SyFox Giant — a broad System-One decision engine (SI substrate, no NN)

One giant SyFox fabric family covering 16 decision domains (customer support, banking intents, spam, emotions, moderation, news topics, product sentiment, entity typing, boolean comprehension, game commands, guardrails) plus per-script substrates for Bengali / Devanagari / Cyrillic with automatic script routing. No transformer, no neural network, no pattern matching: every decision comes from a C++ SI substrate — energy injection → dissipative settle → resonance readout → honest silence. The lanes ARE the knowledge; every one is inspectable in model/*/substrate.bin.

Built from 76,142 real/synthetic-verified training rows across 16 domains, trained with the three failure-mode fixes learned from the model-xl RCA: empty per-row instructions, globally-unique opaque anchors (278), and round-robin domain interleaving. Full measured evaluation: docs/EVALUATION.md.

Headline measured numbers (hidden splits, scored once)

domain	class acc	router	end-to-end
tickets_en	0.904	0.831	0.761
sms spam	0.752	0.233	0.222
enron email spam	0.752	0.394	0.323
hate/offensive	0.653	0.381	0.277
ag_news topics	0.604	0.320	0.206
amazon polarity	0.571	0.339	0.189
dbpedia entities	0.475	0.403	0.204
boolq (boolean)	0.531	0.097	0.059
bank77 (77-way)	0.023	0.931	0.022
tickets_bn (bengali substrate)	0.943	1.000	0.943
tickets_hi (devanagari substrate)	0.920	1.000	0.920
tickets_ru (cyrillic substrate)	0.959	1.000	0.959

Honest notes: the single 16-way router is the end-to-end bottleneck for same-script domains — blind routing alone measures 0.400 on hidden (23,528 rows) and the table's router column is the informed condition (router + the true domain's class question asked together); naming the domain in the API gives the class numbers. bank77's 77-way readout collapses in a shared fabric — see v3.2: the dedicated bank77 fabric below for the architectural fix (0.023 → 0.081 → 0.158); boolq stays near majority — boolean passage comprehension is a measured limit of bag-of-words SI physics. Latency ≈ 6-11 ms/row on the 43k-node latin fabric.

v3.2 — the semantic layer, retrieval by default, the two-stage router

The engine in this repo is SyFox v3.2.0: the substrate now carries a second, deterministic field — a 64-dim semantic vector per concept (signed character-trigram hashing + two fabric-grounding passes over the Hebbian lanes; no ML, no fit) — from which resonance edges leak an energy-conserving share of a source's energy to semantically similar neighbours during settle. Context-sensitive lanes learn required/forbidden context words from the lessons that laid them and carry less when the decision's own tokens do not match. Retrieval is default-on: a memories.jsonl in the model dir primes every decide with the outcomes of the most resonating lived experiences (Hopfield-style settled-field cosine), disclosed in the response as usage.retrieval. Pre-v3.2 fabrics replay bit-for-bit; --no-semantics / --no-retrieval / --no-hierarchy (and the equal options keys on HTTP) are the kill switches.

Two new model dirs ship here:

• model/b77-sem — the DEDICATED banking77 fabric: 9,000 opaque-anchor intent lessons + 9,000 category lessons, 1,648 nodes / 217,729 lanes, 7,728 resonance edges, 141,920 context-signed lanes, 256 retrieval memories, calibrated. Hidden (3,080 rows, scored once, energy-norm): 0.158 top-1 vs 0.081 (v3.1.0 dedicated+opaque) and 0.023 (shared giant). Ablation: --no-semantics 0.039 — the semantic field IS the jump; the hierarchy gate measured-negative on cal and ships floor 1.0 = off.
• model/router16 — the two-stage physics router: 4,742 nodes trained ONLY on router questions (16 domain anchors x 60 rows). Cal 16-way routing 0.444 (chance 0.0625), ECE 0.380 → 0.086. Route with decide --router model/router16 (CLI) or the two-call pattern below — bank77 queries route to model/b77-sem, everything else to giant-latin.

Zero-shot re-run under v3.2 (docs/ZEROSHOT.md Part 4, datasets/zeroshot_v32_results.json): email spam on the tickets-only fabric 6/6 plain and 6/6 semantic rebuild; sms spam 5/6 (the one error at conf 0.103, disclosed); snake 1/8 both; tic-tac-toe picks cell 1 at conf 0 — multi-constraint composition fails exactly as in Part 1 and the semantic field does not fake it.

Files

engine/      C++ SI substrate core v3.2 (core/*.hpp, src/*.cpp) — builds with make
server/      HTTP bridge (v3 API, CLI-equal flags incl. the v3.2 knobs)
inference.py thin python wrapper (auto script routing)
model/       giant-{latin,bengali,devanagari,cyrillic}/substrate.bin
             b77-sem/      dedicated 77-way banking fabric (v3.2 semantic layer)
             router16/     two-stage physics router (router.json: anchors+models)
datasets/    eval hidden splits, probes, schemas, MANIFEST (SHA-256),
             router16_router.json, zeroshot_v32_results.json, DATASETS.md
docs/        EVALUATION.md, ARCHITECTURE.md (§16 = v3.2), PROBING.md, ZEROSHOT.md
examples/    JSON payloads + curl + python usage
Dockerfile   HuggingFace Spaces (builds engine, serves on 7860)

Usage (python)

from inference import SyFox
m = SyFox()                       # loads model/giant-<script> per detected script
r = m.decide("I was charged twice on my last invoice, please refund.")
print(r["answers"]["domain"])     # router: choice + probabilities + confidence

# ask a named domain's question directly (skips the router bottleneck):
r = m.decide("I was charged twice on my last invoice, please refund.",
             domain="bank77")
print(r["answers"]["intent"])     # 77-way choice with probabilities

Usage (HTTP, what the Space runs)

curl -X POST http://localhost:7860/v1/systemone -H 'Content-Type: application/json' -d '{
  "state": "Congratulations! You won a £1000 prize. Call now to claim!",
  "questions": {"domain": {"type": "choice", "instructions": "",
    "criteria": {"rt00": "support billing sales technical", "rt05": "text messages spam",
                  "...": "see datasets/schemas.json"}}},
  "options": {"energy_norm": true, "defer_margin": 0.05}
}'

The server also exposes /v1/health, /v1/models, /v1/deferrals (active-learning mirror) and per-call options energy_norm, salience_gating, miller_window, ngrams, defer_margin, evidence, and the v3.2 knobs semantics, retrieval, retrieval_topk, retrieval_dose, hierarchy.

Docker (HuggingFace Spaces)

The included Dockerfile builds the engine from source and serves the v3 API on port 7860. Create a Docker Space, push this repo, done. CLI-equal flags are passed in the CMD (--lang auto --energy-norm).

Training recipe (reproduce or extend)

See docs/EVALUATION.md §Reproduce. Corpus + tooling live in the main SyFox repo (tools/hf_fetch.py, tools/giant_prepare.py, tools/giant_bench.py, tools/router_sweep.py). Constraint discipline: hidden splits are firewall-enforced (never taught, never calibrated on); all model/wording selection happened on cal; the core physics was never modified.

Limitations

• Single-router e2e is bottlenecked by 16-way choice in one fabric (measured; wording sweep shows it is the readout mechanism, not the wording); the v3.2 dedicated router16 fabric raises blind 16-way routing to 0.444 on its cal carve and routes bank77 to its dedicated fabric.
• 77-way banking intents: dedicated fabric v3.2 measures 0.158 hidden (semantic field ablation 0.039 — the field is the mechanism); higher ladders need arbitration or per-category substrates (hierarchy gating itself measured-negative on cal, shipped off).
• Boolean passage comprehension is near-majority — the substrate reads lexical fields, not logic.
• Zero-shot composition (constraints, topology) is NOT solved by scale or by the semantic field — see docs/ZEROSHOT.md Parts 1 and 4 for the measured two-dimension separation.

License & credit

MIT © 2026 Mr-DS-ML-85. Core mechanics ported from the author's Synthetic-Intelligence research substrate. Dataset licenses: see datasets/DATASETS.md.