-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathdeck.html
More file actions
430 lines (403 loc) · 25.5 KB
/
Copy pathdeck.html
File metadata and controls
430 lines (403 loc) · 25.5 KB
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
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8" />
<meta name="viewport" content="width=device-width, initial-scale=1" />
<title>FlakeWarden · AgentHack 2026 deck</title>
<style>
:root{
--bg:#0a0e1a; --bg2:#121a2e; --panel:#161f38; --ink:#eef2fb; --muted:#9aa6c2;
--line:#26334f; --accent:#ff5a2c; --good:#34d399; --blue:#5b9dff; --amber:#fbbf24;
--mono:"Cascadia Code",ui-monospace,Consolas,"Courier New",monospace;
--sans:"Segoe UI",system-ui,-apple-system,Roboto,Helvetica,Arial,sans-serif;
}
*{box-sizing:border-box;margin:0;padding:0}
html,body{height:100%;background:#05070d;font-family:var(--sans);color:var(--ink)}
body{display:flex;align-items:center;justify-content:center;overflow:hidden}
/* one logical slide = 1280x720, scaled to fit */
#stage{position:relative;width:1280px;height:720px;transform-origin:center center}
.slide{position:absolute;inset:0;display:none;flex-direction:column;
padding:54px 64px 48px;background:
radial-gradient(1200px 600px at 85% -10%, rgba(255,90,44,.10), transparent 60%),
radial-gradient(900px 500px at -10% 110%, rgba(91,157,255,.10), transparent 60%),
var(--bg);border:1px solid var(--line);border-radius:18px;overflow:hidden}
.slide.active{display:flex}
.crit{position:absolute;top:20px;right:24px;font-size:12px;letter-spacing:.04em;
color:var(--muted);border:1px solid var(--line);border-radius:999px;
padding:5px 12px;background:rgba(255,255,255,.02)}
.kicker{font-size:14px;letter-spacing:.18em;text-transform:uppercase;color:var(--accent);font-weight:700}
h1{font-size:40px;line-height:1.1;font-weight:800;letter-spacing:-.01em}
h2{font-size:33px;line-height:1.12;font-weight:800;letter-spacing:-.01em;margin-bottom:6px}
.sub{color:var(--muted);font-size:18px;line-height:1.45}
.accent{color:var(--accent)} .good{color:var(--good)} .blue{color:var(--blue)} .amber{color:var(--amber)}
.mono{font-family:var(--mono)}
.grow{flex:1} .row{display:flex;gap:20px} .col{display:flex;flex-direction:column;gap:14px}
.center{align-items:center;justify-content:center;text-align:center}
ul{list-style:none;display:flex;flex-direction:column;gap:13px;margin-top:6px}
li{position:relative;padding-left:26px;font-size:18px;line-height:1.4;color:#dde4f3}
li::before{content:"";position:absolute;left:2px;top:9px;width:9px;height:9px;border-radius:3px;
background:var(--accent);box-shadow:0 0 0 4px rgba(255,90,44,.14)}
li.g::before{background:var(--good);box-shadow:0 0 0 4px rgba(52,211,153,.16)}
li.b::before{background:var(--blue);box-shadow:0 0 0 4px rgba(91,157,255,.16)}
small,.fine{color:var(--muted);font-size:14px;line-height:1.4}
.card{background:linear-gradient(180deg,rgba(255,255,255,.03),rgba(255,255,255,.01));
border:1px solid var(--line);border-radius:14px;padding:18px 20px}
.chip{display:inline-flex;align-items:center;gap:8px;border:1px solid var(--line);
border-radius:999px;padding:7px 14px;font-size:14px;font-weight:600;background:rgba(255,255,255,.02)}
.dot{width:9px;height:9px;border-radius:50%}
.figframe{border:1px solid var(--line);border-radius:12px;overflow:hidden;background:#0c1322;
box-shadow:0 12px 30px rgba(0,0,0,.35)}
.figframe img{display:block;width:100%;height:100%;object-fit:cover;object-position:top center}
.cap{font-size:13px;color:var(--muted);margin-top:7px;line-height:1.35}
.foot{position:absolute;left:64px;right:64px;bottom:20px;display:flex;justify-content:space-between;
align-items:center;color:var(--muted);font-size:13px;border-top:1px solid var(--line);padding-top:12px}
/* big metric tiles */
.metrics{display:grid;grid-template-columns:repeat(3,1fr);gap:18px}
.metric{background:var(--panel);border:1px solid var(--line);border-radius:16px;padding:22px 20px;text-align:center}
.metric .n{font-size:46px;font-weight:850;letter-spacing:-.02em;line-height:1}
.metric .l{color:var(--muted);font-size:14px;margin-top:8px;line-height:1.3}
/* architecture flow */
.flow{display:flex;align-items:center;gap:10px;flex-wrap:nowrap;margin-top:10px}
.node{background:var(--panel);border:1px solid var(--line);border-radius:12px;padding:12px 14px;
min-width:120px;text-align:center;font-size:13.5px;line-height:1.25}
.node b{display:block;font-size:14px;margin-bottom:3px}
.node .t{color:var(--muted);font-size:11.5px}
.node.acc{border-color:rgba(255,90,44,.5);background:rgba(255,90,44,.07)}
.node.good{border-color:rgba(52,211,153,.45);background:rgba(52,211,153,.06)}
.node.blue{border-color:rgba(91,157,255,.45);background:rgba(91,157,255,.06)}
.arrow{color:var(--muted);font-size:13px;display:flex;flex-direction:column;align-items:center;gap:2px;min-width:64px}
.arrow .lbl{font-size:10.5px;letter-spacing:.02em}
.branchwrap{display:flex;flex-direction:column;gap:8px}
/* two-col compare */
.two{display:grid;grid-template-columns:1fr 1fr;gap:18px;margin-top:6px}
.pane{border:1px solid var(--line);border-radius:14px;padding:18px 20px;background:rgba(255,255,255,.02)}
.pane h3{font-size:18px;margin-bottom:6px;display:flex;align-items:center;gap:8px}
.pane.det{border-color:rgba(52,211,153,.35)} .pane.gen{border-color:rgba(91,157,255,.35)}
.pane ul{margin-top:2px} .pane li{font-size:15.5px;padding-left:22px} .pane li::before{top:8px;width:7px;height:7px}
/* components table */
table{width:100%;border-collapse:collapse;font-size:14.5px}
td,th{text-align:left;padding:8px 10px;border-bottom:1px solid var(--line);vertical-align:top}
th{color:var(--muted);font-weight:600;font-size:12.5px;letter-spacing:.04em;text-transform:uppercase}
td b{color:#fff}
.verdict-row{display:grid;grid-template-columns:repeat(3,1fr);gap:16px}
.verdict .figframe{height:184px}
.vhead{display:flex;align-items:center;gap:8px;margin-bottom:6px;font-weight:700;font-size:15px}
.hint{position:fixed;bottom:14px;left:50%;transform:translateX(-50%);color:#54607a;font-size:12px;z-index:50}
.pager{display:flex;gap:6px;align-items:center}
.pager b{color:var(--ink)}
/* print: one slide per page, all shown */
@media print{
@page{size:1280px 720px;margin:0}
html,body{height:auto;display:block;background:#fff;overflow:visible}
#stage{transform:none!important;width:1280px;height:720px}
.slide{display:flex!important;position:relative;inset:auto;page-break-after:always;
border-radius:0;border:none;width:1280px;height:720px}
.hint{display:none}
}
</style>
</head>
<body>
<div id="stage">
<!-- 1 · TITLE -->
<section class="slide active center">
<div class="kicker">UiPath AgentHack 2026 · Track 3 · UiPath Test Cloud</div>
<h1 style="font-size:74px;margin:18px 0 10px">Flake<span class="accent">Warden</span></h1>
<div style="font-size:24px;color:#dde4f3;font-weight:600">Agentic flaky-test triage for UiPath Test Cloud</div>
<div class="sub" style="max-width:780px;margin:16px auto 0">
Real defect, flaky, or environment? <span class="accent">With a human in charge of every change.</span>
</div>
<div class="row center" style="gap:10px;margin-top:30px;flex-wrap:wrap;justify-content:center">
<span class="chip"><span class="dot" style="background:var(--accent)"></span>Deterministic + generative</span>
<span class="chip"><span class="dot" style="background:var(--good)"></span>0% safety false-positive</span>
<span class="chip"><span class="dot" style="background:var(--blue)"></span>Built with a coding agent</span>
</div>
<div class="foot"><span>Jonathan Andrei</span><span>github.com/JonathanSolvesProblems/flakewarden · MIT</span></div>
</section>
<!-- 2 · PROBLEM -->
<section class="slide">
<span class="crit">Business Impact & Adoption</span>
<div class="kicker">The problem</div>
<h2>Flaky tests are the most corrosive failure mode in CI.</h2>
<p class="sub" style="max-width:920px">When a build goes red, you often can't tell a real regression from noise.</p>
<div class="metrics" style="margin-top:26px">
<div class="metric"><div class="n accent">~16%</div><div class="l">of Google's tests showed some flakiness</div></div>
<div class="metric"><div class="n accent">~84%</div><div class="l">of pass→fail transitions came from flaky tests</div></div>
<div class="metric"><div class="n">~100</div><div class="l">spurious failures per run at 5% flake on a 2,000-test suite</div></div>
</div>
<div class="card" style="margin-top:24px">
<b class="accent">The real cost isn't the triage minutes.</b>
<span class="sub" style="font-size:17px"> Teams stop trusting red builds, start ignoring them, and a genuine regression ships to production.</span>
</div>
<div class="foot"><span>Memon et al., <i>Taming Google-Scale Continuous Testing</i>, ICSE-SEIP 2017</span><span>2 / 12</span></div>
</section>
<!-- 3 · THE IDEA -->
<section class="slide">
<span class="crit">Creativity & Innovation</span>
<div class="kicker">The idea</div>
<h2>Answer the one question that matters, then route the failure.</h2>
<div class="row" style="gap:12px;margin:14px 0 6px">
<span class="chip"><span class="dot" style="background:var(--accent)"></span><b>real_defect</b> → escalate</span>
<span class="chip"><span class="dot" style="background:var(--amber)"></span><b>flaky</b> → propose a fix (human-gated)</span>
<span class="chip"><span class="dot" style="background:var(--blue)"></span><b>environment</b> → re-run on clean infra</span>
</div>
<div class="two" style="grid-template-columns:1fr;margin-top:14px">
<div class="pane">
<h3 class="accent">The uniqueness claim</h3>
<p class="sub" style="font-size:17px;color:#dde4f3">
Detection tools tell you a test is flaky. Healing tools auto-repair selectors.
<b class="accent">FlakeWarden is the governance layer between them:</b>
a reproducible, auditable scorer that spends an LLM only on the ambiguous band, under a hard safety
contract (a real regression is never auto-quarantined as flaky), with every fix a
<b>human-gated proposal written back through UiPath</b>, not an autonomous mutation.
</p>
</div>
</div>
<p class="fine" style="margin-top:14px">The moat is the <b class="accent">composition + safety contract</b>, not any single component. Positioned vs Datadog / Develocity / Trunk (detection) and Healenium / Tricentis / UiPath Autopilot (healing) in <span class="mono">docs/prior-art.md</span>.</p>
<div class="foot"><span>FlakeWarden</span><span>3 / 12</span></div>
</section>
<!-- 4 · THE THESIS -->
<section class="slide">
<span class="crit">AI Factor · Creativity</span>
<div class="kicker">The thesis</div>
<h2>Deterministic where the math must be exact. Generative where the evidence is messy.</h2>
<div class="two">
<div class="pane det">
<h3 class="good">◧ Deterministic flake-scorer</h3>
<ul>
<li class="g">Auditable statistics over run history, no model call</li>
<li class="g">Resolves the confident cases and <b>never guesses</b></li>
<li class="g"><b>52 / 150</b> failures resolved here, ~⅓, zero LLM spent</li>
<li class="g">Reproducible: same input → same verdict, every time</li>
</ul>
</div>
<div class="pane gen">
<h3 class="blue">◍ Grounded Agent Builder classifier</h3>
<ul>
<li class="b">RAG over stack trace, DOM diff, commit message, runner logs</li>
<li class="b">Reasons over <b>only the ambiguous band</b> (the other 98)</li>
<li class="b">Cites the evidence; tie-breaks toward <b>real_defect</b></li>
<li class="b">Runs on UiPath AI Trust Layer (PII redaction + audit)</li>
</ul>
</div>
</div>
<div class="card center" style="margin-top:20px">
<b>And a human approves every change.</b> <span class="sub" style="font-size:16px">The model proposes; it never mutates the suite on its own.</span>
</div>
<div class="foot"><span>The deterministic-vs-generative boundary is the headline AI design decision</span><span>4 / 12</span></div>
</section>
<!-- 5 · ARCHITECTURE -->
<section class="slide">
<span class="crit">Technical Execution</span>
<div class="kicker">Architecture</div>
<h2>One Maestro process. The right actor on the right failure.</h2>
<div class="flow" style="margin-top:30px">
<div class="node blue"><b>Test Cloud</b><span class="t">Test Manager results + execution history</span></div>
<div class="arrow">→</div>
<div class="node good"><b>Flake-scorer</b><span class="t">deterministic, auditable math</span></div>
<div class="arrow"><span>→</span><span class="lbl">ambiguous</span></div>
<div class="node"><b>Grounded<br>Classifier</b><span class="t">Agent Builder + RAG</span></div>
<div class="arrow">→</div>
<div class="node acc"><b>Governed<br>router</b><span class="t">verdict → action</span></div>
<div class="arrow">→</div>
<div class="branchwrap">
<div class="node" style="min-width:150px"><b class="amber">flaky</b><span class="t">Repair Agent drafts fix</span></div>
<div class="node" style="min-width:150px"><b class="accent">real_defect</b><span class="t">escalate</span></div>
<div class="node" style="min-width:150px"><b class="blue">environment</b><span class="t">re-run clean infra</span></div>
</div>
</div>
<div class="row" style="justify-content:center;gap:14px;margin-top:36px;align-items:center">
<span class="chip"><span class="dot" style="background:var(--good)"></span>confident cases auto-resolve</span>
<span class="arrow" style="min-width:30px">→</span>
<span class="chip" style="border-color:rgba(255,90,44,.5)"><span class="dot" style="background:var(--accent)"></span><b>Action Center: mandatory human review gate</b></span>
<span class="arrow" style="min-width:30px">→</span>
<span class="chip"><span class="dot" style="background:var(--blue)"></span>Orchestrator: quarantine / promote / open defect</span>
</div>
<p class="fine center" style="margin-top:26px">Three confidence bands, plus a defect cap: a selector change at the moment of breakage is treated as a real defect, even when the runtime looks jittery.</p>
<div class="foot"><span>Full data flow in <span class="mono">ARCHITECTURE.md</span></span><span>5 / 12</span></div>
</section>
<!-- 6 · LIVE VERDICTS -->
<section class="slide">
<span class="crit">Platform Usage · Technical Execution</span>
<div class="kicker">Live in UiPath Agent Builder</div>
<h2>Same agent, same grounding, three correct calls.</h2>
<div class="verdict-row verdict" style="margin-top:16px">
<div>
<div class="vhead"><span class="dot" style="background:var(--accent)"></span>real_defect · 0.95</div>
<div class="figframe"><img src="preview/verdict-real_defect.png" alt="real defect verdict"></div>
<div class="cap">Wrong total + renamed selector at the break. <b>Proposes no fix.</b> You don't heal a real bug.</div>
</div>
<div>
<div class="vhead"><span class="dot" style="background:var(--amber)"></span>flaky · 0.86</div>
<div class="figframe"><img src="preview/verdict-flaky.png" alt="flaky verdict"></div>
<div class="cap">Stale-element, no code change. <b>Proposes a fix</b> (re-locate + wait), but only as a proposal.</div>
</div>
<div>
<div class="vhead"><span class="dot" style="background:var(--blue)"></span>environment · 0.97</div>
<div class="figframe"><img src="preview/verdict-environment.png" alt="environment verdict"></div>
<div class="cap">503 from an upstream service with an open incident. <b>Re-run on clean infra</b>, blame neither test nor code.</div>
</div>
</div>
<p class="fine" style="margin-top:16px">Each verdict cites the evidence in its rationale. Published v1.0.0 and deployed as an Orchestrator process.</p>
<div class="foot"><span>UiPath Agent Builder (Studio Web), grounded + structured output</span><span>6 / 12</span></div>
</section>
<!-- 7 · SAFETY CONTRACT + RESULTS -->
<section class="slide">
<span class="crit">Business Impact · Technical Execution</span>
<div class="kicker">The safety contract, measured</div>
<h2>A real regression is never quarantined as flaky.</h2>
<div class="metrics" style="margin-top:22px">
<div class="metric"><div class="n">90.7%</div><div class="l">overall accuracy on a 150-case labeled corpus</div></div>
<div class="metric"><div class="n good">0.0%</div><div class="l"><b>safety false-positive rate</b><br>(no real defect ever hidden)</div></div>
<div class="metric"><div class="n amber">12%</div><div class="l">noise (flaky over-escalated as defect), the <i>safe</i> direction</div></div>
</div>
<div class="card" style="margin-top:22px">
<span class="sub" style="font-size:16.5px">
The 0% is <b class="good">enforced by a mechanism</b> (tie-break-to-defect + a flaky-band regression guard), not planted in the data.
Measured on a <b>synthetic-but-adversarial</b> corpus, not a production study.
<b>52 / 150</b> resolve deterministically; the grounded classifier carries the other <b>98</b>.
Run, not claimed: <span class="mono accent">python eval/harness.py</span> · gated by <span class="mono">eval/negative_control.py</span>.
</span>
</div>
<div class="foot"><span>Honest framing per <span class="mono">docs/limitations.md</span></span><span>7 / 12</span></div>
</section>
<!-- 8 · PLATFORM -->
<section class="slide">
<span class="crit">Platform Usage</span>
<div class="kicker">Running on the UiPath platform</div>
<h2>Built, grounded, orchestrated, and deployed on UiPath.</h2>
<div class="row" style="gap:22px;margin-top:14px">
<div style="flex:1.15">
<table>
<tr><th>Component</th><th>Role</th></tr>
<tr><td><b>Agent Builder</b></td><td>Hosts the grounded Triage Classifier + Repair Agent</td></tr>
<tr><td><b>Maestro</b></td><td>Orchestrates scorer → classifier → router → human gate</td></tr>
<tr><td><b>Context Grounding</b></td><td>Hybrid RAG over Test Manager artifacts</td></tr>
<tr><td><b>Action Center</b></td><td>Mandatory human review before any change</td></tr>
<tr><td><b>Orchestrator</b></td><td>Hosts the deployed solution; governed write-backs</td></tr>
<tr><td><b>AI Trust Layer</b></td><td>PII redaction + audit logging on every call</td></tr>
</table>
</div>
<div style="flex:1;display:flex;flex-direction:column;gap:12px">
<div>
<div class="figframe" style="height:210px"><img src="preview/maestro-run.png" alt="live Maestro run"></div>
<div class="cap"><b class="good">Live Maestro run, end-to-end Successful.</b> Agent classifies, the Verdict? gateway routes to the flaky branch.</div>
</div>
<div>
<div class="figframe" style="height:120px"><img src="preview/deployed-process.png" alt="deployed agent"></div>
<div class="cap">Deployed on Orchestrator, v1.0.0.</div>
</div>
</div>
</div>
<div class="foot"><span>Coded scorer + low-code Agent Builder agents, orchestrated by Maestro</span><span>8 / 12</span></div>
</section>
<!-- 9 · GOVERNANCE -->
<section class="slide">
<span class="crit">Business Impact · Governance</span>
<div class="kicker">Governance & the human gate</div>
<h2>The agent informs the decision. The human makes it.</h2>
<div class="two">
<div class="pane">
<ul>
<li><b>No autonomous mutation.</b> Every heal, quarantine, or baseline change is a proposal → Action Center.</li>
<li><b>Fail-safe default.</b> On low confidence the agent frames the failure as a defect, never silently quarantines.</li>
</ul>
</div>
<div class="pane">
<ul>
<li class="g"><b>Eval-driven release gate.</b> The agent can't publish below the accuracy + safety threshold.</li>
<li class="g"><b>Negative control.</b> 0 real defects hidden and 0 defects auto-healed across all 150 cases.</li>
</ul>
</div>
</div>
<div class="card center" style="margin-top:22px">
<span class="sub" style="font-size:17px">This is the line a QA lead has to trust: <b class="accent">the system escalates when it's unsure, instead of hiding.</b></span>
</div>
<div class="foot"><span>Aligns with "information, not automation": AI works <i>with</i> you, not instead of you</span><span>9 / 12</span></div>
</section>
<!-- 10 · CODING AGENT (BONUS) -->
<section class="slide">
<span class="crit">Platform Usage · Coding-Agent Bonus</span>
<div class="kicker">Built with a coding agent</div>
<h2>The whole build-and-deploy loop, from the terminal.</h2>
<div class="row" style="gap:22px;margin-top:12px">
<div style="flex:1.05;display:flex;flex-direction:column;gap:14px">
<ul>
<li class="b"><b>UiPath for Coding Agents:</b> a coding agent drove the <span class="mono">uip</span> CLI to deploy the agent, author the Maestro BPMN, and validate it.</li>
<li class="b"><span class="mono" style="font-size:14px">uip login · agent deploy · maestro bpmn registry / init / validate</span></li>
<li class="g">The eval loop caught a <b>real design flaw</b> (environment failures auto-classified as defects), lifting accuracy 70% → 90.7%.</li>
<li class="g">A self-review caught a <b>circular-corpus</b> problem, then de-rigged it: real engineering and honest methodology, not a one-shot generation.</li>
</ul>
</div>
<div style="flex:1">
<div class="figframe" style="height:330px"><img src="preview/coding-agent-cli.png" alt="uip CLI validate" style="object-position:center"></div>
<div class="cap"><span class="mono">uip maestro bpmn validate</span> → <b class="good">"Status": "Valid"</b>, 12 UiPath extensions, 1 process, 1 start event.</div>
</div>
</div>
<div class="foot"><span>Coded logic + low-code Agent Builder agent, orchestrated by Maestro · <span class="mono">docs/coding-agents.md</span></span><span>10 / 12</span></div>
</section>
<!-- 11 · LIMITATIONS -->
<section class="slide">
<span class="crit">Feasibility · Completeness</span>
<div class="kicker">Honest limitations & path to production</div>
<h2>What's real today, and what production needs.</h2>
<div class="two">
<div class="pane gen">
<h3 class="blue">Real today</h3>
<ul>
<li class="b">Live grounded agent classifying all three classes on UiPath</li>
<li class="b">Maestro orchestration running end-to-end, Successful</li>
<li class="b">Measured accuracy + a mechanism-enforced 0% safety rate</li>
<li class="b">Production-shaped architecture, governance, and eval methodology</li>
</ul>
</div>
<div class="pane">
<h3 class="accent">The gap is the data</h3>
<ul>
<li>Corpus is synthetic-but-adversarial: a solo builder can't ship a real enterprise's CI history</li>
<li>Production needs the live Test Manager results API + gold-standard labels</li>
<li>Then a shadow-mode prospective study to report real-world accuracy</li>
<li>Drift monitoring is scaffolded, not solved (named honestly)</li>
</ul>
</div>
</div>
<p class="fine center" style="margin-top:18px">The architecture, governance gates, and eval methodology are production-shaped. The data is the honest gap.</p>
<div class="foot"><span><span class="mono">docs/limitations.md</span></span><span>11 / 12</span></div>
</section>
<!-- 12 · CLOSE -->
<section class="slide center">
<div class="kicker">FlakeWarden</div>
<h1 style="font-size:40px;max-width:980px;margin:18px 0 8px;line-height:1.18">
Deterministic where it must be exact. Generative where the context is messy.
<span class="accent">A human in charge of every change.</span>
</h1>
<p class="sub" style="max-width:760px;margin-top:8px">
The agent, the orchestration, the labeled corpus, and the setup instructions are all on GitHub under MIT.
</p>
<div class="row center" style="gap:10px;margin-top:28px;flex-wrap:wrap;justify-content:center">
<span class="chip mono">github.com/JonathanSolvesProblems/flakewarden</span>
<span class="chip">UiPath AgentHack 2026 · Track 3</span>
</div>
<div class="foot center" style="justify-content:center"><span>Thank you · Jonathan Andrei</span></div>
</section>
</div>
<div class="hint">← → / Space to navigate · press <b>P</b> to print to PDF · <b>F</b> fullscreen</div>
<script>
const slides=[...document.querySelectorAll('.slide')];
let i=0;
function show(n){i=Math.max(0,Math.min(slides.length-1,n));
slides.forEach((s,k)=>s.classList.toggle('active',k===i));}
function fit(){const s=Math.min(innerWidth/1280,innerHeight/720)*0.97;
document.getElementById('stage').style.transform=`scale(${s})`;}
addEventListener('resize',fit); fit();
const h=parseInt((location.hash.match(/\d+/)||[1])[0],10); show((h||1)-1);
addEventListener('keydown',e=>{
if(['ArrowRight','ArrowDown',' ','PageDown'].includes(e.key)){show(i+1);e.preventDefault();}
else if(['ArrowLeft','ArrowUp','PageUp'].includes(e.key)){show(i-1);e.preventDefault();}
else if(e.key==='Home'){show(0);} else if(e.key==='End'){show(slides.length-1);}
else if(e.key==='p'||e.key==='P'){print();}
else if(e.key==='f'||e.key==='F'){if(!document.fullscreenElement)document.documentElement.requestFullscreen();else document.exitFullscreen();}
});
addEventListener('click',e=>{if(e.clientX<innerWidth*0.32)show(i-1);else show(i+1);});
</script>
</body>
</html>