorGanapolsky/ThumbGate/blob/main/evals/risk-model-report.json">evals/risk-model-report.json</a>. A clean clone gives you that report, the methodology write-up, the CI quality gate (synthetic fixtures / no private data), and the same harness to run against <em>your</em> feedback log. The figures on this page are first-party operator measurements documented in <a style="color:var(--cyan)" href="https://github.com/IgorGanapolsky/ThumbGate/blob/main/docs/ML-EVALUATION.md">docs/ML-EVALUATION.md</a> — not a committed private corpus dump.</p>
+  <pre>git clone https://github.com/IgorGanapolsky/ThumbGate
+cd ThumbGate &amp;&amp; npm ci
+npm run test:risk-quality              # CI quality gate — no private data needed
+# Optional: re-run the harness on your own feedback log
+node scripts/eval-risk-model.js --corpus &lt;your-own-feedback-log&gt; --resamples 12 --json</pre>
+  <p>Full methodology: <a style="color:var(--cyan)" href="https://github.com/IgorGanapolsky/ThumbGate/blob/main/docs/ML-EVALUATION.md">docs/ML-EVALUATION.md</a> · case studies: <a style="color:var(--cyan)" href="/case-studies">/case-studies</a> · gate scorecard: <a style="color:var(--cyan)" href="/eval-scorecard">/eval-scorecard</a></p>
+  <p class="footer-note">ThumbGate is built in the open by <a href="https://github.com/IgorGanapolsky">Igor Ganapolsky</a>. This page reports first-party measurements only — no customer counts, no install volume, no revenue claims.</p>
+</main>
+</body>
+</html>

package/public/index.html CHANGED Viewed

@@ -5,7 +5,7 @@
   <meta name="viewport" content="width=device-width, initial-scale=1.0">
   <meta name="generator" content="ThumbGate">
   <meta name="author" content="Igor Ganapolsky">
-  <meta name="thumbgate-version" content="1.29.2">
+  <meta name="thumbgate-version" content="1.31.0">
   __GOOGLE_SITE_VERIFICATION_META__
   <link rel="icon" type="image/png" href="/thumbgate-icon.png">
   <link rel="canonical" href="__APP_ORIGIN__/">
@@ -39,7 +39,7 @@
     "applicationCategory": "DeveloperApplication",
     "operatingSystem": "macOS, Linux, Windows",
     "url": "__APP_ORIGIN__",
-    "description": "Self-improving firewall for AI coding agents. ThumbGate captures feedback as local lessons, re-ranks relevant lessons for each action, promotes repeated negative patterns into warning or blocking gates, expires stale auto-promoted gates, and runs pre-action checks before shell, file, git, deploy, or API tool calls. Detected secret exfiltration and gate-bypass commands are denied by default; strict mode blocks matching destructive-action rules.",
+    "description": "Self-improving firewall for AI coding agents. ThumbGate captures feedback as local lessons, re-ranks relevant lessons for each action, promotes repeated negative patterns into warning or blocking gates, expires stale auto-promoted gates, and runs pre-action checks before shell, file, git, deploy, or API tool calls. Detected secret exfiltration and process-kill/environment-override self-disable commands (gate-process bypass) are denied by default; strict mode blocks matching destructive-action rules.",
     "codeRepository": "https://github.com/IgorGanapolsky/ThumbGate",
     "featureList": [
       "Pre-action PreToolUse checks for coding agents",
@@ -559,6 +559,104 @@
       h1 { font-size: 46px; }
       .checkout-card, .pro-card { padding: 22px 18px; }
     }
+    /* Visual-first conversion: thumbs + diagrams */
+    .hero-thumbs {
+      display: flex;
+      align-items: center;
+      justify-content: flex-start;
+      gap: 14px;
+      margin: 0 0 18px;
+      font-size: clamp(42px, 6vw, 64px);
+      line-height: 1;
+      filter: drop-shadow(0 0 28px rgba(34, 211, 238, 0.28));
+    }
+    .hero-thumbs span {
+      display: inline-flex;
+      width: 1.15em;
+      height: 1.15em;
+      align-items: center;
+      justify-content: center;
+      border-radius: 22%;
+      background: rgba(17, 23, 34, 0.9);
+      border: 1px solid rgba(39, 50, 70, 0.9);
+    }
+    .hero-thumbs .up { border-color: rgba(86, 227, 159, 0.55); box-shadow: 0 0 0 1px rgba(86, 227, 159, 0.12); }
+    .hero-thumbs .down { border-color: rgba(255, 100, 124, 0.55); box-shadow: 0 0 0 1px rgba(255, 100, 124, 0.12); }
+    .hero-visual {
+      margin: 22px 0 0;
+      border: 1px solid rgba(56, 215, 255, 0.22);
+      border-radius: 18px;
+      overflow: hidden;
+      background: #0a1018;
+      box-shadow: 0 24px 70px rgba(0, 0, 0, 0.35);
+    }
+    .hero-visual img, .diagram-frame img {
+      display: block;
+      width: 100%;
+      height: auto;
+    }
+    .diagram-strip { padding: 56px 0 12px; }
+    .diagram-frame {
+      margin: 28px 0 0;
+      border: 1px solid rgba(56, 215, 255, 0.2);
+      border-radius: 20px;
+      overflow-x: auto;
+      overflow-y: hidden;
+      -webkit-overflow-scrolling: touch;
+      background: #0a0a0b;
+      box-shadow: 0 20px 60px rgba(0, 0, 0, 0.3);
+    }
+    .diagram-frame img {
+      min-width: 640px;
+    }
+    .hero-visual img {
+      min-width: 0;
+    }
+    @media (max-width: 720px) {
+      .diagram-frame img {
+        min-width: 720px;
+      }
+      .hero-visual {
+        margin-left: -4px;
+        margin-right: -4px;
+      }
+    }
+    .diagram-caption {
+      margin: 14px 0 0;
+      color: var(--muted);
+      font-size: 14px;
+      text-align: center;
+    }
+    .self-improve-callout {
+      margin: 28px 0 0;
+      padding: 20px 22px;
+      border: 1px solid rgba(86, 227, 159, 0.3);
+      border-radius: 14px;
+      background: rgba(86, 227, 159, 0.06);
+      color: #c5ceda;
+      font-size: 16px;
+      line-height: 1.55;
+      max-width: 820px;
+    }
+    .self-improve-callout strong { color: var(--green); }
+    .loop-step .thumb-icon {
+      display: inline-flex;
+      font-size: 28px;
+      line-height: 1;
+      margin-bottom: 4px;
+    }
+    .visual-pair {
+      display: grid;
+      grid-template-columns: 1.1fr 0.9fr;
+      gap: 18px;
+      margin-top: 28px;
+      align-items: stretch;
+    }
+    @media (max-width: 900px) {
+      .visual-pair { grid-template-columns: 1fr; }
+      .hero-thumbs { justify-content: center; }
+    }
   </style>
 </head>
 <body data-revenue-assist="off">
@@ -584,6 +682,7 @@
       <div class="shell hero-grid">
         <div class="hero-copy">
           <div class="eyebrow">Now live · Pre-action gates · Not a prompt · Not a postmortem</div>
+          <div class="hero-thumbs" aria-hidden="true"><span class="up">👍</span><span class="down">👎</span></div>
           <h1>Stop AI agent mistakes before they cost you.</h1>
           <p class="hero-lede">Hard allow/deny at the <strong>tool-call boundary</strong>. Audit entry written at decision time. Every approval teaches the next gate.</p>
           <p class="fit-line"><strong>Dual path (buy + book):</strong> <strong>$499 Diagnostic</strong> maps one expensive failure on Claude Code, Cursor, Codex, or similar agents and installs a hard gate with proof — or self-serve <strong>Pro at $19/mo</strong> if you already know the loop. Free local evaluate stays free.</p>
@@ -593,7 +692,11 @@
             <span class="spec-chip"><strong>Promote</strong> gates</span>
             <span class="spec-chip"><strong>Block</strong> next action</span>
           </div>
+          <figure class="hero-visual">
+            <img src="/assets/diagrams/hero-thumbs.svg" width="640" height="360" alt="ThumbGate: thumbs up and thumbs down become a pre-action gate that can ALLOW, WARN, or DENY before the tool call runs" loading="eager" decoding="async">
+          </figure>
           <a class="proof-link" href="#proof">See a strict-mode deny example ↓</a>
+          <a class="proof-link-secondary" href="#how-it-works" style="margin-left:14px;color:var(--cyan);font-weight:750;text-decoration:none">See how the self-improving loop works ↓</a>
           <p class="install-hint"><code>npx thumbgate init</code> free local evaluate · first hard gate usually minutes after install · <a href="/pricing">how we stack up</a></p>
           <p class="install-hint" id="thumbgate-dashboard-command" style="margin-top:10px;">Dashboard: <code>npx thumbgate dashboard --open</code> · <code>/thumbgate-dashboard</code> · bin <code>thumbgate-dashboard</code> if global · <a href="/dashboard#insights">demo</a></p>
         </div>
@@ -638,34 +741,60 @@
       </div>
     </section>
+    <section class="diagram-strip" id="before-after" aria-label="Before and after ThumbGate">
+      <div class="shell">
+        <div class="section-kicker">Why it’s called ThumbGate</div>
+        <h2 class="section-title">Thumbs teach. The gate enforces.</h2>
+        <div class="visual-pair">
+          <figure class="diagram-frame">
+            <img src="/assets/diagrams/before-after.svg" width="960" height="300" alt="Without ThumbGate the same AI agent mistake repeats. With ThumbGate one thumbs-down becomes a local pre-action gate that can refresh or expire." loading="lazy" decoding="async">
+          </figure>
+          <figure class="diagram-frame">
+            <img src="/assets/diagrams/loop.svg" width="960" height="300" alt="Agent proposes a tool, ThumbGate decides ALLOW WARN or DENY, thumbs feedback becomes a rule that closes the loop." loading="lazy" decoding="async">
+          </figure>
+        </div>
+      </div>
+    </section>
     <section id="how-it-works">
       <div class="shell">
         <div class="section-kicker">Self-improving product loop</div>
         <h2 class="section-title">Pre-action checks—and the systems that refine them.</h2>
-        <p class="section-lede">No governance novel. Corrections become reviewable local lessons; relevant lessons are re-ranked; repeated negative patterns can become explicit gates; and the next action is checked before execution. <strong style="color:var(--text)">Click a step</strong> to see what happens under the hood.</p>
+        <p class="section-lede">Corrections become local lessons; repeated negatives become gates; the next action is checked before execution. <strong style="color:var(--text)">Click a step</strong> for the under-the-hood demo.</p>
+        <figure class="diagram-frame" style="margin-top:28px">
+          <img src="/assets/diagrams/self-improving-thumbs-loop.svg" width="1100" height="420" alt="Self-improving ThumbGate loop: thumbs up or thumbs down become a local lesson, then a prevention rule, then ALLOW WARN or DENY on the next tool call. The loop closes under your control without retraining the model." loading="lazy" decoding="async">
+        </figure>
+        <div class="self-improve-callout">
+          <strong>Is it really self-improving?</strong>
+          Yes — the control layer, not the LLM. 👍/👎 → local lesson → ranked rule → next tool call gated. No model retrain.
+        </div>
         <div class="loop" role="tablist" aria-label="Self-improving product loop — click a step for under-the-hood demo">
           <!-- Cards are divs (not <button>) so headings/chips stay valid HTML; browsers
                auto-close <button> around <h3>/<div>, which broke click targets. -->
           <div class="loop-step" role="tab" tabindex="0" id="loop-tab-1" data-loop-step="1" aria-selected="false" aria-controls="loop-panel" aria-expanded="false">
             <span class="loop-number">1</span>
+            <div class="thumb-icon" aria-hidden="true">👍👎</div>
             <h3>Capture feedback</h3>
-            <p>Record explicit 👍 or 👎 feedback with the action and outcome context that made it useful or wrong.</p>
+            <p>Record explicit 👍 or 👎 with the action and outcome context.</p>
             <span class="loop-hint">Under the hood ↓</span>
           </div>
           <div class="loop-step" role="tab" tabindex="0" id="loop-tab-2" data-loop-step="2" aria-selected="false" aria-controls="loop-panel" aria-expanded="false">
             <span class="loop-number">2</span>
+            <div class="thumb-icon" aria-hidden="true">🧠</div>
             <h3>Remember locally</h3>
-            <p>Store a reviewable lesson that survives sessions and model changes without sending the control history into model weights.</p>
+            <p>Store a reviewable lesson that survives sessions without writing into model weights.</p>
             <span class="loop-hint">Under the hood ↓</span>
           </div>
           <div class="loop-step" role="tab" tabindex="0" id="loop-tab-3" data-loop-step="3" aria-selected="false" aria-controls="loop-panel" aria-expanded="false">
             <span class="loop-number">3</span>
+            <div class="thumb-icon" aria-hidden="true">📈</div>
             <h3>Rank and refine</h3>
-            <p>Re-rank relevant lessons for the action. Repeated negative patterns can promote from warnings to blocking gates; stale gates expire.</p>
+            <p>Re-rank lessons for the action. Repeated negatives can promote to blocking gates; stale gates expire.</p>
             <span class="loop-hint">Under the hood ↓</span>
           </div>
           <div class="loop-step" role="tab" tabindex="0" id="loop-tab-4" data-loop-step="4" aria-selected="false" aria-controls="loop-panel" aria-expanded="false">
             <span class="loop-number">4</span>
+            <div class="thumb-icon" aria-hidden="true">🛡️</div>
             <h3>Gate the next action</h3>
             <p>The action is allowed, warned, or denied before the tool proceeds.</p>
             <div class="decisions" aria-label="Possible decisions">
@@ -717,7 +846,7 @@
       <div class="shell">
         <div class="section-kicker">What the $499 enterprise gate buys</div>
         <h2 class="section-title">A managed gate for one painful workflow.</h2>
-        <p class="section-lede">This is the enterprise entry offer for one workflow, not an org-wide hosted platform license.</p>
+        <p class="section-lede">Enterprise entry for one workflow—not an org-wide platform license.</p>
         <div class="deliverables">
           <article class="deliverable">
             <h3>Failure map</h3>
@@ -760,9 +889,10 @@ next      decision recorded before execution</pre>
             </div>
           </div>
           <div class="proof-copy">
-            <p>Detected secret exfiltration and attempts to kill or bypass the gate process are denied by default. Matching destructive actions warn by default and deny in strict mode.</p>
-            <p>The example shows the strict-mode decision the managed install would specify and test. It is not a claim that every free install blocks every risky command automatically.</p>
+            <p>Detected secret exfiltration and gate-process bypass attempts are denied by default (process-kill/environment-override self-disable floors). Matching destructive actions warn by default and deny in strict mode.</p>
+            <p>Managed strict-mode example—not a claim that every free install blocks every risky command automatically.</p>
             <a href="https://github.com/IgorGanapolsky/ThumbGate/blob/main/docs/VERIFICATION_EVIDENCE.md" target="_blank" rel="noopener">Read the test-backed verification evidence →</a>
+            <p style="margin-top:12px;"><a href="/whitepaper">White paper</a> · <a href="/eval-scorecard">Scorecard</a> · <a href="/architecture">Architecture</a> · <a href="/case-studies">Cases</a></p>
           </div>
         </div>
       </div>
@@ -801,7 +931,7 @@ next      decision recorded before execution</pre>
       <div class="shell">
         <div class="section-kicker">Two cash paths</div>
         <h2 class="section-title">Start self-serve, or bring the failure that already cost you time.</h2>
-        <p class="section-lede">Pro is the self-serve subscription for operators. The $499 enterprise gate is a managed install for one painful workflow—not another policy deck.</p>
+        <p class="section-lede">Pro is self-serve for operators. The $499 gate is a managed install for one painful workflow—not a policy deck.</p>
         <div style="display:flex;flex-wrap:wrap;gap:12px;justify-content:center;margin-top:22px">
           <a class="nav-buy" href="/checkout/pro?utm_source=website&amp;utm_medium=homepage_final&amp;utm_campaign=pro_self_serve&amp;cta_id=final_pro_buy&amp;cta_placement=final&amp;plan_id=pro" data-offer-link data-cta-id="final_pro_buy">Start Pro — $19/mo</a>
           <a class="nav-enterprise" href="#enterprise-gate" data-offer-link data-cta-id="final_diagnostic_buy">Buy the $499 enterprise gate</a>
@@ -812,7 +942,7 @@ next      decision recorded before execution</pre>
   <footer>
     <div class="shell footer-inner">
-      <span>ThumbGate · MIT License · npm v1.29.2</span>
+      <span>ThumbGate · MIT License · npm v1.31.0</span>
       <div class="footer-links">
         <a href="https://github.com/IgorGanapolsky/ThumbGate" target="_blank" rel="noopener">GitHub</a>
         <a href="/guide">Technical setup</a>
@@ -1019,7 +1149,7 @@ next      decision recorded before execution</pre>
             'mode      <span class="y">warn-by-default</span>\n' +
             'decision  <span class="y">WARN</span>\n' +
             'note      logged + flagged; tool may still proceed\n\n' +
-            'secret exfil / gate-bypass floors  →  <span class="r">DENY</span> regardless',
+            'secret exfil / gate process bypass floors  →  <span class="r">DENY</span> regardless',
           bodyStrict:
             'proposed  git push --force origin main\n' +
             'hook      PreToolUse → gates-engine\n' +
@@ -1029,7 +1159,7 @@ next      decision recorded before execution</pre>
             'next      blocked before execution',
           facts: [
             'PreToolUse hooks check Bash / Edit / Write before the tool runs',
-            'Outcomes: ALLOW, WARN, or DENY',
+            'Outcomes: allowed, warned, or denied (ALLOW / WARN / DENY)',
             'Default install is warn + audit; hard deny needs strict mode (plus secret floors)',
           ],
           honest: 'Default is not “block everything.” Default warns. Strict mode denies matching high-risk actions. Secrets/exfil still hard-deny.',

package/public/numbers.html CHANGED Viewed

@@ -25,7 +25,7 @@
   "alternateName": "thumbgate",
   "applicationCategory": "DeveloperApplication",
   "operatingSystem": "Cross-platform, Node.js >=18.18.0",
-  "softwareVersion": "1.29.2",
+  "softwareVersion": "1.31.0",
   "url": "https://thumbgate.ai/numbers",
   "dateModified": "2026-05-07",
   "creator": {
@@ -195,6 +195,7 @@
   <a href="/learn">Learn</a>
   <a href="/compare">Compare</a>
   <a href="/numbers">Numbers</a>
+  <a href="/evaluations">Evaluations</a>
   <a href="/dashboard">Dashboard</a>
   <a href="/pro">Pro</a>
 </nav>
@@ -202,7 +203,7 @@
 <main class="container">
   <h1>The Numbers</h1>
   <p class="subtitle">Generated first-party operational snapshot from the ThumbGate runtime. This is not customer traction, install volume, revenue, or proof that a configured gate has fired.</p>
-  <div class="freshness">Updated: 2026-05-07 · Version 1.29.2</div>
+  <div class="freshness">Updated: 2026-05-07 · Version 1.31.0</div>
   <div class="truth-note"><strong>Read this first:</strong> configured checks are inventory. Recorded blocks and warnings are usage evidence. This snapshot currently reports 0 recorded hard-block event(s) and 0 recorded warning event(s).</div>
   <h2>Gate enforcement</h2>