Hugging Face's logo

Spaces:
ratingtesting
/
keelwright
Running

App Files Community
1
keelwright
File size: 25,029 Bytes
c54957a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
<!DOCTYPE html>

<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> &nbsp;→&nbsp; GOAL + ACCEPTANCE CRITERIA &nbsp;&nbsp;◄──&nbsp;&nbsp; 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">▼ &nbsp;perimeter supervision&nbsp; ▼</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. &nbsp;·&nbsp;

      <b>PERSONAS</b> (delegate_task): @architect @implementer @tester @reviewer @architecture-critic @visual-qa &nbsp;·&nbsp;

      <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 &nbsp;·&nbsp; ⚡ 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>