Building ENIGMA: a deterministic agent simulation, solo, via free-tier LLM access

In March 2026 I started building a deterministic causal simulation engine — agents that act on their own incomplete and potentially erroneous models of the world, not on objective truth. The principle is WORLD TRUTH ≠ NPC MEMORY ≠ NPC BELIEF, enforced at runtime.

Today, 7.5 months later: v0.5.4.4.5, 729 commits, 18 enforced architectural laws (§ENIGMA-001..018), 12-layer canonical construction order, AGPL-3.0, Zenodo DOI.

This is not a game. A scientific software artifact in active development.

This is not a tutorial. Architectural documentation for researchers and engineers working on agent systems, epistemic logic, deterministic simulation, or LLM-assisted engineering.

This is not a product. Solo-maintained research project, no support SLAs, no roadmap commitments to external stakeholders.

I want to share the engineering methodology, because it's the most interesting part and probably the most reproducible.

I don't use agent swarms. I don't use API access. I use free-tier web access to multiple models in parallel, three browser windows, and a human-as-router architecture:

The models argue with each other through me. I synthesize. The final decision goes to GLM as an architect with a strict format contract (original Russian names kept):

БЫЛО: <one line of code to be replaced> СТАЛО: <new code> IPT: <invariant checked> КРАСНЫЕ ИНВАРИАНТЫ: N red → M red Enter fullscreen mode Exit fullscreen mode (Translation: БЫЛО = before, СТАЛО = after, IPT = invariant probe test, КРАСНЫЕ ИНВАРИАНТЫ = red invariants count)

This is faster than any agent swarm I've tried. Context window optimization is done by me — I extract relevant fragments from 200K+ lines of code, the models never see the full codebase. Latency hides through parallelism (while model A is thinking, I formulate task for model B, review result from model C).

I've observed that newer model versions tend to "drift". GLM 5.3 vs 5.2 — 5.3 wants to design inside its own environment, ignoring the working mode protocol. Similar patterns in Claude and GPT between versions.

The architecture is model-agnostic. If GLM disappears tomorrow, I switch to another model and continue — because my asset is the discipline, not the model.

Each layer must use the causal work of the layer below, never imitate it.

0. Authoring → Compiler (LLM translates, Python validates) 1. Identity / History (past exists as data, not text) 2.

Body → Need (states have consequences, not decoration) 3. Desire → Object → Activity (NPC wants something, chooses activity) 4. Decision → Action → Outcome (CAN ≠ WANT ≠ INTEND ≠ ATTEMPT ≠ SUCCESS) 5.

Perception → Experience → Memory (NPC knows only what it could perceive) 6. Salience → Consolidation (memory becomes selective) 7. Relationships as History (relationship = consequence of history) 8.

Belief → Expectation → Decision (NPC acts on its picture of the world) 9. NPC as Interface (player learns complexity through NPC) 10. Self-Interpretation (WHAT HAPPENED ≠ HOW EXPLAINS) 11.

Appraisal / Prediction (surprise as measurable prediction error) 12. Deception (lie ≠ ignorance ≠ error ≠ withholding) Enter fullscreen mode Exit fullscreen mode Anti-Bond rule: each layer must prove causal work no other layer performs. No parallel truth-stores.

No decorative state.

Dialogue ≠ Memory Memory ≠ Causality Belief ≠ Decision Relationship scalar ≠ History Self-interpretation ≠ Motivation LLM ≠ Canonical Truth Enter fullscreen mode Exit fullscreen mode Key architectural principles Epistemic grounding — LLM is never a source of canonical truth; Python = causal authority Deterministic replay — KernelRNG-protocol, no random.* or wall-clock in simulation layer (§15) Causal layering — upper layers must use the causal work of lower layers, never imitate Anti-Bond — no parallel truth-stores, no decorative state Belief non-mutation — beliefs are lenses (weight modifiers), not genes (scalars); personality drift via interpretation is forbidden (§16) Epistemological orthogonality — Reality and Observer Knowledge are orthogonal axes; consumers receive only epistemic representations (§17) What's open right now Active fronts:

The canonical vertical proof is the "Goran? Yes..." memory scenario (track MEM §16.5): an NPC reconstructs its individual history with perspective, temporal structure, surviving and lost details, possible errors, and prediction of future behavior — distinct from WORLD TRUTH, reproducible by the deterministic engine.

Mostly to share the methodology. The "human-as-router with three LLM windows" workflow is something I haven't seen documented, and it's been more effective for me than any agent framework.

If you're doing similar long-context LLM-assisted work, I'd like to hear your patterns. The comments are open.

The project itself is research-grade — AGPL-3.0, so if you serve it over network, share the source.

@software { gritsaenko_2026_enigma, author = {Gritsaenko, M.}, title = {ENIGMA: Executable Epistemic Contracts in Deterministic Agent Simulation}, version = {0.5.4.4.5}, year = {2026}, url = {https://github.com/4Um4/Enigma}, doi = {10.5281/zenodo.22831764}, license = {AGPL-3.0} } Enter fullscreen mode Exit fullscreen mode Happy to answer technical questions about the architecture, the deterministic replay protocol, the KernelRNG, the anti-Bond discipline, or the LLM workflow.