Lab note — the studio's internal working record; for the service see the home page.
Spark · producing theses
Spark is not a chat; when the round ends, a scorecard is written.
Spark (Kıvılcım in Turkish) is the seven-step procedure for producing theses and theories: Anu (Claude Fable) records the pre-expectation and writes the anonymous brief, Atlas (Gemini 4 Argon · Vizier) runs the round, the Author (S0) · field verdict · final say layer forms the backbone of field syntheses, and isolated channels where voices hear nothing of each other · objection / counter-thesis only (Atlas Argon own voice, context-free Gemini Web, Claude arms Opus · Sonnet, Astra, Grok, and inner voices Enki · Enlil · Enkidu / local Gemma · Qwen) pass through the Reaper sieve: source verification (Crossref · PubMed · DOI) + rote filter + red team. Testing a ready-made verdict is a separate job; that is called a referee round, and the two are not mixed.
This page has three parts: the seven-step procedure, the scorecard of the rounds so far (67 rounds · 73 log rows), and four real examples of the procedure. The procedure is dated 21 Sep 2026 (19 Oct 2026 update: author voice · Reaper sieve · evidence cards).
67 rounds (73 scorecard rows), 7 mandatory steps, and 4 end-to-end documented synthesis/experiment examples.360/394 citations across 10 external voices were verified (25 fabricated references eliminated; consolidated 386/420).0% → 90.6% on locked n = 32) · Synaptic Harvest (N = 96 experiments, 93.8% vs 75.0%) · Autonomy architecture · 16 machine-readable records · Author: Yusuf Aziz.What it showsA round begins with Anu (Claude Fable) writing the pre-expectation as a single line to the scorecard file (data/kivilcim_karnesi.jsonl); the state is measured, and Anu writes the anonymous brief while Atlas (Gemini 4 Argon) runs the round. Step 04 separates two layers: at the top, Author (S0) · field verdict · final say forms the backbone of the synthesis; below, six dashed channels where voices hear nothing of each other · objection / counter-thesis only (Atlas Argon own voice, context-free Gemini Web, Claude arms Opus·Sonnet, Astra selected·contextual, Grok red team, and Inner voices Enki·Enlil·Enkidu + Local Gemma·Qwen) supply objections and counter-theses. In Step 05, citations fall onto the Reaper sieve: source verification (Crossref · PubMed · DOI) + rote filter + red team; verified sources pass down, while unverified or rote ones stay stamped UNVERIFIED / ROTE. In Step 06, the synthesis is built in the author’s voice, mapping stances into a voice × thesis matrix and Evidence cards (file:line + dilemma line) categorized under “Spark verdict”. In Step 07, the scorecard is written on the same line as the pre-expectation (YES | SHARPENED | NO) with the Reaper stamp.
Run in order. The round starts before the brief is sent and does not end when the answers arrive.
Before the round starts, Anu (Claude Fable) writes the expected answer in a single line and seals it to the scorecard file (data/kivilcim_karnesi.jsonl). No brief is sent until it is written.
The aim is to measure honestly, afterwards, whether the round changed anything.
The STATE section of the brief is built from measured items with their sources. What cannot be measured is written as “UNKNOWN”.
Anu writes the anonymous brief and Atlas (Gemini 4 Argon) runs the round. In production mode no ready-made verdict is given; every voice is asked for the same form:
No line cap is set; the answer is not reduced to a checklist.
In authored articles and field syntheses, two layers are separated: at the top, Author (S0) · field verdict · final say forms the backbone of the text and is not “one of the voices”. Below, channels where voices hear nothing of each other · objection / counter-thesis only — Atlas (Gemini 4 Argon, own voice and Vizier), context-free Gemini Web, Claude arms (Opus · Sonnet), Astra (selected · contextual), Grok (red team), the inner voices Enki · Enlil · Enkidu, and local models Gemma · Qwen — produce objections and counter-theses only. In Wave 1 none sees another’s answer.
No number is taken from a model whose identity has not been confirmed. Every source passes through the Reaper sieve: (1) source verification (Crossref · PubMed · DOI) — unfound citations stay stamped UNVERIFIED, and fabricated citations (including light class) mechanically invalidate the answer; (2) rote filter (EZBER) — unsourced outputs or generic literature summaries that fail to answer the brief’s specific question are filtered out as ROTE / EZBER; (3) red team — the strongest sourced objection against consensus (Grok when quota is open, alongside Astra).
“The brightest reference is the one most often made up.”
· ✓ Verified by Reaper · 360/394 · 19 Oct 2026
The synthesis is built in the author’s voice (Atlas runs it; the verdict rests with the author + Anu). Every voice’s stance on each thesis is mapped into a voice × thesis matrix and Evidence cards (file:line + dilemma line); external and inner voices appear only in counter-thesis boxes. Syntheses carrying a verdict are categorized with the “Spark verdict” (Kıvılcım hükmü) label. The raw answers are kept.
On a deep topic a second round opens: the measurement in between is run, then a second brief goes out.
When the round ends, “did the decision change” is written on the same line as the pre-expectation: YES | SHARPENED | NO plus a one-line note; published syntheses carry the Reaper stamp (verified/total · date).
A real count (data/kivilcim_karnesi.jsonl: 73 log rows, 67 deduplicated rounds / 66 unique topics). The question is the same in every round: did the decision change.
What it showsdata/kivilcim_karnesi.jsonl contains 73 rows; because 6 rounds recorded pre-expectation and synthesis on two consecutive lines, the deduplicated count is 67 rounds (66 unique topics since one topic’s first attempt was KOSMADI and second KESKINLESTI; 64 closed rounds / 65 non-null rows: 44 Sharpened, 13 Yes [14 in raw rows], 3 No, 2 Worthless/withdrawn, 1 Did not run, 1 Unknown; plus 3 open/blank rounds [8 null rows in raw file including 5 paired pre-expectation lines]). Blank, worthless and unknown rows are counted too; they are not hidden.
Gate: once a month three things are counted: the decision-changed rate, the gap between pre-expectation and synthesis, and the rate of unverifiable sources. If the rate cannot be told apart from chance, the round is an expensive chat.
Real rounds.
Built all together in one night by three agent families and Yusuf: rings P0–P6 and G1–G5, the 4-model handover reading benchmark, and local watchdog gates went live under mechanical self-tests. The full narrative essay is on yusufaziz.com.tr, while the technical Evidence Note with all architectural tables is published here.
Read the essay (yusufaziz.com.tr) → · Technical Evidence Note →
Four voices each wrote, separately, the second evolution of a family of three agents. The decision was made where they split.
Seven blind voices collided; in Round 2 a 75.9-second MLX QLoRA weight-training experiment ran on M3 Ultra: a 3B model trained with Atlas's laws (93.8%) beat an unruled 31B model (75.0%).
Against an untrained model that reads a square jaw as “oval”: 478-landmark 3D cephalometry + 1,000-iteration QLoRA (n = 32 locked test, raw log published). A human eye caught the confounder; a 4-arm ablation tested it.
Read the experiment, the live 3D module and the ablation table →
Why generators cannot audit their own reasoning in-context, how Reaper caught 25 fabricated references across 10 models, and how 67 Spark rounds enforce Popperian falsifiability.
Read the proof note → · Event-Driven Task Bus → · Applied Spark (on yusufaziz.com.tr →) · Identity Freeze & Three-Family → · QAJAN 6-Gate →
The author’s (S0) 24 field verdicts, isolated AI voices providing objections/counter-theses, voice×thesis matrix with 12 evidence cards, and 90+ verified DOI/PubMed sources on the Türkiye and global equation. ✓ Verified by Reaper · 360/394 · 19 Oct 2026.
The EVOLUTION-2 Spark · 7 Oct 2026.
The brief went out in production mode, anonymous. The question: what should the second evolution of the family of three agents (Enki · Enlil · Enkidu) be? The state was written as measured:
Four voices wrote: Gemini (no context, two seats), Gemma (local) and Grok Bot. They did not hear each other.
The pre-expectation was written before the round: no new agent; a proposal to attach “friction” to Enlil will come and will be rejected; the blind mutation engine will be Enkidu.
The voices split in four places; the decision was there. The hardest counter-thesis was the manual alternative: if the owner looked at the list of decaying rules for 15 minutes once a week, the same result might come out.
Split · who produces the mutation
“100 local variations a day”. The zero-cost assumption collapsed the same day, under GPU starvation.
“One a day, ≤3 candidates + placebo 4:1”.
Placebo was chosen.
Split · scale
Three steps in 30 days. That was the 30-day plan of three voices.
The arithmetic: at a 76% base rate, seeing a 10-point difference with 80% power takes ≈240 episodes per arm. Three steps do not fit into 30 days.
30 days = one step. The plan of the three voices was optimistic.
Split · the manual alternative
The automatic loop: mutation → selection → consolidation.
A look by hand: if the owner looked at the list of decaying rules for 15 minutes once a week, the same result might come out.
Accepted as the kill test: two weeks in parallel; if they are equal, the automation stops.
Split · contrast embedding analysis
“Contrast embedding analysis”.
An experiment the same day: the embedding model saw contrary verdicts as 0.90 similar.
Removed.
All four of the four voices agreed on these (4/4):
Source verification: two academic citations were verified on the web; one had been given with the wrong year (2024 instead of 2026). One citation could not be verified and was not used as support.
Two measurements kill the thesis. The first is the manual alternative: it runs in parallel for two weeks, and if the two are equal the automation stops. The second is the main measure: if the number of closed loops does not rise in four weeks, there is no evolution.
The synthesis turned into an order: EVOLUTION-2 Step 1, in shadow.
It was built the same day: 7 tools, 153 tests. The model that built it left 9 objections; the objections went on record. One of them: “the backward arm, at a 24-hour horizon, takes ~218 days to reach 240 episodes → a structural zero in this arm”.
SHARPENED. The expectation held; the new parts that sharpened it: the placebo, the 240-episode arithmetic and the manual alternative as kill test.
Primary technical literature and open-source specifications supporting the architectural and experimental measurements on this page:
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).