OPEN SOURCE · C++17 · ZERO DEPENDENCIES

Decisions that
collapse out of physics

SyFox is an open System One decision engine. State in. Typed decisions out. No text generation. Ever. Built on Synthetic-Intelligence substrate physics — not transformers, not neural nets, not pattern matching.

Bit-deterministic Honest silence 100+ languages CPU microseconds
syfox decide
$ ./build/syfox decide --model model-tickets \
  --state "You charged me twice, please refund" \
  --questions '{"department":{"type":"choice",...}}'

{
  "department": {
    "choice": "billing",
    "confidence": 0.204,
    "deferred": false
  },
  "settled_energy": 0.83,
  "lanes": 5786
}
µs
~40 µs
p50 latency
✓
1.00
OOD defer rate
0
Transformers
Neural nets
~40µs
p50 decision
latency
1.00
OOD defer
rate
100+
Languages
supported
0
Dependencies
Runtime cost
C++17
Pure physics
substrate
How it works

One settle pass.
Decisions emerge.

State tokens inject energy into a ConceptField. The field settles through dissipative diffusion along Hebbian lanes. Typed answers are read out from the same settled field. Unknown vocabulary produces honest silence.

01 — INJECT

Energy in

State tokens deposit energy onto concept nodes. Acoustic mass and source-cap keep the working set bounded.

source_cap=241/√mass
02 — SETTLE

Diffuse & damp

Energy propagates along Hebbian lanes. Dissipation + cavity salience sparsify the field. Optional Miller window.

K_settletanh salience
03 — READOUT

Typed answers

Choice, score, and noul questions read resonance from the identical settled field. Flat marginal latency.

specificsupport
04 — SILENCE

Honest defer

Nothing settled above floor? Defer. Never invent an answer. Unknown words stay dark.

energy floorno guess

Core = SI physics only. No transformer, no neural network, no pattern-matching classifier. Tools around the core (JSON, CLI, HTTP) are ordinary engineering. The substrate itself is pure energy dynamics.

Why it matters

Not another model.
A different substrate.

SyFox will never generalize like a foundation model. It sees only what you taught it — inspectable lane by lane. The trade is deliberate: narrow, fully ownable, auditable, offline, and free.

LLM (System 2) Jev SyFox
Output prose to parse typed values typed values
Compute token by token one parallel pass one settle pass
Can hallucinate text yes no (schema-bound) no + honest silence
Core transformer trained foundation model energy-field physics (SI)
Weights billions, opaque opaque your lanes, inspectable
Runs offline maybe no yes, zero dependencies
Cost per decision $$ very low ~0 (CPU µs)
Evidence

Measured on held-out data.
Not resubstitution.

v2.1 introduced a real 70/30 held-out split. Coverage-vs-accuracy is the headline metric. Derivation is gated so it cannot silently regress taught knowledge.

Held-out operating points

Coverage Accuracy
100% 83.3%
≥70% 90.0%
≥50% 100%

Tickets domain. Confident subset is trustworthy by measurement.

What the harness enforces

  • No-regression gate — any taught-row argmax flip or manufactured certainty on ambiguous probes reverts the fabric bit-exactly.
  • Honest silence — OOD / unknown vocabulary defers at rate 1.00. Never invents.
  • Calibration on held-out — multi-class NLL + adoption guard. Tickets ECE 0.665 → 0.012.
  • Bit-deterministic — same state produces the same settled field on every machine.
v3.0

Multilingual + trigrams

UTF-8 script routing across 35 families. Traction-gated character trigram bridges for typo robustness. Active-learning loop.

1.00Bengali held-out routing
v2.1

Held-out methodology

Real 70/30 split, coverage curve, multi-class calibration, Porter + synonyms at the boundary.

0.012Tickets ECE after fit
v0.2

Derivation & gate

Compose + harvest, dreamer, analogs. Transactional no-regression gate with bit-exact snapshot/restore.

PASSGame model ships derived
Boundary

100+ languages.
One physics core.

v3.0 rebuilt the input boundary. Pure-ASCII keeps the original path. Non-ASCII routes by Unicode script family into separate substrates. The SI core stays byte-identical.

Latin বাংলা हिन्दी Русский العربية 中文 日本語 한국어 Ελληνικά עברית ไทย Tiếng Việt + 20 more families
Script families 35
Bengali held-out routing 1.000
OOD defer (all scripts) 1.00
Trigram bridges traction-gated

Legitimate unseen words stay dark. Corrupted forms can recover only when ≥25% of their trigrams already exist in the fabric.

Get started

Build. Teach. Decide.

Zero dependencies. One make. Your data stays yours.

terminal
# clone & build
$ git clone https://github.com/Mr-DS-ML-85/SyFox.git && cd SyFox
$ make                           # builds syfox + tests
$ make test                       # physics + honesty checks
$ make models                     # teach the three seed domains

# decide
$ ./build/syfox decide --model model-tickets \
  --state "charged twice for the same invoice" \
  --questions '{"department":{"type":"choice","criteria":{...}}}'

# optional HTTP (Jev-compatible shape)
$ make serve                       # http://127.0.0.1:8010