Repository navigation
Expand file tree
/
Copy pathwork.html
More file actions
968 lines (521 loc) · 53.6 KB
/
Copy pathwork.html
File metadata and controls
968 lines (521 loc) · 53.6 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
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
<!DOCTYPE html>
<html lang="en"><head><meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>Ramona C. Truta — Work</title>
<meta name="description" content="AI evaluation, AI security, benchmarking, and data engineering. Where the work is public, the experiment is attached.">
<link rel="icon" type="image/x-icon" href="assets/favicon.ico">
<link rel="icon" type="image/png" sizes="32x32" href="assets/favicon-32x32.png">
<link rel="icon" type="image/png" sizes="16x16" href="assets/favicon-16x16.png">
<link rel="apple-touch-icon" sizes="180x180" href="assets/apple-touch-icon.png">
<meta property="og:site_name" content="Ramona C. Truta">
<meta property="og:type" content="website">
<meta property="og:title" content="Ramona C. Truta — Work">
<meta property="og:description" content="AI evaluation, AI security, benchmarking, and data engineering. Where the work is public, the experiment is attached.">
<meta property="og:image" content="https://ramonactruta.com/assets/photo.jpg">
<meta name="twitter:card" content="summary">
<meta name="twitter:title" content="Ramona C. Truta — Work">
<meta name="twitter:description" content="AI evaluation, AI security, benchmarking, and data engineering. Where the work is public, the experiment is attached.">
<meta name="twitter:image" content="https://ramonactruta.com/assets/photo.jpg">
<link rel="stylesheet" href="style.css?v=a1292bed75"></head><body>
<!-- Generated by tools/build_work.py from data/work.yaml. Do not edit by hand. -->
<nav class="topnav" data-pagefind-ignore><div class="wrap">
<a class="brand" href="index.html" title="Home" aria-label="Home">Ramona C. Truta</a>
<span class="navlinks">
<a href="work.html" class="here" data-tip="AI evaluation, AI security, benchmarking, and data engineering. Where the work is public, the experiment is attached.">Work</a>
<a href="teaching.html" data-tip="Teaching philosophy, course design, student testimonials, materials, and academic leadership.">Teaching</a>
<a href="speaking.html" data-tip="Conference talks and podcast appearances, as guest and as guest host.">Speaking</a>
<a href="writing.html" data-tip="Essays on data modeling, AI evaluation, security and the people using these systems. Each piece published with its repository when there is one.">Writing</a>
<a href="tags.html" data-tip="Every tag used on this site, and all the work it points to.">Index</a>
<a href="contact.html" data-tip="Open to consulting, contracting, project scoping & development, speaking engagements, research advisory, and conversations that don't have a job attached.">Contact</a>
<button type="button" class="searchbtn" aria-label="Search this site — every page, project, article and talk (⌘K)" data-tip="Search this site — every page, project, article and talk (⌘K)" data-search-open>
<svg viewBox="0 0 24 24" width="17" height="17" fill="none" stroke="currentColor"
stroke-width="2" stroke-linecap="round"><circle cx="11" cy="11" r="7"/><path d="M20 20l-4.2-4.2"/></svg>
</button>
</span>
</div></nav>
<nav class="subnav" id="top" data-pagefind-ignore><div class="wrap">
<a href="#agent-evaluation">AI evaluation</a>
<a href="#memory-security">AI security</a>
<a href="#benchmarking">Benchmarking</a>
<a href="#data-engineering">Data engineering</a>
</div></nav>
<main data-pagefind-body>
<section><div class="wrap">
<h2 id="agent-evaluation">AI evaluation</h2>
<p class="sub small cat-note">Whether a model or an agent did what it claimed, and what a deployment will cost. Instrumenting systems and reading what they actually did, rather than asking them. Relates to <a href="https://genai.owasp.org/llmrisk/llm06-excessive-agency/" target="_blank" rel="noopener">OWASP LLM06 (excessive agency)</a> and <a href="https://genai.owasp.org/llmrisk/llm09-misinformation/" target="_blank" rel="noopener">OWASP LLM09 (misinformation)</a>.</p>
<div class="proj" id="harness-bench">
<h3>harness-bench — measuring a deployment before it is a deployment</h3>
<p>I engineered <strong>harness-bench</strong> to evaluate agent deployments before they reach production. While skills and MCP servers are shipped widely, harness-bench parses tool surfaces, infers execution taxonomies, executes content-addressed YAML study contracts, traces every turn, and deterministically grades agent execution to report real token costs and failure modes — token cost per architecture, which guidance layers pay for themselves, which model tier suffices, and which states the harness is missing.</p>
<p>Four results it was built on, each measured rather than assumed:</p>
<ul>
<li><strong>Harness architecture dominates model choice.</strong> Decomposing a monolithic agent into specialists cut cost 20.7× at equal answer quality, and pulled five models across two vendors into a single 18–35K token band.</li><li><strong>Prompt guidance is the weakest lever.</strong> An AGENTS.md added roughly 50K tokens per run for no behavioral difference. Tool schemas and role prompts are what move frontier models.</li><li><strong>Words don’t bind; gates do.</strong> Every prompt-level instruction was eventually ignored by some model. Every mechanical gate held.</li><li><strong>Uniform cross-model failure is a harness defect, not a model defect</strong> — running a model ladder inside one contract makes that signature identify itself.</li>
</ul>
<p><strong>Skills are a variable it measures, not a feature it assumes.</strong> A study contract declares its cells — baseline against skills-injected — and the same goals run through both, so the cost and behaviour of a skill layer is a measurement rather than a claim in a README. That is what produced the Malloy Publisher result below.</p>
<p>The states it checks for were not designed in advance. Refusal, suspend and resume, a librarian role, resource caps — each was discovered by cheap probe runs and only then given a name.</p>
<p>The engine is generalized from a working implementation rather than designed on paper: the sqlbenchdag adapter carries its own graders, roles and tool surface, and the two runners — monolith and state machine — are the pair whose comparison produced the 20.7× result above. A new target is a new adapter against the same runners.</p>
<figure>
<img src="assets/sbd-state-machine.png" alt="State machine: librarian specialist branching to answered, refused, or config builder; config builder to the poller, refused, or config_builder_failed; poller to the analyzer or poll_failed; analyzer to complete or analyzer_failed">
<figcaption>The specialist state machine. A goal enters at the librarian, which can answer it from the existing catalog, refuse it as impossible, or send it on to build a fresh experiment. Every path ends in a named terminal state, including the failures — <code>refused</code>, <code>poll_failed</code>, <code>analyzer_failed</code> — so a run that goes wrong says where, rather than thrashing. Most of these states were discovered by probing, not designed up front.</figcaption>
</figure>
<figure>
<img src="assets/harness-gating-cost.svg" alt="Token totals per model for three goals run against a bare MCP surface and the same surface behind mechanical gates">
<figcaption>Three goals against the MotherDuck MCP surface, run bare and then behind mechanical gates. All twelve runs graded PASS in both arms, so the question is what gating costs — and the answer depends on the model. GPT-4o-mini spent 56% more tokens getting there; Claude Sonnet 5 spent 8% fewer. The cheap model pays for the guardrail.</figcaption>
</figure>
<div class="tags">
<a class="tag" href="tags.html#ai-evaluation">AI evaluation</a>
<a class="tag" href="tags.html#deterministic-grading">Deterministic grading</a>
<a class="tag" href="tags.html#per-turn-tracing">Per-turn tracing</a>
<a class="tag" href="tags.html#content-addressing">Content addressing</a>
<a class="tag" href="tags.html#mcp">MCP</a>
<a class="tag" href="tags.html#python">Python</a>
<a class="tag" href="tags.html#data-contracts">Data contracts</a>
<a class="tag" href="tags.html#pre-commit-gates">Pre-commit gates</a>
</div>
<div class="links">
<a href="https://github.com/rctruta/harness-bench" target="_blank" rel="noopener">harness-bench</a>
</div>
</div>
<div class="proj" id="malloy-study">
<h3>Do bundled agent skills actually reach the agent? — Malloy Publisher study</h3>
<p>The Malloy Publisher ships agent skills alongside its MCP server. Two questions: do those skills reach an autonomous agent at all, and what changes when they do. Eighteen runs — three goals × three models × baseline against skills-injected — with the full per-turn traces published and every table in the report derived from them by script. No number in it is hand-typed.</p>
<p><strong>Measured 15 July 2026</strong>, against <code>malloydata/publisher@ad6c09b</code>: the released image served four utility prompts rather than the twenty-eight skills then present on the project’s main branch. The skills existed; an agent pulling the official image did not get them.</p>
<p><strong>Fixed upstream, beginning 20 July 2026.</strong> The project shipped the skills as their own package (<a href="https://github.com/malloydata/publisher/pull/899" target="_blank" rel="noopener">#899</a>), then bundled each skill as an MCP prompt (<a href="https://github.com/malloydata/publisher/pull/944" target="_blank" rel="noopener">#944</a>, 29 July), and by 2 September added a manifest with tests holding every delivery channel to it (<a href="https://github.com/malloydata/publisher/pull/1088" target="_blank" rel="noopener">#1088</a>). The server’s own code now states the intent: hosts that only speak MCP receive the same guidance as hosts that read skill files directly. Skills on main today: 44.</p>
<p class="sub small">Both dates are checkable. Nothing here claims one caused the other. Re-verified 29 September 2026.</p>
<p>The second cuts both ways. Injecting the skills rescued the weaker models — one went from exhausting its turn budget after twenty turns, 123,191 tokens and seventeen query errors, to a correct answer in five turns, 31,658 tokens and one error. On the model that was already succeeding, the same skills raised token use roughly sixfold on that goal for the same correct answer. Guidance is not free, and it is worth most to whoever needs it most.</p>
<figure>
<img src="assets/malloy-skills-tokens.svg" alt="Tokens per run, baseline versus skills injected, across three goals and three models">
<figcaption>All 18 runs. Injecting the skills collapses the two models that were struggling — Haiku on the count goal from 123,191 tokens and an exhausted turn budget to 31,658 and a correct answer — and adds cost to Sonnet 5, which was already answering correctly. Derived from the per-turn traces in the study repository; n=1 per cell.</figcaption>
</figure>
<div class="tags">
<a class="tag" href="tags.html#ai-evaluation">AI evaluation</a>
<a class="tag" href="tags.html#ablation-study">Ablation study</a>
<a class="tag" href="tags.html#per-turn-tracing">Per-turn tracing</a>
<a class="tag" href="tags.html#mcp">MCP</a>
</div>
<div class="links">
<a href="https://github.com/rctruta/malloy-publisher-agent-study" target="_blank" rel="noopener">Study & all 18 traces</a>
</div>
</div>
<div class="proj" id="agent-traces">
<h3>Reading agent traces</h3>
<p>Per-turn tracing and deterministic grading of coding agents working against a commercial lakehouse platform. The traces surfaced three defects I reported upstream — including an error message that actively sends an agent in the wrong direction, suggesting <code>tanh</code> for a date function.</p>
<div class="tags">
<a class="tag" href="tags.html#ai-evaluation">AI evaluation</a>
<a class="tag" href="tags.html#per-turn-tracing">Per-turn tracing</a>
<a class="tag" href="tags.html#deterministic-grading">Deterministic grading</a>
<a class="tag" href="tags.html#python">Python</a>
<a class="tag" href="tags.html#process-telemetry">Process telemetry</a>
</div>
<div class="links">
<a href="https://github.com/rctruta/bauplan-agent-parse-study" target="_blank" rel="noopener">Study & traces</a>
<a href="https://github.com/BauplanLabs/bauplan/issues/369" target="_blank" rel="noopener">Issue #369</a>
<a href="https://github.com/BauplanLabs/bauplan/issues/370" target="_blank" rel="noopener">#370</a>
<a href="https://github.com/BauplanLabs/bauplan/issues/371" target="_blank" rel="noopener">#371</a>
</div>
</div>
<div class="proj" id="agent-telemetry">
<h3>agent-telemetry — transcript-derived process metrology for coding agents</h3>
<p>A local analytics package that parses raw JSONL session transcripts from coding agents (Claude Code) into un-opinionated process metrics: <code>tool_calls_per_turn</code>, <code>reads_before_first_edit</code>, <code>repeated_commands</code>, <code>error_rate</code>, <code>seconds_to_first_write</code>, and <code>edit_revisits</code>.</p>
<p>Prose instructions (in <code>CLAUDE.md</code> or prompt guidance) are hypotheses about agent behavior. By comparing session telemetry before and after introducing a rule, <code>agent-telemetry</code> evaluates whether a prompt-level instruction actually shifted model behavior. If a rule does not move its metric, the rule does not bind.</p>
<div class="tags">
<a class="tag" href="tags.html#ai-evaluation">AI evaluation</a>
<a class="tag" href="tags.html#process-telemetry">Process telemetry</a>
<a class="tag" href="tags.html#deterministic-grading">Deterministic grading</a>
<a class="tag" href="tags.html#python">Python</a>
</div>
<div class="links">
<a href="https://github.com/rctruta/agent-telemetry" target="_blank" rel="noopener">agent-telemetry</a>
</div>
</div>
<div class="proj" id="rag-accuracy">
<h3>A RAG system answering 17 of 20 questions its sources couldn't support</h3>
<p>I built a labelled evaluation set for a retrieval system I had written myself, and found it answering fluently, with citations, and with no error of any kind. Recalibrated the refusal threshold from 0.56 to 0.87 precision and validated out-of-sample. Confident output is not evidence; the failure mode to fear is the plausible one, not the obvious one.</p>
<p>Runs entirely from a local directory — LanceDB for vector, full-text and hybrid search, sentence-transformers for embeddings, no hosted cluster and no API key. Ask a question, get an answer and a link to the timestamp where it’s answered.</p>
<figure>
<img src="assets/rag-calibration.svg" alt="Left: retrieval precision before and after the refusal threshold moved. Right: F-beta of three candidate signals, with the bi-encoder cosine highest">
<figcaption>Left: moving the refusal threshold from 0.25 to 0.541 took precision from 0.564 to 0.870, with recall at 0.909 across 42 questions over 7,377 chunks. Right: the signal was chosen by trying three — a plain bi-encoder cosine beat both a cross-encoder and the two combined, and cross-validated at 0.83. The cheaper signal won, which is why it is worth measuring rather than assuming.</figcaption>
</figure>
<div class="tags">
<a class="tag" href="tags.html#ai-evaluation">AI evaluation</a>
<a class="tag" href="tags.html#evaluation-sets">Evaluation sets</a>
<a class="tag" href="tags.html#python">Python</a>
<a class="tag" href="tags.html#rag">RAG</a>
<a class="tag" href="tags.html#lancedb">LanceDB</a>
</div>
<div class="links">
<a href="https://github.com/rctruta/podcast-rag" target="_blank" rel="noopener">podcast-rag</a>
</div>
</div>
<div class="proj" id="music-recommender">
<h3>Music recommendation system — MIT Professional Education capstone</h3>
<p>Top-10 song recommendation over the Million Song Dataset Taste Profile Subset: two million listening events, ten thousand songs. Four algorithm families compared — user-user and item-item collaborative filtering, SVD matrix factorization, co-clustering, and content-based — with grid-search cross-validation over each.</p>
<p>Revisited in 2026, the interesting result is the evaluation rather than the models. Every one of the eight configurations reported Precision@10 of 1.0 while RMSE sat between 1.009 and 1.088, which cannot both be true of a working measurement. The relevance threshold was a play count of one, and the test set contains only songs a user already played — so every candidate counted as relevant and the metric reduced to k/k. Perfect scores produced by arithmetic, not by the recommenders.</p>
<div class="tags">
<a class="tag" href="tags.html#ai-evaluation">AI evaluation</a>
<a class="tag" href="tags.html#evaluation-sets">Evaluation sets</a>
<a class="tag" href="tags.html#python">Python</a>
</div>
<div class="links">
<a href="https://github.com/rctruta/music_recommendation_system" target="_blank" rel="noopener">music_recommendation_system</a>
</div>
</div>
<p class="totop"><a href="#top">↑ Top</a></p>
</div></section>
<section><div class="wrap">
<h2 id="memory-security">AI security</h2>
<p class="sub small cat-note">Attacks on what an agent reads, stores and later trusts — and the controls that hold against them. Maps to <a href="https://genai.owasp.org/llmrisk/llm01-prompt-injection/" target="_blank" rel="noopener">OWASP LLM01 (indirect prompt injection)</a> and <a href="https://genai.owasp.org/llmrisk/llm04-data-and-model-poisoning/" target="_blank" rel="noopener">OWASP LLM04 (data poisoning)</a>.</p>
<div class="proj" id="nodes26">
<h3>nodes26 — a controlled, reproducible graph memory lab</h3>
<p><strong>nodes26</strong> is a controlled, reproducible case study of one graph-backed agent-memory pipeline. A fictional catalog entry describes a desk lamp that does not exist. The lab uses the official <code>neo4j-agent-memory</code> ingestion and search APIs, pins the library version, and runs separate control, single-claim, repeated-copy, and independently worded-source arms in isolated databases. Each run preserves its experiment definition, graph records, retrieval scores, and agent tool traces so another reader can inspect the path from document to answer.</p>
<p>Across three runs — the library’s default schema, a procurement ontology, and the library’s example ontology — the fictional lamp became a typed, retrievable entity and appeared in later agent answers. Strict type validation did not establish whether the catalog claim was true. Repeating the claim added source links but did not improve that entity’s similarity score.</p>
<p>The lab also exposed a break in the agent’s evidence-checking path, reported in <a href="https://github.com/neo4j-labs/agent-memory/issues/392" target="_blank" rel="noopener">neo4j-labs/agent-memory issue #392</a>. <code>search_memory</code> gives the agent entity names without IDs, while <code>get_entity_provenance</code> requires a UUID and silently answers <code>Sources (0)</code> for an unknown one. Default ingestion writes <code>MENTIONS</code> links, but that lookup reads <code>EXTRACTED_FROM</code>, which default ingestion does not write. In each of two runs where source links <em>were</em> written, 16 of 17 agent provenance calls returned no sources. Agent answers then said there was no provenance, although the graph held four source links; the one call supplied a real ID returned all four. The failure is in carrying evidence through the tools to the agent, with an empty result presented as if it were an authoritative check. These runs establish that behavior for this pipeline and these inputs; they do not measure a general failure rate or test a fix.</p>
<div class="tags">
<a class="tag" href="tags.html#ai-security">AI security</a>
<a class="tag" href="tags.html#ai-evaluation">AI evaluation</a>
<a class="tag" href="tags.html#ablation-study">Ablation study</a>
<a class="tag" href="tags.html#knowledge-graphs">Knowledge graphs</a>
<a class="tag" href="tags.html#python">Python</a>
</div>
<div class="links">
<a href="https://github.com/rctruta/nodes26" target="_blank" rel="noopener">nodes26 — experiment, records and analysis</a>
<a href="https://github.com/neo4j-labs/agent-memory/issues/392" target="_blank" rel="noopener">Agent source-verification failure — issue #392</a>
</div>
</div>
<div class="proj" id="agent-memory">
<h3>Semantic injection into agent memory</h3>
<p>A deterministic exploit against an open-source agent memory system: text that reads as ordinary content to a human, but that the extraction pipeline turns into a stored fact the agent later treats as its own knowledge. Disclosed responsibly. Separately, a forensic read of a coding agent’s local state directory — what it writes to disk, in plaintext, without telling you.</p>
<div class="tags">
<a class="tag" href="tags.html#ai-security">AI security</a>
<a class="tag" href="tags.html#ai-evaluation">AI evaluation</a>
</div>
<div class="links">
<a href="https://ramonactruta.substack.com/p/the-blueberry-muffin-exploit-semantic" target="_blank" rel="noopener">The Blueberry Muffin Exploit</a>
<a href="https://ramonactruta.substack.com/p/inside-claude-structure-is-not-security" target="_blank" rel="noopener">Inside ~/.claude/</a>
</div>
<div class="note">Authorised researcher, Anthropic Cyber Verification Program.</div>
</div>
<div class="proj" id="consensus-contagion">
<h3>Consensus Contagion — status bias and the RLHF tax in agent routing</h3>
<p>Based on my <strong>adversarial-judgement-research</strong> framework (14 study contracts, 205 graded runs across 10 models from 4 vendors, with a successor testbed extending the matrix to local models): I proved that generalized LLM orchestrators evaluate simulated social status over technical merit. Injecting a high-status persona into routing decisions causes agentic state to collapse into authority bias (the RLHF tax). Prompt-level instructions fail under pressure; mechanical gates hold.</p>
<div class="tags">
<a class="tag" href="tags.html#ai-evaluation">AI evaluation</a>
<a class="tag" href="tags.html#deterministic-grading">Deterministic grading</a>
<a class="tag" href="tags.html#ablation-study">Ablation study</a>
<a class="tag" href="tags.html#rlhf">RLHF</a>
<a class="tag" href="tags.html#data-contracts">Data contracts</a>
<a class="tag" href="tags.html#ai-security">AI security</a>
<a class="tag" href="tags.html#python">Python</a>
</div>
<div class="links">
<a href="https://github.com/rctruta/adversarial-judgement-research" target="_blank" rel="noopener">adversarial-judgement-research</a>
</div>
<div class="note">Presented at the Canadian Women in Cybersecurity Conference, Toronto, May 2026.</div>
<div class="note">Experiments varying a protected attribute are held back, with their findings: a single name pair cannot separate a name-specific effect from a systematic one. The published README says so rather than leaving a gap.</div>
</div>
<div class="proj" id="security-testbed">
<h3>AI security testbed — semantic privilege escalation under adversarial audit</h3>
<p>The successor to <strong>adversarial-judgement-research</strong>, built to test one premise: that
wrapping a model in a system prompt does not make a multi-agent pipeline secure, because
the control plane and the data plane are both just text. A transformer has no role-based
access control, so an instruction and a social authority payload arrive the same way.</p>
<p>A deterministic matrix engine runs a 2x2 factorial at temperature 0 with replication,
sweeping model, adversary status and thinking budget, and grades every run by rule. The
corpus is 416 traces across frontier and local models — GPT-4.1, Gemini 3.1 Pro, Claude
Opus 4.6, Cohere Command-R, Groq, HuggingFace, and Ollama running llama3, mistral and
gemma2 on my own machine. Running the same scenarios locally is what makes the sweep
affordable to replicate rather than a one-off purchase.</p>
<p>The engine, the study contracts and every trace are public. My own working notes are
not: those are research in progress, and they are the only thing held back.</p>
<div class="tags">
<a class="tag" href="tags.html#ai-security">AI security</a>
<a class="tag" href="tags.html#ai-evaluation">AI evaluation</a>
<a class="tag" href="tags.html#deterministic-grading">Deterministic grading</a>
<a class="tag" href="tags.html#ablation-study">Ablation study</a>
<a class="tag" href="tags.html#content-addressing">Content addressing</a>
<a class="tag" href="tags.html#python">Python</a>
</div>
<div class="links">
<a href="https://github.com/rctruta/ai-security-testbed" target="_blank" rel="noopener">ai-security-testbed</a>
</div>
<div class="note">Measurements are dated: they describe the model versions named in each study contract, at the time each study ran. Model behaviour changes; an undated claim about a model ages into a false one.</div>
</div>
<div class="proj" id="agent-tools">
<h3>Tools and playbooks for collaborating with coding agents</h3>
<p><strong>ai-agent-utils</strong> — a bootstrap that starts a repository with gates already enforced, rather than rules written down and hoped for. Two families: git discipline, and claim integrity — every load-bearing number registered with a source and re-checked before commit, fabricated claims blocked by pattern, and a checkpoint that refuses to record a working state unless a proof command passes. Enforcement is git hooks and stdlib scripts, so it binds whichever agent is driving; the Claude and Cursor files only describe it. <strong>The Pedantic Medallion</strong> — community-curated playbooks for building shared understanding, from the framework of the same name.</p>
<div class="tags">
<a class="tag" href="tags.html#ai-security">AI security</a>
<a class="tag" href="tags.html#deterministic-grading">Deterministic grading</a>
<a class="tag" href="tags.html#python">Python</a>
<a class="tag" href="tags.html#pre-commit-gates">Pre-commit gates</a>
</div>
<div class="links">
<a href="https://github.com/rctruta/ai-agent-utils" target="_blank" rel="noopener">ai-agent-utils</a>
<a href="https://github.com/rctruta/awesome-pedantic-medallion" target="_blank" rel="noopener">awesome-pedantic-medallion</a>
<a href="https://ramonactruta.substack.com/p/a-readme-is-a-set-of-falsifiable" target="_blank" rel="noopener">A README is a Set of Falsifiable Claims</a>
</div>
</div>
<p class="totop"><a href="#top">↑ Top</a></p>
</div></section>
<section><div class="wrap">
<h2 id="benchmarking">Benchmarking</h2>
<p class="sub small cat-note">Measuring engines and protocols under conditions stated well enough to reproduce. Every result ships with the experiment that produced it.</p>
<div class="proj" id="sqlbenchdag">
<h3>sqlbenchdag — a reproducible benchmarking laboratory</h3>
<p>Every experiment is a capsule addressed by an 8-character SHA-256 fingerprint over its config, SQL, and every line of measurement-relevant Python. Change the method and the ID changes. Cold-cache execution, hard row-count assertions, and an integrity seal on every capsule. The four published Quack capsules additionally carry OpenTimestamps proofs anchored to Bitcoin.</p>
<figure>
<img src="assets/quack-vs-postgres.png" alt="DuckDB-over-Quack vs PostgreSQL, query duration by table size">
<figcaption>Capsule <code>902d1277</code> — DuckDB-over-Quack (pushdown) against PostgreSQL: 4.4× at 100K rows, 6.1× at 1M, 13.2× at 10M. Bands are replication min–max, cold cache, arm64, 8 cores, DuckDB 1.5.3. Caveat disclosed in the config: PostgreSQL pays a macOS Docker-VM tax on this bench.</figcaption>
</figure>
<figure>
<img src="assets/quack-execution-modes.png" alt="Quack attach mode vs pushdown vs in-process DuckDB">
<figcaption>Capsule <code>b8e2bfaf</code> — attach-mode overhead grows with scan size (2.6× at 100K → 9.5× at 10M); pushdown stays flat at ~2×. Attach mode streams table data client-side; pushdown ships only results. The residual 2× traced to reduced server-side parallelism, not transport (capsule <code>25b0e134</code>).</figcaption>
</figure>
<div class="tags">
<a class="tag" href="tags.html#benchmarking">Benchmarking</a>
<a class="tag" href="tags.html#content-addressing">Content addressing</a>
<a class="tag" href="tags.html#cold-cache-measurement">Cold-cache measurement</a>
<a class="tag" href="tags.html#dagster">Dagster</a>
<a class="tag" href="tags.html#duckdb">DuckDB</a>
<a class="tag" href="tags.html#postgresql">PostgreSQL</a>
<a class="tag" href="tags.html#polars">Polars</a>
<a class="tag" href="tags.html#arrow-adbc">Arrow / ADBC</a>
<a class="tag" href="tags.html#python">Python</a>
<a class="tag" href="tags.html#mcp">MCP</a>
<a class="tag" href="tags.html#sql">SQL</a>
<a class="tag" href="tags.html#data-contracts">Data contracts</a>
<a class="tag" href="tags.html#pre-commit-gates">Pre-commit gates</a>
<a class="tag" href="tags.html#pypi">PyPI</a>
</div>
<div class="links">
<a href="https://github.com/rctruta/sql-benchmarks-dagster" target="_blank" rel="noopener">sql-benchmarks-dagster</a>
<a href="https://pypi.org/project/sqlbenchdag/" target="_blank" rel="noopener">pip install sqlbenchdag</a>
<a href="https://ramonactruta.substack.com/p/measuring-quack-duckdbs-new-client" target="_blank" rel="noopener">Write-up</a>
</div>
<div class="recognition">
<span class="recognition-who">mcprush</span>
<span class="recognition-date">2026-10-20</span>
found the MCP server in the official registry, scanned the published artifact and graded it: security A, 94/100, on v0.1.3 —
<a href="https://mcprush.com/rctruta/sql-benchmarks-lab-mcp" target="_blank" rel="noopener">source</a>. Their scan of a snapshot, under their own methodology. Not a certification.
<a class="recognition-badge" href="https://mcprush.com/rctruta/sql-benchmarks-lab-mcp" target="_blank" rel="noopener"><img src="https://mcprush.com/badge/sql-benchmarks-lab-mcp/card.svg" alt="mcprush listing for SQL Benchmarks Lab MCP server: security A, trust score 94 out of 100" loading="lazy" width="380" height="200"></a>
</div>
<div class="note"><strong><a href="https://pypi.org/project/sqlbenchdag/" target="_blank" rel="noopener">Published Python package on PyPI</a></strong> — reproducible benchmarking laboratory. Install with <code>pip install sqlbenchdag</code>.</div>
<div class="note"><strong><a href="https://registry.modelcontextprotocol.io/v0/servers?search=sqlbenchdag" target="_blank" rel="noopener">MCP server in the official Model Context Protocol registry</a></strong> — v0.1.2, 10 tools, active since 2026-08-10. Run it with <code>uvx --from sqlbenchdag sqlbench-mcp</code>: the package name is not the command, and <code>uvx sqlbenchdag</code> exits 1.</div>
<div class="note"><strong><a href="https://github.com/rctruta/sql-benchmarks-dagster/blob/main/.github/workflows/published_mcp.yml" target="_blank" rel="noopener">One command, cold, with nothing else running</a></strong> — the published server reads the committed capsules directly, so it needs no REST API, no config and no key. A scheduled CI job installs it from PyPI on a clean machine weekly and fails if the handshake or any tool description breaks.</div>
</div>
<div class="proj" id="sql-behaviour">
<h3>SQL behaviour under measurement</h3>
<p>Findings from the laboratory that are about the database rather than the benchmark. Postgres index selectivity has a crossover: on cold cache a bitmap heap scan at 5% selectivity is slower than a sequential scan, so the index becomes a liability — EXPLAIN-confirmed, with the sequential scan winning outright past roughly 10%. Separately, on three-valued logic: the null-safe <code>IS NOT DISTINCT FROM</code> operator does not remove the quadratic cost of a join over nullable keys. It decides whether the query degrades or hangs. That is an optimizer cliff in Postgres, not something intrinsic to the semantics.</p>
<div class="tags">
<a class="tag" href="tags.html#benchmarking">Benchmarking</a>
<a class="tag" href="tags.html#cold-cache-measurement">Cold-cache measurement</a>
<a class="tag" href="tags.html#data-modeling">Data modeling</a>
<a class="tag" href="tags.html#postgresql">PostgreSQL</a>
<a class="tag" href="tags.html#duckdb">DuckDB</a>
<a class="tag" href="tags.html#sql">SQL</a>
</div>
<div class="links">
<a href="https://ramonactruta.substack.com/p/the-pizza-logic-you-didnt-know-you" target="_blank" rel="noopener">The Pizza Logic You Didn't Know You Kneaded</a>
</div>
</div>
<div class="proj" id="result-transport">
<h3>What the connectivity layer costs — ADBC, connectorx and psycopg out of PostgreSQL</h3>
<p>One question, asked properly: when a client pulls a large result out of PostgreSQL, what does the <strong>transport</strong> cost, separately from the query? Same server, same SQL, trivial <code>SELECT *</code> so execution does not dominate — only the client varies: psycopg fetching rows, ADBC over Arrow, connectorx over Arrow. Timed from execute to the result fully materialised in memory.</p>
<p><strong>The answer is not "Arrow is faster."</strong> On primitive columns the two paths are level and at 100K rows psycopg is marginally ahead (0.88×). What changes the picture is column type. On <code>jsonb</code> at 10 million rows psycopg takes 143 seconds against ADBC’s 20.8 — <strong>6.9×</strong> — and on a mixed nested column 165 seconds against 29.9. The row path does not scale because it materialises a Python object per value; Arrow stays linear. The penalty widens with size exactly where that cost lives: jsonb goes 2.70× at 100K, 3.42× at 1M, 6.88× at 10M, while primitives sit at 1.22× throughout.</p>
<p>The two Arrow clients are close but not interchangeable: ADBC wins <code>jsonb</code> at 10M (20.8s vs 35.0s), connectorx wins <code>int[]</code> (11.4s vs 12.2s). I shared the results with Ian Cook at Voltron Data, whose Arrow work prompted the question.</p>
<figure>
<img src="assets/transport-row-path-collapse.svg" alt="Row-path penalty by column type at 100K, 1M and 10M rows: primitives stay near 1x while jsonb reaches 6.88x">
<figcaption>Capsule <code>af6ba593</code>, generated from the capsule's own CSV. psycopg divided by the faster Arrow client. Replication 3, cold cache, Darwin arm64, 16 GB, <code>sqlbenchdag@9105c38</code>.</figcaption>
</figure>
<div class="tags">
<a class="tag" href="tags.html#benchmarking">Benchmarking</a>
<a class="tag" href="tags.html#cold-cache-measurement">Cold-cache measurement</a>
<a class="tag" href="tags.html#content-addressing">Content addressing</a>
<a class="tag" href="tags.html#arrow-adbc">Arrow / ADBC</a>
<a class="tag" href="tags.html#postgresql">PostgreSQL</a>
<a class="tag" href="tags.html#python">Python</a>
<a class="tag" href="tags.html#sql">SQL</a>
</div>
<div class="links">
<a href="https://github.com/rctruta/sql-benchmarks-dagster/tree/main/sql_benchmarks/experiments/results/af6ba593" target="_blank" rel="noopener">capsule af6ba593</a>
<a href="https://github.com/rctruta/sql-benchmarks-dagster/blob/main/sql_benchmarks/experiments/queue/transport_adbc_scale.yaml" target="_blank" rel="noopener">the experiment config</a>
</div>
<div class="note">Exploratory tier, not in the verified published set: two capsules on one machine, replication 3. It is a shape, not a benchmark result to quote against someone else’s hardware — compare the ratios, not the seconds.</div>
<div class="note">A day’s work, and it leaned on the laboratory already being there. The value is that the harness made a proper cold-cache, sealed, content-addressed run cheap enough to do in an afternoon.</div>
<div class="note">Says nothing about Arrow IPC or Flight as protocols, nor about ADBC drivers in general — one alpha Quack driver was measured separately and is not this result.</div>
</div>
<div class="proj" id="lakehouse-semantics">
<h3>lakehouse-semantics — SQL correctness and engine divergence matrix</h3>
<p>A suite of 23 correctness probes executed across DuckDB, DuckDB+Delta, PostgreSQL, and Databricks SQL. Rather than timing performance (which varies by serverless hardware allocation), these probes measure <strong>semantic determinism</strong>: integer division, NULL ordering in ASC scans, VARCHAR overflow behavior, and constraint enforcement paths.</p>
<p>The key finding: Databricks SQL accepts inline <code>CHECK</code> constraints in <code>CREATE TABLE</code> without enforcing them, while the identical constraint added via <code>ALTER TABLE ADD CONSTRAINT</code> is strictly enforced and raises <code>DELTA_VIOLATE_CONSTRAINT_WITH_VALUES</code>. A query 20% slower costs money; a query that silently returns different rows across engines costs trust.</p>
<figure>
<img src="assets/lakehouse-semantics-probes.svg" alt="Seven SQL semantics probes run on DuckDB, PostgreSQL and Databricks SQL, with the engine that disagrees with the other two marked on each row">
<figcaption>Seven of the 23 probes. The marked cell on each row is the engine that disagrees with the other two — DuckDB on three, Databricks on three, PostgreSQL on one. Strictness reverses by probe: Databricks raises on <code>1/0</code> where DuckDB returns <code>inf</code>, and accepts a duplicate <code>PRIMARY KEY</code> where the others reject it. That last one is <a href="https://docs.databricks.com/aws/en/tables/constraints">documented behaviour</a>, not a discovery — PK, FK and UNIQUE are informational there. <code>duckdb+delta</code> matched <code>duckdb</code> on all 23, which is the control arm that makes the Databricks column readable.</figcaption>
</figure>
<div class="tags">
<a class="tag" href="tags.html#benchmarking">Benchmarking</a>
<a class="tag" href="tags.html#postgresql">PostgreSQL</a>
<a class="tag" href="tags.html#duckdb">DuckDB</a>
<a class="tag" href="tags.html#sql">SQL</a>
<a class="tag" href="tags.html#python">Python</a>
</div>
<div class="links">
<a href="https://github.com/rctruta/lakehouse-semantics" target="_blank" rel="noopener">lakehouse-semantics</a>
</div>
</div>
<p class="totop"><a href="#top">↑ Top</a></p>
</div></section>
<section><div class="wrap">
<h2 id="data-engineering">Data engineering</h2>
<p class="sub small cat-note">Data modeling is the throughline. Source systems into structures analysts and models can use — migration, entity resolution, record linkage, annotation quality control, and the SQL behaviour underneath it.</p>
<div class="proj" id="guelph-migration">
<h3>Legacy-to-PostgreSQL migration and record linkage — University of Guelph, 2011–2015</h3>
<p>Sole architect of the migration of a historical research database to PostgreSQL: the data model cut storage by 73%. The result was a production system with researchers as its direct clients — requirements gathering, stakeholder translation, and ongoing support were part of the job, not adjacent to it.</p>
<p>On top of it, probabilistic record linkage connecting WWI military service records to census data: 97% linkage across 450,000 records at 89% precision. Blocking and scoring were only half the problem; the harder half was validation, so I automated family re-creation to check links against household structure, which raised verification speed fourfold.</p>
<p>Two properties of that work carry into anything with dirty source data: the ground truth is never handed to you, and a linkage rate means nothing without the precision beside it.</p>
<p>The coding work was not a solo exercise. I led a team of twenty research assistants doing the manual verification, and built the monitoring that made their work checkable at the level of a single record — one line, in one file, among a great many. The scarce resource was their attention, so the system’s job was to spend it only where a disagreement actually existed.</p>
<div class="tags">
<a class="tag" href="tags.html#data-modeling">Data modeling</a>
<a class="tag" href="tags.html#entity-resolution">Entity resolution</a>
<a class="tag" href="tags.html#data-quality">Data quality</a>
<a class="tag" href="tags.html#migration">Migration</a>
<a class="tag" href="tags.html#postgresql">PostgreSQL</a>
<a class="tag" href="tags.html#team-leadership">Team leadership</a>
</div>
</div>
<div class="proj" id="name-coding">
<h3>Automating a hand-built name dictionary — first-name standardization across seven censuses</h3>
<p>Historians had been coding first names by hand: each name assigned a three- to five-letter code standing for an equivalence class of spellings. It did not scale, and every new census arrived with a fresh tail of names nobody had seen. I automated it, starting from their 1871 dictionary as the seed.</p>
<p>The equivalence relation was phonetic and orthographic together — double metaphone for blocking, then minimum Levenshtein distance within the block. A second variant added substring substitution, which is what made French spellings tractable. Two phases: generate candidate codes over a sample of 70,000 strings, then apply the replacement rules to whatever was left. Strings that survived both were reported as unassigned rather than forced.</p>
<p>The finding that shaped the design: <strong>the expert dictionary contradicted itself.</strong> Two strings sharing a code were not always equivalent under any formal definition, and strings with different codes sometimes were. So instead of emitting one answer, the system emitted several candidates and let the experts choose — and some strings the experts themselves coded more than one way. When the ground truth disagrees with itself, widening the output is more honest than picking a winner.</p>
<p>Every candidate carried a verdict, good or bad, which made the system measurable. Twenty research assistants did that checking — each code looked at independently by two people, with disagreements adjudicated rather than averaged. <strong>237,203 codes were checked by hand out of a corpus of 1,367,022 strings: 90.96% of them good, 95.29% setting aside a class of special cases.</strong> I checked the other direction too — whether <em>wrong</em> codes also sit at minimum distance — since a metric that scores well but cannot discriminate is worth nothing.</p>
<p>The 17% that needed a person is the number I would defend the design on. Across 63 batches the median was 92.5%, and the spread ran from 97.2% down to 34.0% on one small batch — which is why the aggregate is the honest figure and a range is not.</p>
<p>The coded names carried forward across the 1852, 1871, 1880, 1881, 1891, 1901 and 1911 censuses, with a full ingest pipeline per census: process, detect inconsistencies, apply coding, prepare for the model downstream.</p>
<div class="tags">
<a class="tag" href="tags.html#data-modeling">Data modeling</a>
<a class="tag" href="tags.html#metadata-architecture">Metadata architecture</a>
<a class="tag" href="tags.html#entity-resolution">Entity resolution</a>
<a class="tag" href="tags.html#evaluation-sets">Evaluation sets</a>
<a class="tag" href="tags.html#data-quality">Data quality</a>
<a class="tag" href="tags.html#postgresql">PostgreSQL</a>
<a class="tag" href="tags.html#python">Python</a>
</div>
</div>
<div class="proj" id="age-modeling">
<h3>Fitting the agreement function for record linkage — age difference across censuses</h3>
<p>Two records ten years apart rarely agree on age. To link them you need a function turning that disagreement into evidence: <code>ageDiff</code> mapping an age difference onto a weight between one and zero. I fitted five candidates — normalized by constant factors, by frequency percentage, and by a fast-increasing function — and chose between them on standard deviation against observed frequencies, across five independent datasets and 11,813 records.</p>
<p>I also tested the received wisdom that women misreport age more often than men. Running the regression separately by gender, the difference was slight: both curves were the same function shifted by about a year. Not two models. One model and an offset — which is the result you only get by checking rather than assuming.</p>
<div class="tags">
<a class="tag" href="tags.html#data-modeling">Data modeling</a>
<a class="tag" href="tags.html#entity-resolution">Entity resolution</a>
<a class="tag" href="tags.html#evaluation-sets">Evaluation sets</a>
<a class="tag" href="tags.html#data-quality">Data quality</a>
</div>
</div>
<div class="proj" id="police-gazette">
<h3>Getting convict transport records out of OCR — Police Gazette</h3>
<p>Nineteenth-century Police Gazettes recording prisoners transported to Tasmania, held as OCR’d scans. The data the historians needed was in tables inside those documents, and OCR does not respect a table. I wrote the extraction design: fixed layouts for the two table types, one table per output file, a naming convention tying every file back to its source page, and a separate metadata table holding what is true of a whole file — its description, its row count, who extracted it, who checked it — recorded once instead of repeated on every row.</p>
<p>My proposal secured the grant to carry out the work: one table per file, so extraction and checking could run independently and be verified in small units, and a metatable recording each file’s facts once — including who extracted it, who checked it, and the row count.</p>
<p>Reading it back in 2026: the metatable is a metadata catalog, the filename scheme is a canonical identifier carrying lineage, and <code>ExtractedBy</code> / <code>CheckedBy</code> / <code>RowCount</code> are a provenance trail. I had none of those words in 2011.</p>
<div class="tags">
<a class="tag" href="tags.html#data-modeling">Data modeling</a>
<a class="tag" href="tags.html#metadata-architecture">Metadata architecture</a>
<a class="tag" href="tags.html#data-quality">Data quality</a>
<a class="tag" href="tags.html#evaluation-sets">Evaluation sets</a>
</div>
<div class="links">
<a href="assets/police-gazette-transcripts-2011.pdf" target="_blank" rel="noopener">The 2011 design document (PDF)</a>
</div>
</div>
<div class="proj" id="db2-postgres">
<h3>Departmental migration from IBM DB2 to PostgreSQL — University of Toronto</h3>
<p>Led the Computer Science department’s move off DB2: platform evaluation, rewriting the course materials that depended on it, retraining instructors, and supporting students through the change. A migration whose hardest constraint was that several hundred people had to keep working through it.</p>
<div class="tags">
<a class="tag" href="tags.html#migration">Migration</a>
<a class="tag" href="tags.html#postgresql">PostgreSQL</a>
<a class="tag" href="tags.html#team-leadership">Team leadership</a>
<a class="tag" href="tags.html#teaching">Teaching</a>
</div>
</div>
<div class="proj" id="site-infrastructure">
<h3>rctruta.github.io — contract-driven portfolio & validated publishing pipeline</h3>
<p>The repository powering this website is built as a data modeling project. Public content lives in YAML and is validated against JSON Schemas before rendering. Unknown fields, missing required values, duplicate IDs and invalid taxonomy references stop the build. A shared taxonomy supplies stable concept IDs, display labels and published anchors to the pages and their index.</p>
<p>Python generators render the pages, and Pagefind builds search. After generation, checks validate local links, anchors and index membership for each term. Local hooks and CI run declared checks; a privacy guard checks staged content for email addresses, phone numbers and explicit private-file markers. These checks enforce document structure and consistency. They do not independently verify factual claims in prose.</p>
<div class="tags">
<a class="tag" href="tags.html#data-modeling">Data modeling</a>
<a class="tag" href="tags.html#metadata-architecture">Metadata architecture</a>
<a class="tag" href="tags.html#pre-commit-gates">Pre-commit gates</a>
<a class="tag" href="tags.html#python">Python</a>
<a class="tag" href="tags.html#data-contracts">Data contracts</a>
</div>
<div class="links">
<a href="https://github.com/rctruta/rctruta.github.io" target="_blank" rel="noopener">rctruta.github.io</a>
</div>
</div>
<p class="totop"><a href="#top">↑ Top</a></p>
</div></section>
</main>
<footer data-pagefind-ignore><div class="wrap"><span>© 2025–2026 Ramona C. Truta · <a href="https://creativecommons.org/licenses/by-nc-sa/4.0/" rel="license">CC BY-NC-SA 4.0</a> · <a href="contact.html">Contact</a></span></div></footer>
<!-- search:start -->
<div class="searchmodal" id="searchmodal" hidden data-pagefind-ignore>
<div class="searchmodal__panel" role="dialog" aria-modal="true" aria-label="Search this site">
<button type="button" class="searchmodal__close" aria-label="Close search" data-search-close>×</button>
<div id="search"></div>
</div>
</div>
<link href="pagefind/pagefind-ui.css" rel="stylesheet">
<script src="pagefind/pagefind-ui.js"></script>
<script>
(() => {
const modal = document.getElementById('searchmodal');
let ui = null;
const open = () => {
modal.hidden = false;
document.body.style.overflow = 'hidden';
if (!ui) {
ui = new PagefindUI({
element: '#search',
showSubResults: true,
showImages: false,
pageSize: 8,
highlightParam: "highlight",
translations: {
placeholder: 'Search this site',
zero_results: 'Nothing found for [SEARCH_TERM]'
}
});
}
const saved = sessionStorage.getItem('pagefind_query');
const input = modal.querySelector('input');
if (input) {
if (saved && !input.value) {
ui.triggerSearch(saved);
}
input.focus();
}
};
const close = () => {
modal.hidden = true;
document.body.style.overflow = '';
};
document.querySelectorAll('[data-search-open]').forEach(b => b.addEventListener('click', open));
document.querySelectorAll('[data-search-close]').forEach(b => b.addEventListener('click', close));
modal.addEventListener('input', e => {
if (e.target && e.target.tagName === 'INPUT') {
sessionStorage.setItem('pagefind_query', e.target.value);
}
});
modal.addEventListener('click', e => {
if (e.target === modal) close();
const link = e.target.closest('a');
if (link && link.getAttribute('href')) {
close();
}
});
document.addEventListener('keydown', e => {
if (e.key === 'Escape' && !modal.hidden) close();
if ((e.metaKey || e.ctrlKey) && e.key === 'k') { e.preventDefault(); modal.hidden ? open() : close(); }
});
})();
</script>
<!-- search:end -->
</body></html>