SYSTEM GUIDE / v0.4

Build models that learn
with control.

A practical guide to Vqela's six-system cognitive host: signal timing, durable memory, vessel isolation, and measurable action.

What Vqela is

Vqela coordinates six cooperating systems around a resident population of active units. Instead of allocating one dense object for every possible token, it keeps a sparse working set and steps only the units that are relevant to the current signal.

Design principle

Use timing and salience to decide what matters. Keep memory durable, actions measurable, and the runtime isolated.

The six systems

1. Attention

Competes for salience and gates focus.

2. Episodic systems

Tracks episodes, spatial context, and transfer.

3. Threat

Checks fast lexicon and anomaly signals.

4. Action and habit

Selects policies and applies reward updates.

5. Recall and affect

Reinstates memories and supplies valence.

6. Vessel

Provides the isolated user-space and I/O.

Quickstart

Start an interactive window or run one complete cycle from the repository root.

Shell

PYTHONPATH=. python3 -m system16 window

PYTHONPATH=. python3 -m system16 step \
  "there is a fire in the north room"

How a signal moves

Drive combines inputs, excitability, and ion-channel state. EPSP traces decay faster; IPSP traces decay more slowly. Higher drive shortens response latency within configured bounds.

Signal model

drive = sum(inputs) + excitability + ion_state
latency = base_latency / (1 + k * relu(drive))

Memory commands

Inspect the current state

Memory updates after each line unless you supply an explicit reward.

Window

S16> there is a fire in the north room
S16> :inspect
S16> :reward 0.4
S16> :memory
S16> :forget

Run inside a vessel

The vessel is not Docker. It is a tarball image plus a runtime that packs the package and manifest, extracts an isolated workdir, strips the environment, applies resource limits, and attaches a framed text window over stdin/stdout.

Pack and launch

PYTHONPATH=. python3 -m system16 pack
PYTHONPATH=. python3 -m system16 run

Optional distillation

Dialogue JSONL can teach a small character language model. Without a model key, the PhraseBook fallback still lets you smoke-test the flow.

Smoke tests

PYTHONPATH=. python tools/teach_llm.py --dry-run --pairs 8
PYTHONPATH=. python tools/train_small_lm.py --dry-run

Evaluation and promotion

Keep learning, evaluation, and deployment separate. Promote a candidate only after checking memory behavior, threat handling, resource limits, and rollback readiness.

Security and privacy

Run the vessel with a stripped environment and explicit resource limits. Keep durable memory in the vessel workspace and treat user-provided text as untrusted input.

Build models that learnwith control.