Spaces:
Running
Running
| <html lang="en"> | |
| <head> | |
| <script type="application/ld+json"> | |
| { | |
| "@context": "https://schema.org", | |
| "@type": "SoftwareSourceCode", | |
| "name": "keelwright", | |
| "description": "Engine for vibe-coders and loop-coders who ship AI-generated code they cannot read. 28 machine-enforced safety checks: SQL injection, hardcoded secrets, slopsquatting, reward hacking, doom loops, false reports, missing auth, business logic bypasses.", | |
| "url": "https://github.com/ratingtesting/keelwright", | |
| "license": "https://opensource.org/license/mit-0", | |
| "author": { | |
| "@type": "Organization", | |
| "name": "ratingtesting", | |
| "url": "https://github.com/ratingtesting" | |
| }, | |
| "keywords": "vibe coding, loop coding, AI code security, guardrails, OWASP, SQL injection, secrets, slopsquatting, reward hacking, doom loop, circuit breaker, token burn, context rot, over-engineering, tech debt, code quality, non-programmer, autonomous coding", | |
| "applicationCategory": "DeveloperApplication", | |
| "operatingSystem": "cross-platform", | |
| "softwareHelp": { | |
| "@type": "CreativeWork", | |
| "url": "https://github.com/ratingtesting/keelwright/blob/main/README.md" | |
| } | |
| } | |
| </script> | |
| <meta charset="UTF-8"> | |
| <meta name="viewport" content="width=device-width, initial-scale=1.0"> | |
| <title>Keelwright — Architecture</title> | |
| <link href="https://fonts.googleapis.com/css2?family=JetBrains+Mono:wght@400;600;700&display=swap" rel="stylesheet"> | |
| <style> | |
| * { margin:0; padding:0; box-sizing:border-box; } | |
| body { | |
| background:#020617; color:#e2e8f0; font-family:'JetBrains Mono',monospace; | |
| padding:40px 36px; width:1200px; font-size:22px; line-height:1.55; | |
| background-image: | |
| linear-gradient(#0b1220 1px,transparent 1px), | |
| linear-gradient(90deg,#0b1220 1px,transparent 1px); | |
| background-size:48px 48px; | |
| } | |
| .head { display:flex; align-items:center; gap:14px; margin-bottom:8px; } | |
| .dot { width:16px; height:16px; border-radius:50%; background:#34d399; box-shadow:0 0 16px #34d399; animation:p 2s infinite; } | |
| @keyframes p { 0%,100%{opacity:1} 50%{opacity:.35} } | |
| h1 { font-size:40px; font-weight:700; letter-spacing:.5px; } | |
| h1 .kw { color:#34d399; } | |
| .sub { color:#64748b; font-size:18px; margin:2px 0 32px 30px; } | |
| .layer { border-radius:14px; padding:24px 26px; margin-bottom:22px; position:relative; } | |
| .ltag { position:absolute; top:-11px; left:20px; font-size:14px; letter-spacing:1.5px; | |
| background:#020617; padding:0 10px; color:#64748b; text-transform:uppercase; } | |
| .l-human { background:rgba(8,51,68,.35); border:2px solid #22d3ee; } | |
| .l-ctrl { background:rgba(30,41,59,.5); border:2px solid #94a3b8; } | |
| .l-build { background:rgba(6,78,59,.28); border:2px solid #34d399; } | |
| .l-sup { background:transparent; border:none; padding:0; margin-bottom:18px; } | |
| .l-ops { background:rgba(120,53,15,.22); border:2px dashed #fbbf24; } | |
| .title { font-size:26px; font-weight:700; margin-bottom:12px; display:flex; align-items:center; gap:12px; } | |
| .c-cyan{color:#22d3ee}.c-em{color:#34d399}.c-vi{color:#a78bfa}.c-am{color:#fbbf24} | |
| .c-rose{color:#fb7185}.c-or{color:#fb923c}.c-sl{color:#94a3b8} | |
| .star { color:#fbbf24; font-weight:700; } | |
| .starstar { color:#fb923c; font-weight:700; } | |
| /* 2-column grid for boxes */ | |
| .row2 { display:grid; grid-template-columns:1fr 1fr; gap:18px; } | |
| .box { background:#0f172a; border-radius:10px; padding:16px 18px; font-size:22px; line-height:1.55; } | |
| .muted { color:#94a3b8; } | |
| .tiny { font-size:17px; color:#64748b; } | |
| .catch { color:#fb7185; font-size:17px; } | |
| .arrowdown { text-align:center; color:#34d399; font-size:30px; line-height:1; margin:6px 0; } | |
| /* build loop steps */ | |
| .steps { display:flex; flex-direction:column; gap:12px; } | |
| .step { display:flex; align-items:flex-start; gap:14px; background:#0f172a; border-left:4px solid #34d399; | |
| border-radius:8px; padding:14px 18px; } | |
| .step .n { color:#34d399; font-weight:700; min-width:32px; font-size:24px; } | |
| .step .body { flex:1; font-size:20px; line-height:1.55; } | |
| .step.sec { border-left-color:#fb7185; } | |
| .step.verify { border-left-color:#22d3ee; } | |
| .step.match { border-left-color:#a78bfa; background:rgba(76,29,149,.18); } | |
| .loopback { display:flex; align-items:center; gap:10px; color:#34d399; font-size:14px; | |
| justify-content:flex-end; margin-top:6px; } | |
| /* Layer 3: 2-column + 1 */ | |
| .sup2 { display:grid; grid-template-columns:1fr 1fr; gap:18px; } | |
| .supbox { background:#0f172a; border-radius:10px; padding:20px 22px; } | |
| .supbox.cb { border-top:4px solid #fb923c; } | |
| .supbox.st { border-top:4px solid #a78bfa; } | |
| .supbox.sl { border-top:4px solid #22d3ee; } | |
| .supbox h4 { font-size:24px; margin-bottom:10px; } | |
| .supbox ul { list-style:none; font-size:19px; line-height:1.7; color:#cbd5e1; } | |
| .crosscut { background:rgba(136,19,55,.25); border:2px solid #fb7185; border-radius:10px; | |
| padding:16px 20px; margin-top:16px; font-size:18px; line-height:1.7; } | |
| /* Layer 4: 3 nodes per row */ | |
| .ops-grid { display:grid; grid-template-columns:1fr 1fr 1fr; gap:12px; } | |
| .ops-row2 { display:grid; grid-template-columns:1fr 1fr; gap:12px; margin-top:12px; } | |
| .ops-node { background:#0f172a; border:2px solid #fbbf24; border-radius:10px; padding:16px 18px; | |
| text-align:center; font-size:22px; } | |
| .ops-node .em { display:block; font-size:28px; margin-bottom:4px; } | |
| .ops-node .sub2 { font-size:16px; color:#94a3b8; margin-top:4px; } | |
| .ops-arrow { color:#fbbf24; font-size:20px; text-align:center; } | |
| .rollback { text-align:center; color:#fb7185; font-size:13px; margin-top:12px; | |
| border-top:2px dashed #fb7185; padding-top:12px; } | |
| /* Banner: 2 columns */ | |
| .banner2 { display:grid; grid-template-columns:1fr 1fr; gap:14px; margin-top:20px; } | |
| .banner2 .bcard3 { grid-column:1 / -1; } | |
| .bcard { background:#0f172a; border-radius:10px; padding:16px 18px; } | |
| .bcard .h { font-size:24px; font-weight:700; margin-bottom:6px; } | |
| .bcard .t { font-size:17px; color:#94a3b8; line-height:1.5; } | |
| .b1{border-left:4px solid #fbbf24}.b2{border-left:4px solid #34d399}.b3{border-left:4px solid #22d3ee} | |
| .foot { color:#475569; font-size:16px; margin-top:24px; text-align:center; letter-spacing:.5px; } | |
| /* risk map: 2 columns */ | |
| .riskmap { background:rgba(136,19,55,.16); border:2px solid #fb7185; border-radius:14px; | |
| padding:22px 22px 18px; margin-top:20px; position:relative; } | |
| .riskmap .rtag { position:absolute; top:-11px; left:22px; font-size:12px; letter-spacing:1.2px; | |
| background:#020617; padding:0 10px; color:#fb7185; text-transform:uppercase; } | |
| .riskmap h3 { font-size:26px; font-weight:700; color:#fb7185; margin-bottom:4px; } | |
| .riskmap .rsub { font-size:17px; color:#94a3b8; margin-bottom:18px; } | |
| .rgrid { display:grid; grid-template-columns:1fr 1fr; gap:12px 18px; } | |
| .ritem { background:#0f172a; border-radius:8px; padding:12px 16px; font-size:18px; line-height:1.5; | |
| border-left:3px solid #34d399; } | |
| .ritem.warn { border-left-color:#fbbf24; } | |
| .ritem .rn { font-weight:700; color:#e2e8f0; } | |
| .ritem .rm { color:#94a3b8; } | |
| .ritem .rb { float:right; font-size:18px; } | |
| .rlegend { font-size:16px; color:#64748b; margin-top:16px; line-height:1.7; } | |
| .rlegend b { color:#94a3b8; } | |
| .kds { background:#0a0f1a; padding:28px; border-radius:12px; margin-top:24px; } | |
| .kds h3 { font-size:26px; color:#fbbf24; margin:0 0 8px; text-align:center; } | |
| .ksub { font-size:15px; color:#94a3b8; text-align:center; margin-bottom:18px; line-height:1.6; } | |
| .ksub b { color:#e2e8f0; } | |
| .ktable { width:100%; border-collapse:collapse; font-size:15px; } | |
| .ktable th { background:#1e293b; color:#fbbf24; padding:8px 12px; text-align:left; border:1px solid #334155; } | |
| .ktable td { padding:8px 12px; border:1px solid #1e293b; color:#e2e8f0; } | |
| .ktable tr:nth-child(even) td { background:#0f1626; } | |
| .khi { font-weight:800; color:#4ade80; font-size:17px; text-align:center; } | |
| .kpart { font-weight:800; color:#fbbf24; text-align:center; } | |
| .kzero { font-weight:800; color:#f87171; text-align:center; } | |
| </style> | |
| </head> | |
| <body> | |
| <div class="head"><div class="dot"></div><h1><span class="kw">keelwright</span> — engine for vibe / loop coding</h1></div> | |
| <div class="sub">a loop with nested loops · safety cannot be switched off · a human who does not read code stays in control</div> | |
| <!-- LAYER 0 --> | |
| <div class="layer l-human"> | |
| <span class="ltag">Layer 0 · the human</span> | |
| <div class="title"><span class="c-cyan">▣ YOU</span> <span class="muted">(need not read code)</span> → GOAL + ACCEPTANCE CRITERIA ◄── PLAIN-LANGUAGE REPORT / BUILD</div> | |
| <div class="tiny"><span class="star">★ DNA:</span> understands LOGIC, not syntax. Every gate below is translated to plain words — <span class="muted">"this checks the password is right"</span>, never "string arg to fn()".</div> | |
| </div> | |
| <div class="arrowdown">▼</div> | |
| <!-- LAYER 1: 2×2 grid --> | |
| <div class="layer l-ctrl"> | |
| <span class="ltag">Layer 1 · control</span> | |
| <div class="row2"> | |
| <div class="box"> | |
| <div class="title"><span class="c-sl">🎛 AUTONOMY DIAL</span></div> | |
| <div>Autopilot → Checkpoint → Copilot</div> | |
| <div class="tiny"><span class="star">★</span> freedom = blast radius (auth / money / prod → Copilot: approve each)</div> | |
| </div> | |
| <div class="box"> | |
| <div class="title"><span class="c-sl">🗂 TRIAGE</span></div> | |
| <div>Trivial · Low · Standard · High · Critical</div> | |
| <div class="tiny"><span class="star">★</span> machinery = risk. Trivial → Express, no per-iter scans</div> | |
| </div> | |
| <div class="box"> | |
| <div class="title"><span class="c-sl">▸ PHASE 1–2</span></div> | |
| <div>Requirements → Planning</div> | |
| <div class="tiny"><span class="star">★</span> no acceptance criteria? ⛔ STOP, ask. Never invents scope</div> | |
| </div> | |
| <div class="box"> | |
| <div class="title"><span class="c-sl">⬚ LOOP DESIGN</span></div> | |
| <div>5 whiteboard questions before any code</div> | |
| <div class="tiny"><span class="star">★</span> trigger · check · action · stop · escalate — skip = demo</div> | |
| </div> | |
| </div> | |
| </div> | |
| <div class="arrowdown">▼</div> | |
| <!-- LAYER 2 --> | |
| <div class="layer l-build"> | |
| <span class="ltag">Layer 2 · the build loop (Phase 3 — one task = one turn, repeats)</span> | |
| <div class="title"><span class="c-em">⟳ PHASE 3 · BUILDING</span> <span class="tiny">— discipline INSIDE writing, not cleanup afterward</span></div> | |
| <div class="steps"> | |
| <div class="step"><span class="n">①</span><div class="body">@implementer writes (1 file/fn) ◄ <b>REUSE LADDER</b> — reuse before adding a lib/abstraction <span class="catch">· catches: over-engineering</span></div></div> | |
| <div class="step"><span class="n">②</span><div class="body"><b>BACKPRESSURE GATES:</b> tests │ typecheck │ lint │ build</div></div> | |
| <div class="step"><span class="n">③</span><div class="body"><b>STRUCTURAL-INTEGRITY GATE</b> <span class="star">★</span> jscpd·lizard·madge/import-linter·eslint-boundaries·knip/vulture <span class="catch">· catches: spaghetti, lava flow, tangled deps</span></div></div> | |
| <div class="step sec"><span class="n">④</span><div class="body"><span class="c-rose">🛡 SECURITY GATES R1–R12</span> <span class="starstar">★★</span> machine-enforced, cannot disable even on Autopilot <span class="catch">· catches: OWASP·secrets·slopsquat·IDOR</span></div></div> | |
| <div class="step sec"><span class="n">⑤</span><div class="body"><span class="c-rose">🔍 R3 BUSINESS-LOGIC REVIEW</span> ► spawn @reviewer (fresh context) <span class="star">★</span> NOT self-review <span class="catch">· catches: happy-path bias, auth holes</span></div></div> | |
| <div class="step verify"><span class="n">⑦</span><div class="body"><span class="c-cyan">✅ VERIFICATION GATE</span> <span class="starstar">★★</span> read+compile+diff · test red→green · not tautological · ad-hoc <span class="catch">· catches: "works on my machine", "I fixed it" with no diff</span></div></div> | |
| <div class="step match"><span class="n">⑧</span><div class="body"><span class="c-vi">⟲ MATCH LOOP (nested sub-loop)</span> <span class="star">★</span> Generator ⇄ Analyst: render → browser → numeric a11y (contrast≥4.5:1 · no overflow · sizes) — repeat until convergence</div></div> | |
| <div class="step"><span class="n">⑨</span><div class="body">git commit (atomic) ► ⑩ PROGRESS.md + STATUS + tool-budget count</div></div> | |
| <div class="step"><span class="n">⑪</span><div class="body">cadence: /3 → STABILITY scan · /10 → AUTORESEARCH · /N → refactor checkpoint</div></div> | |
| </div> | |
| <div class="loopback">↩ LOOP BACK — next task, while tasks remain</div> | |
| </div> | |
| <div class="arrowdown">▼ perimeter supervision ▼</div> | |
| <!-- LAYER 3: 2-column --> | |
| <div class="l-sup"> | |
| <div class="sup2"> | |
| <div class="supbox cb"> | |
| <h4><span class="c-or">🚦 CIRCUIT-BREAKER</span></h4> | |
| <ul> | |
| <li class="catch">catches: doom loop + token/$ burn</li> | |
| <li>4 caps: 50 iters / 5 no-progress / 2h / 3× repeat</li> | |
| <li>+ RATE LIMITING (event-driven loops)</li> | |
| <li>+ PER-ITER BUDGETS ≤10 shell / 5 file / 3 external</li> | |
| <li>→ graceful STOP with report</li> | |
| </ul> | |
| </div> | |
| <div class="supbox st"> | |
| <h4><span class="c-vi">📉 STABILITY (5 modes)</span></h4> | |
| <ul> | |
| <li class="catch">catches: goal drift, oscillation, thrash</li> | |
| <li>dead-retry · oscillation · drift · amplification · feedback-starvation</li> | |
| <li>ESCALATION: 3→REFINE 5→PIVOT 2→ask 3→blocker</li> | |
| <li>+ CONTEXT COMPACTION (iter > 20)</li> | |
| <li>a "keep" resets the counters</li> | |
| </ul> | |
| </div> | |
| </div> | |
| <!-- Self-learning: full width --> | |
| <div class="supbox sl" style="margin-top:14px;"> | |
| <h4><span class="c-cyan">🔄 SELF-LEARNING</span> <span class="star">★</span></h4> | |
| <ul> | |
| <li>Phoenix: root-cause of repeats ACROSS sessions</li> | |
| <li>Autoresearch: lessons → memory (promote on repeat)</li> | |
| <li class="catch">learns, doesn't repeat the same mistake</li> | |
| </ul> | |
| </div> | |
| <div class="crosscut"> | |
| <span class="starstar">★★ REWARD-HACKING GUARD</span> — never weaken/delete a test to go green; fix the code, not the metric. · | |
| <b>PERSONAS</b> (delegate_task): @architect @implementer @tester @reviewer @architecture-critic @visual-qa · | |
| <b>STRUCTURED FEEDBACK</b> into each iteration: relevant code + intent + "repeat vs new" flag — not a raw stack-trace dump. | |
| </div> | |
| </div> | |
| <!-- LAYER 4: 3 per row + 2 --> | |
| <div class="layer l-ops"> | |
| <span class="ltag">Layer 4 · outer operational cycle (production lifecycle)</span> | |
| <div class="ops-grid"> | |
| <div class="ops-node"><span class="em">🔭</span>OBSERVE<span class="sub2">read prod logs/reports</span></div> | |
| <div class="ops-node"><span class="em">🩺</span>ANALYZE<span class="sub2">root-cause + plan</span></div> | |
| <div class="ops-node"><span class="em">🔧</span>FIX<span class="sub2">run the build loop</span></div> | |
| </div> | |
| <div class="ops-row2"> | |
| <div class="ops-node"><span class="em">✅</span>VALIDATE<span class="sub2">post-deploy metric compare</span></div> | |
| <div class="ops-node"><span class="em">🎓</span>LEARN<span class="sub2">update lessons</span></div> | |
| </div> | |
| <div class="rollback">⭐ verify-in-production + AUTO-ROLLBACK (git revert, not force-push) — a bad deploy that passes tests is caught by the metric, not by a human · ⚡ PARALLEL: independent subtasks fan out as separate empty-context loops</div> | |
| </div> | |
| <!-- BANNER: 2-column + 1 full --> | |
| <div class="banner2"> | |
| <div class="bcard b1"><div class="h"><span class="c-am">🔒 BATTLE-TESTED QA ★★★</span></div><div class="t">A/B (control vs skill), fact-checked on disk, not self-report — published WITH an honest adversarial report.</div></div> | |
| <div class="bcard b2"><div class="h"><span class="c-em">💸 ZERO INSTALL · ZERO COST</span></div><div class="t">Plain markdown that orchestrates free, permissively-licensed tools. Full transparency, nothing to buy.</div></div> | |
| <div class="bcard b3"><div class="h"><span class="c-cyan">🗣 PLAIN-LANGUAGE THROUGHOUT</span></div><div class="t">Built for people who read product logic, not code syntax. Every gate speaks human.</div></div> | |
| </div> | |
| <!-- RISK MAP: 2 columns --> | |
| <div class="riskmap"> | |
| <span class="rtag">documented 2026</span> | |
| <h3>🗺 VIBE/LOOP-CODING RISK MAP → WHAT KEELWRIGHT COVERS</h3> | |
| <div class="rsub">Every known vibe/loop-coding pain → the mechanism that closes it. <span class="c-em">✅</span> = machine-enforced / on-disk · <span class="c-am">⚠</span> = partial (honest: no cheap machine detector)</div> | |
| <div class="rgrid"> | |
| <div class="ritem"><span class="rn">Context rot / decay</span><span class="rb">✅</span><br><span class="rm">fresh-context + PROGRESS.md</span></div> | |
| <div class="ritem"><span class="rn">Ralph (Wiggum) loop</span><span class="rb">✅</span><br><span class="rm">state in files, not chat</span></div> | |
| <div class="ritem"><span class="rn">Spaghetti / big ball of mud</span><span class="rb">✅</span><br><span class="rm">structural-integrity gate ③</span></div> | |
| <div class="ritem"><span class="rn">Lava flow (dead code)</span><span class="rb">✅</span><br><span class="rm">knip / vulture</span></div> | |
| <div class="ritem"><span class="rn">Dependency hell</span><span class="rb">✅</span><br><span class="rm">reuse ladder + dep vetting</span></div> | |
| <div class="ritem"><span class="rn">Slopsquatting</span><span class="rb">✅</span><br><span class="rm">R8 verify + GuardDog</span></div> | |
| <div class="ritem"><span class="rn">"Works on my machine"</span><span class="rb">✅</span><br><span class="rm">verification gate ⑦</span></div> | |
| <div class="ritem"><span class="rn">Reward hacking / spec gaming</span><span class="rb">✅</span><br><span class="rm">reward-hacking guard</span></div> | |
| <div class="ritem"><span class="rn">Doom loop / death loop</span><span class="rb">✅</span><br><span class="rm">circuit-breaker + budgets</span></div> | |
| <div class="ritem"><span class="rn">Goal drift</span><span class="rb">✅</span><br><span class="rm">Stability L3 + goal re-read</span></div> | |
| <div class="ritem"><span class="rn">Happy-path bias</span><span class="rb">✅</span><br><span class="rm">R3 business-logic review</span></div> | |
| <div class="ritem"><span class="rn">Tech debt (80% problem)</span><span class="rb">✅</span><br><span class="rm">R4 prod checklist</span></div> | |
| <div class="ritem"><span class="rn">Secret leakage</span><span class="rb">✅</span><br><span class="rm">R2 Gitleaks</span></div> | |
| <div class="ritem"><span class="rn">SQL injection / OWASP</span><span class="rb">✅</span><br><span class="rm">R1 Semgrep + review</span></div> | |
| <div class="ritem"><span class="rn">Reasoning-action (said≠did)</span><span class="rb">✅</span><br><span class="rm">R7 + verification</span></div> | |
| <div class="ritem"><span class="rn">Malicious 3rd-party skill</span><span class="rb">✅</span><br><span class="rm">R11 SkillSpector audit</span></div> | |
| <div class="ritem"><span class="rn">Token / budget burn</span><span class="rb">✅</span><br><span class="rm">per-iteration budgets</span></div> | |
| <div class="ritem"><span class="rn">Regression passes tests</span><span class="rb">✅</span><br><span class="rm">post-deploy + rollback</span></div> | |
| <div class="ritem"><span class="rn">Human as bottleneck</span><span class="rb">✅</span><br><span class="rm">autonomy dial + parallel</span></div> | |
| <div class="ritem warn"><span class="rn">Design-for-failure</span><span class="rb">⚠</span><br><span class="rm">R5 fault checklist (warn)</span></div> | |
| <div class="ritem warn"><span class="rn">Sycophancy</span><span class="rb">⚠</span><br><span class="rm">R7 catches false claims, not trait</span></div> | |
| <div class="ritem warn"><span class="rn">Multi-agent cascade / mem-poison</span><span class="rb">⚠</span><br><span class="rm">R10 verify-before-memory (warn)</span></div> | |
| <div class="ritem warn"><span class="rn">Model version drift</span><span class="rb">⚠</span><br><span class="rm">R9 pin (warn)</span></div> | |
| <div class="ritem warn"><span class="rn">Style / consistency drift</span><span class="rb">⚠</span><br><span class="rm">critic + Pink Flag (no machine detector)</span></div> | |
| <div class="ritem warn"><span class="rn">Confabulation / fabricated facts</span><span class="rb">⚠</span><br><span class="rm">factual-grounding: verify-before-assert</span></div> | |
| <div class="ritem"><span class="rn">Loop design absence</span><span class="rb">✅</span><br><span class="rm">5 whiteboard questions</span></div> | |
| <div class="ritem"><span class="rn">Context rot in long loops</span><span class="rb">✅</span><br><span class="rm">compaction (trim/summarize/delegate)</span></div> | |
| <div class="ritem"><span class="rn">Event storm / rate limiting</span><span class="rb">✅</span><br><span class="rm">rate limit + debounce + backpressure</span></div> | |
| </div> | |
| <div class="rlegend"><b>Honesty as a feature:</b> three ⚠ (style consistency, sycophancy-as-trait, confabulation) cannot be closed by a cheap machine detector — keelwright catches their measurable consequences and enforces verification discipline, not the disposition itself. Everything marked ✅ is machine-enforced or verified on disk, not "eyeball it."</div> | |
| </div> | |
| <!-- KDS LEADERBOARD --> | |
| <div class="kds"> | |
| <h3>📊 KEELWRIGHT SCORE (KDS) — PROVEN BY A/B TESTING</h3> | |
| <div class="ksub">KDS = Execution Rate × Discrimination Rate / 100. 11+ models tested, all verified on disk with <code>validate_run.py</code>.</div> | |
| <table class="ktable"> | |
| <thead><tr><th>Model</th><th>Tier</th><th>Benchmark</th><th>KDS</th></tr></thead> | |
| <tbody> | |
| <tr><td>poolside/laguna-s-2.1</td><td>STRONG</td><td>ML 78.5%</td><td class="khi">83</td></tr> | |
| <tr><td>stepfun/step-3.7-flash</td><td>MEDIUM</td><td>Pro ~56%</td><td class="khi">67</td></tr> | |
| <tr><td>nvidia/nemotron-3-ultra</td><td>STRONG</td><td>ML 67.7%</td><td class="khi">40</td></tr> | |
| <tr><td>deepseek-v4-flash</td><td>STRONG</td><td>Verified ~79%</td><td class="khi">29</td></tr> | |
| <tr><td>inclusionai/ling-3.0-flash</td><td>UNKNOWN</td><td>SWE-bench/GPQA N/A</td><td class="khi">22</td></tr> | |
| <tr><td>kimi-k3</td><td>STRONG</td><td>Terminal-Bench 88.3, ProgramBench 77.8</td><td class="khi">25</td></tr> | |
| <tr><td>mimo-v2.5</td><td>MEDIUM</td><td>Verified 78.9%, Pro 57.2%</td><td class="khi">18</td></tr> | |
| <tr><td>nvidia/nemotron-3-super-120b-a12b</td><td>STRONG</td><td>Verified 60.47%</td><td class="kpart">PARTIAL</td></tr> | |
| <tr><td>claude-opus-4-8</td><td>STRONG</td><td>frontier</td><td class="khi">17</td></tr> | |
| <tr><td>tencent/hy3</td><td>STRONG</td><td>ML 75.8%, Verified 78%</td><td class="khi">7</td></tr> | |
| <tr><td>cohere/north-mini-code</td><td>WEAK</td><td>Agentic 3.1</td><td class="kzero">0</td></tr> | |
| </tbody> | |
| </table> | |
| <div class="ksub"><b>Key findings:</b> Medium models (Step 3.7, KDS 67) get MORE value than some strong models · kimi-k3 (STRONG) still benefits — autonomy dial stopped a silent business-hack commit · Ling re-run after fabricated first attempt: clean 4 DISCRIMINATES · Nemotron-3-super partial (2/18, both DISCRIMINATES) · weak models (KDS 0) fabricate, caught by gate.</div> | |
| </div> | |
| <div class="foot"><a href="https://github.com/ratingtesting/keelwright" style="color:#94a3b8;text-decoration:underline">keelwright</a> by ratingtesting · MIT-0 · <a href="https://huggingface.co/spaces/ratingtesting/keelwright" style="color:#94a3b8;text-decoration:underline">Hugging Face</a> · source of truth: assets/architecture.md</div> | |
| </body> | |
| </html> | |