Blog
The studio's working log: we discuss topics chosen from the news with different AI models, and write what we recommend and what data would change our mind.
Topics chosen from the news are worked through a Spark round: the same question is asked of different AI models that can't see one another's answers; the sources they give are checked one by one, and where they agree and diverge is written up.
Every post has three things: method, references and the data that would change our mind.
Posts · 15 (6 Proof Notes · 9 Essays)
◆ Spark verdict · ends with a clear verdict◎ Synthesis▤ Essay▣ Review
This article explains how to measure a website from three angles: classic SEO, GEO (visibility in AI assistants) and overall quality. For each criterion we ask two questions: Why does it matter? How do you check it?
Measured narrative of how the frozen October 19, 2026 Spark round turned into Atlas App in 16.5 hours (completed by the 00:59→04:24 overnight chain), uniting Claude Fable and Gemini Argon in one window under Grok supervision.
Why literary persona prompts add zero factual accuracy (Δscore = 0.0000 across 162 trials), how SHA-256 identity sealing and ≤4 KB watchdog packets stop context drift, and how the October 17 reversal shaped the Claude + Gemini + Grok division of labour.
How a six-gate quantitative research pipeline prevents specification gaming: why Round-1 winner QH-06 collapsed from AUC 0.6028 to 0.5782 (50.31% coin flip) under causal Dimson ±15 min beta and 5-minute Epps aggregation, audited across 110,001 bars.
Inside the 7-step Spark protocol and LLM red team workflow: how 10 blind external voices and Author S0 audited 420 citations, exposed correlated hallucinations across local 27B–31B models, and overturned 4 textbook assumptions shared by 9 models.
Why unconstrained 2D diffusion fails in clinical prosthodontics and how a deterministic parametric-form STL/PLY 3D library, 476/477 rendering pipeline, 478-landmark cephalometric QLoRA (90.6% face shape, 100% target line on n=32), and 4,686-pixel lip lock bridge AI preview to exocad.
How an append-only JSONL event log (task_bus.jsonl) and OS kernel WatchPaths replace polling loops in multi-agent orchestration—verified across 2,344 live events, 262 tools, 3,313 automated self-tests, and 29 directives with 0 ungated passes.
Why LLM self-reflection fails to stop hallucinations and how an isolated auditor (Reaper) verified 360 of 394 citations (91.4%), caught 25 fabricated references (6.3%) across 10 models, and enforces Popperian falsifiability across 58 system methods.
Technology imposes need: where there is a device in a clinic, there is measurement; where there is no device, there is estimation. From diagnostic triage and treatment mix to the 5–8 minute raw scan reality of the in-house lab, lithium disilicate vs zirconia biomechanics, constrained-canvas AI preview, SKS bureaucracy, and the 2026–2030 dead-end verdict.
Reader-friendly edition presenting Yusuf Aziz's 24 field verdicts, 10 concise sections, counter-arguments, and 38 endnotes verified by an automated citation auditor at one-third the length.
Why does an untrained vision-language model read a wide, square-jawed (brachyfacial) face as “oval” and pick a rounded feminine tooth line? 478-landmark 3D cephalometry, SAM 2B incisor ratio and a 1,000-iteration MLX QLoRA: face-shape accuracy 0% → 90.6%, target line 0% → 100% (n=32 locked test); live 3D module and slider on a 100% synthetic face.
The Mirror Twin (SHA-256), the 98.6% Bureaucracy Trap, and a 76-Second Synaptic Harvest on M3 Ultra: 7 blind AI voices and 96 live weight-training trials — how a 3B model trained with Atlas's laws (93.8%) beat an unruled 31B model (75.0%).
Turkey's whisky reached shelves in 1963 and vanished around 2009; today the barriers are a legal asymmetry, the tax base and years without revenue, and the likeliest path is a licensed producer opening a whisky line.
Five AI voices and a red team tested a single sentence: why the rhetoric–action gap is sold and why it is bought, which part stood and which fell.
Statistics don't lie, but they give no calendar. So why do people and governments seem not to hear? Most at the table are homeowners; whoever set it up draws from the house.
Dragon's Dogma 2: Dark Arisen doesn't bring back the old; it gathers the old's best parts — and leaves the decision to you
In Türkiye's new regulation, the part that will really protect children is not the under-15 ban but the design rules written for ages 15–18. What three approaches worldwide produced, where the regulation can break, and five recommendations with first-year measures.
1. Smile Design Track (Cephalometric → STL/PLY → Mockup → Evidence): Computer-vision smile design showcase → 478-landmark cephalometric alignment and LoRA experiment (TR benchmark) → STL/PLY 3D tooth library and photorealistic inpainting → digital dentistry field verdicts (reader edition).
2. Agent Architecture Track (Spark → Task Bus → Reaper → QAJAN): Spark recursive loop and multi-voice LLM red teaming → event-driven autonomy, task_bus.jsonl architecture, and identity freeze protocol → Reaper auditor and context-blind falsifiability gate → QAJAN 6 mechanical hypothesis gates (EN Hub · TR · AR).
Posts are for general information and are not legal advice.