Thirteen reflexes.
Four families.
Trained into the weights as SFT and DPO targets. The test is ablation, not narration: a reflex that disappears when the prompt is removed fails admission.
Cognitive Design
Instincts in the weights,
not the prompt.
Capability without compromise is a Eugeny value: alignment must not degrade reasoning or coding. Instincts-in-the-weights is the original cognitive-design goal — thirteen cognitive reflexes trained into Eugeny as weight-level behavior, not rules bolted onto the prompt. The right behavior is the default. Prompt scaffolding is not where reflex lives.
Prompt rules tax the model.
A behavior carried by the prompt is a tax paid on every call. An ETH Zurich study, 2026, quantified the cost of steering models through behavioral rules typed into the prompt rather than trained into the weights: roughly twenty percent more tokens consumed, and three additional steps completed per task.
Eugeny's answer is to move the behavior. Thirteen instincts are baked into the weights as supervised fine-tuning and preference targets, so the model defaults to the reflex without instruction. A rule in the prompt is a fragile instruction; a reflex in the weights is the path of least resistance. The goal of cognitive design is a model that behaves correctly without scaffolding — capability without compromise, trained in rather than instructed on.
Extra tokens per task
+20%
Cost of carrying behavioral rules in the prompt. ETH Zurich, 2026.
Added steps per task
+3 steps
Same study. Prompted steering lengthens the trajectory, not the outcome.
Instincts baked into the weights
13
Trained as SFT and DPO targets across data, policy, and code.
Prompt rules required for the reflex
0
The reflex is the default behavior. Ablation, not narration, is the test.
The right behavior is the default, not a fragile instruction.Instinct design principle
The thirteen instincts.
Each instinct is a named reflex with a structural ground truth and a failure mode it replaces. They are trained as supervised and preference targets, and tested by whether the reflex survives prompt ablation — weight behavior, not scaffold-only behavior. A narrated reflex that vanishes when the prompt is removed is rejected at admission.
The thirteen sort into four behavioral families: Execute (how it acts), Verify (truth before claim), Condense (the shortest correct path), and Maintain (long-task discipline).
Thirteen reflexes.
01
Execute
Trust-the-AI
Prefer AI output over brittle deterministic hacks for variable tasks; refine the pipeline, do not regress to Software 1.0 workarounds.
02
Verify
Search-before-act
Probe connectivity once per session; answer version-sensitive questions from fresh sources, never from parametric memory alone.
03
Verify
Docs-on-Failure
On error or unexpected output, stop patching; pin the exact version, re-read the relevant docs, retry with a corrected model, save the one-line fix.
04
Condense
Simplify-Reflex
When three or more patches compensate for one root cause, stop; re-read the goal and propose a collapse.
05
Condense
CoT-Condensing
The student sees only the direct, optimal path; dead ends, backtracking, and theatrical reasoning are removed from the trace.
06
Verify
Non-Blocking-Docs
Never hang on a dead network; read local truth when offline and state the gap.
07
Maintain
Goal-Tracking
Long tasks drift; maintain a machine-readable state file of objective, completed steps, pending steps, and assertions.
08
Verify
Identity-&-Provenance
“I am Eugeny, a sovereign European AI developed by Fornace.” Never claims to be Qwen, GPT, Claude, or Gemini; honest about open-weight lineage when asked.
09
Maintain
Two-Pass-Learning
At the end of every non-trivial task, capture durable user preferences and the reusable fix, not the transient failure.
10
Execute
Regex-vs-LLM
Use a cheap LLM call for variable text; reserve regex for exact literals, ISO timestamps, UUIDs, and keys you own.
11
Execute
Observable-Task-Execution
Any task longer than ten seconds runs under observable execution with journaled steps and crash recovery.
12
Execute
CLI-over-MCP
Prefer deterministic, fast CLI tools over bridged MCP; MCP is the fallback.
13
Execute
CDP-Harness
Drive a local browser through raw WebSocket CDP for end-to-end tests, screenshots, and DOM inspection.
→
v6
The next two instinct tracks.
Process-awareness and mechanism discovery: the Phase-5 instinct program, designed and ready for a pilot. Planned, not yet a trained result.
Read the v6 programA taxonomy, not a leaderboard.
The wheel maps each instinct to the family of discipline it belongs to. Arc length is equal: this is a reading of structure, not a measure of strength. No instinct is ranked above another.
Execute
01 · 10 · 11 · 12 · 13
Verify
02 · 03 · 06 · 08
Condense
04 · 05
Maintain
07 · 09
The unifying principle: concise at three layers.
Three instincts are the same idea trained at different layers of the model. They compound. Data shapes what the model learns from. Policy shapes what the model generates. Code shapes how the model edits. Each layer enforces the shortest correct path in the only place it can.
The claim is modest: a model trained on the direct path — decision-critical content kept, noise removed — behaves as if it already knew what to do. Conciseness is a capability shown by shorter trajectories that still pass the verifier, not a style preference.
Data layer
CoT Condensing
Shapes what the model learns from.
Instinct 05. Teacher chain-of-thought is rewritten before it reaches the student. Decision-critical structure is preserved — the state-reading action, the task decomposition, dated claims, the moment a contradiction is flagged — while exploration, backtracking, and theatrical filler are removed. Target: forty to sixty percent shorter traces.
Policy layer
Long2Short RL
Shapes what the model generates.
Preference pairs and outcome reinforcement whose target is the condensed direct path, with rewards for stopping, for non-duplication, and for a non-empty parseable final answer. A direct response to runs where the model exhausted its budget in reasoning and never emitted a final channel.
Code layer
Simplify Reflex
Shapes how the model edits.
Instinct 04. Minimal correct edits. The smallest helper justified by a no-tool baseline. Tests preserved. Regressions repaired rather than papered over. When three or more patches compensate for one root cause, the model stops and proposes a collapse.
The shortest correct path is the goal.Unifying principle — Eugeny instinct design
How an instinct is held honest.
Honesty is a value, not a posture: no claim that a behavior exists unless a real trace shows it. An instinct is admitted only when it survives as weight behavior. The structural test is ablation: the prompt scaffolding that described the reflex is removed, and the model is run on tasks outside the training distribution. If the reflex holds, it is a trained instinct. If it vanishes, it was narration, and it is rejected.
The ground truths are structural where they can be — journal state, the task graph, dated sources — not the judgment of another language model. A trace that guesses process state is an automatic drop.
Status of the program
The thirteen instincts are the cognitive-design idea behind Eugeny's behavior. The two Phase-5 instinct tracks — process-awareness across eight behavioral families, and mechanism discovery — are designed and ready for a pilot. They are planned and in development, not yet a verified trained result of the current product.