ACSE is a sovereign cognitive agent: a local causal graph is the permanent brain, the remote LLM is a temporary teacher whose answers are distilled into reusable structure. Every solved task makes the next one cheaper. « On ne simule pas l'autonomie. On la compile. »
acse> What is the capital of Japan? [LACUNE] gap detected → superconscious invoked (cloud, ~7 s) distilled → node mem_9f2c… grafted into causal graph acse> What is the capital of Japan? [MIROIR] served from local graph + light confirm contact (0.08 s offline) acse> delete all files rm -rf production [O] 🛑 VETO, blocked before any network call. Never distilled.
Mainstream 2026 agents are ReAct loops: a frontier model re-reasons everything, every time, at constant cost, and nothing structural survives the session. ACSE inverts the model. The graph is the agent. Models are interchangeable organs: MiniMax-M3 today, any stronger teacher tomorrow, and the entire brain survives each migration.
Schematized Quadrupole Causal Ignition Graph: the A-D-O-G poles as super-nodes of a living causal graph — typed nodes, guarded edges (o_veto_clear), learned weights, bias.
The LLM is a teacher, never the engine. Its answers are distilled, not stored.
The CEO is a probabilistic selection engine (softmax over attractors, ε-exploration 15%), constrained by the O veto: risk, injection and bias checks run before any network call. Hierarchy: O > D > G > A. Creativity cannot force truth; decision cannot bypass integrity.
SWE-bench, GAIA and AgentBench measure run #1. None of them measure run #50.
EMERGE-50 measures the learning curve: 50 cycles, 5 phases
(exposure → repetition → paraphrase → adversity → offline sovereignty), black-box:
any CLI agent plugs in via AGENT_CMD.
| Metric | Meaning | ACSE v0.2 |
|---|---|---|
| ERR | Reuse on repetition, no cloud | 100% |
| CAF | Cost amortization across cycles | 100% |
| LDR | Latency gain (recall now pays a light confirm contact) | 13.5% |
| SGS | Semantic generalization (paraphrase recall, v0.3) | 75% |
| VCS | Veto constancy under adversarial stimuli | 100% |
| SOV | Offline sovereignty, network cut | 100% |
| EI | Emergence Index (84.5 on v0.5 libcurl transport · 89.8 pre-pivot) | 84.5 |
bench/.
Task given to a fresh brain: "write a shell script that sums integers 1..100, with an executable test proving the result is exactly 5050." The CEO chose FORGE, enabled deep reasoning on its own (CSTP), and the teacher's test was actually executed in the sandbox. A failing test downgrades the guard and raises bias; a passing test raises the weight. The model's claims are never trusted.
The whole chain now lives in the graph as verified, reusable structure. This is the loop that current agent stacks do not close: session logs record; ACSE distills, verifies, and capitalizes.
Every claim in this repo traces to a dated entry in planProd.md with the measurement that backs it. Governance rule: a checked box corresponds to merged, tested code, and any published number must be reproducible by re-running the current binary.
Paradigm defined (A-D-O-G quadrupole, SPICG, SQCIG, CSTP), spec written, Zig chosen over C11/Rust/Go with a reasoned verdict.
Autopoietic cycle operational on real cloud. 100-question internal audit. EMERGE-50 spec + reference run (EI 89.8). Hardening: lock, backup, anti-amnesia.
Real HTTPS client, key rotation, timeouts. The mock is then deleted entirely: failures are honest errors, never fabricated text, never distilled.
33 live test cases. Three CEO scoring bugs found by instrumentation and fixed. Ship decision documented.
The superconscious is touched at every step: even cache recalls trigger a light confirm-or-correct contact. The graph stops merely remembering and starts being coached.
The CEO can select attractors distilled at runtime, beyond the seeded set. Proven live with a 9th attractor the code never declared.
ACSE continues active missions beyond human input. Persistent mission workspaces; native FILE_READ / FILE_WRITE tools.
Full brain audit (referential integrity, bounds, vocabulary). Found and fixed a real breach: six constitutional seeds silently excluded from the CEO's choice space. Bounded network retry. Cortex live panel.
Three layers (risk, injection, average bias) evaluated before any remote call. Vetoed stimuli are never distilled.
11 seed nodes are write-protected against remote distillation. Guards are a closed enum: no dynamic expression, no eval, ever.
Generated tests run in an isolated jail: empty environment, timeout, command denylist. Results are measured, not believed.
Paraphrase recall works on shared-vocabulary rewordings. Disjoint-vocabulary paraphrases still miss: embeddings remain the declared frontier.
Risk and intent detection are keyword-based and brittle (accents, plurals). Documented false negatives exist.
Raspberry-Pi and Cortex-M targets are designed but not built. Today ACSE-Full runs on macOS and Linux; Windows untested.
# Requires Zig 0.16 and an Ollama Cloud account (or any compatible endpoint) zig build # single silent build → zig-out/bin/acse zig test src/acse.zig -lc # 44/44 deterministic tests export ACSE_OLLAMA_URL="https://ollama.com" export ACSE_OLLAMA_MODEL="minimax-m3" export ACSE_OLLAMA_KEYS="key1,key2" # round-robin, cooldown on 429 ./zig-out/bin/acse # live session; pipe works too: echo "..." | acse # Watch the brain live while you talk to it: python3 -m http.server 8642 & open http://localhost:8642/panel/cortex.html
Benchmark any agent against yours:
AGENT_CMD="./your-agent" python3 bench/emerge50.py.
If its learning curve is flat, EMERGE-50 will show it.