Trioron — an epigenetic-inspired self-expanding architecture

A continual-learning architecture built around the trioron: a node with three coupled state variables (weight, plasticity coefficient, utility) that grows, prunes, and consolidates under a per-curriculum byte budget. Designed for device-conscious deployment on agentic-AI / IoT / embedded hardware.

The full design is in trioron_blueprint.md. The paper draft is in paper/.

• Want a 5-min reproduction of the paper headline? → QUICKSTART.md
• Want to build your own trioron network and deploy it as an agent? → MANUAL.md
• Just the cross-modal bridge / encoders? → BRIDGE.md
• Want a visual walk-through of the mechanisms? → tour/ — a static Canvas petri-dish demo, thirteen chapters with one knob each. Live at https://marcrockhat.github.io/trioron-project/tour/.
• Want an example of what it can do? → Check out https://huggingface.co/spaces/Marcrockhat/trioron-demo.

Headline

On a 30-class class-incremental curriculum (chained-15: MNIST → Fashion-MNIST → EMNIST-letters), with growth + dreaming + manifold replay enabled:

• 0.601 ± 0.008 full-softmax / 0.677 ± 0.007 domain-aware / 0.961 ± 0.001 task-aware at 30 KB of replay storage (n = 10 seeds, paired).
• σ-confident wins over PackNet, HAT, Online EWC, and LwF + EWC on full-softmax and domain-aware (+10σ to +28σ paired). Matches a K = 50 hippocampal exemplar buffer within 0.04 absolute full-softmax at 1/25th the storage.
• BF16 + int8 dream-archive: 157 KB total deployment (network + manifold buffer), Δ ≤ 0.0008 lossless.
• Ship-wake-extend loop validated end-to-end at 23 tasks / 46 classes, 168 KB total, original tasks preserved at task-aware ≥ 0.93.
• Multi-branch absorption: zero-shot composition of independently-trained donors via a 4-byte L0 handshake (R · S factorization); SOFT routing tracks the per-donor upper bound (Δ ≤ 0.0002 task-aware) out to N = 5 donors.

Method and result details: paper/paper.pdf (built from paper/paper.tex).

Install

pip install trioron            # 0.3.2+ — 0.3.1 build_donor crashes; earlier wheels lack trioron.pcll/api

Or straight from GitHub for the latest unreleased changes:

pip install git+https://github.com/marcrockhat/trioron-project.git

Everything a user should import lives in trioron.api. There are three ways in, and — the thing that trips people up — each is fed differently. Trioron is not a fit(X, y) library; only the first path takes a dataset.

1. Continual classification / donors (dataset in, donor out)

The paper's flow. You bring (X, y) per task; the network grows, locks, dreams and rehearses on its own under a byte budget.

from trioron.api import TaskData, TrioronConfig, build_donor

tasks = [
    TaskData(name="cats_vs_dogs",
             X_train=Xtr, y_train=ytr,   # (N, 784) float32, (N,) int64
             X_test=Xte,  y_test=yte,
             classes=[0, 1]),
    TaskData(name="birds_vs_fish",
             X_train=..., y_train=..., X_test=..., y_test=...,
             classes=[2, 3]),
]
donor = build_donor(label="my_donor", tasks=tasks, seed=42,
                    config=TrioronConfig(cap_bytes=32_000),
                    out_path="my_donor.pt")

Compose donors with absorb, keep teaching with extend, deploy with deploy_agent — all from trioron.api; see MANUAL.md. (absorb(rec, donor, ...) on 2.0 Substrate objects is the head-merged graft — one substrate whose forward is the exact sum of its siblings'; MANUAL §13.7.)

2. The substrate itself (a growing net you train like any torch module)

The 2.0 core: cells with a triparametric node (weight, epigenetic lock λ, axonal gain), a hard parameter envelope, growth/pruning/locking as lifecycle events. Trained by consequence — a loss you choose, TD, anything that produces a gradient.

import torch
from trioron.api import construct, seeded, Envelope, default_dispatch_table

sub = construct(base=seeded(784, 10, interior_cells=32, nonlinear=True),
                envelope=Envelope(max_parameter_bytes=200_000),
                dispatch_table=default_dispatch_table(),
                capacity=1024, sparsity_k=0)
sub.compile(); sub.prepare_training()          # prepare_training() is required
opt = torch.optim.Adam(sub.trainable_tensors(), lr=3e-3)
loss = torch.nn.functional.cross_entropy(sub(x), y)
opt.zero_grad(); loss.backward(); sub.zero_dormant_grads(); opt.step()

Spec: paper/v3/spec.md (§2–§6); canonical short reference: docs/TRIORON_MANUAL.md.

Deploying it — the substrate is a training-time structure, not an inference-time cost. The live forward walks the arena (that is what lets it grow); for serving, fold it to a fixed module:

from trioron.api import export_dense
module = export_dense(sub)                                  # exact, buffers-only
fast = torch.jit.freeze(torch.jit.trace(module, x[:1]))

Measured on the world organism (1 CPU thread, batch 1): arena forward ~485 µs → exported ~50 µs — the same per-call latency as a 27 K-param DQN MLP, at 1/5 the parameters. The export does not learn; keep the arena checkpoint for learning and re-export after each wake/extend/dream cycle.

3. Phasecyte — the gradient-free learner (a stream in, no labels required)

The second learner on the same body: single-pass, phase-coherent lock-in over a stream, sufficient statistics only (no stored data). Leaves are enrolled as domains appear; a manifold router arbitrates; a gradient substrate can then be dreamed from the leaves' own sketches with no wake gradients (chained-15: dreamed 0.540 vs phasecyte-nest 0.474 vs monolith 0.403, n = 3).

from trioron.api import PhasecyteNest, dream_distill, dreamed_predict

nest = PhasecyteNest(sense)                # sense: X -> descriptor tensor
nest.enroll(group=0, genesis_pool=X0)      # when domain 0 first appears
nest.observe(0, X_batch, labels)           # single pass, label-free tolerant
router = nest.fit_router()
pred_group = nest.route(X_query)

Spec §10; docs/design/pcll_substrate_integration.md.

What is not in the package yet: the embodied organism

The survival showcase (https://marcrockhat.github.io/trioron-project/tour/phasecyte.html) — drives → primitive leaves → consequence-taught router → structural dreaming from its own cause-of-death table — still lives in archive/experiments/world/ and requires hand-written skill masters. It is being reduced to a "declare your drives" contract (drive-only vocabulary reaches 112 ± 23 survival vs 148 ± 13 master-built, n = 3, zero policy code; see docs/handoff/HANDOFF.md). Until that ships, use the recipe in docs/learning_methods.md and the scripts under archive/.

Setup (WSL2)

The section below is for reproducing the paper, not library use. If you only need the API, the quick install above is enough.

# Move into WSL filesystem (NOT /mnt/c — that's slow)
cd ~/trioron-project

# Use a venv
python3 -m venv .venv
source .venv/bin/activate

# Install
pip install -r requirements.txt

Torch CPU wheel is ~750 MB. First install is the slow part.

Run the unit tests

python3 -m pytest tests -q          # v2 substrate + phenotype + Phasecyte tests

Four known pre-existing failures (test_learning TestCredit ×2, test_lifecycle ×2) are tracked in docs/handoff/HANDOFF.md; everything else passes.

Reproduce the headline results

The n = 10 panels driving the paper's headline table run unattended via:

bash experiments/run_n10_paper.sh

This sequentially runs the manifold-replay panel, the five-family competitor sweep (PackNet / HAT / Online EWC / LwF + EWC / hippocampal K = 50), and the dream-archive panel. Individual panels can also be launched directly:

# chained-15 manifold-grown panel, n = 10
python3 experiments/bench_manifold_replay_n10.py

# Competitor sweep on chained-15 (n = 10)
python3 experiments/bench_packnet_chained_15_n10.py
python3 experiments/bench_hat_chained_15_n10.py
python3 experiments/bench_online_ewc_chained_15_n10.py
python3 experiments/bench_lwf_chained_15_n10.py

# Dream-archive Phase 1 + Phase 2 (storage win, n = 3 pending rerun)
python3 experiments/bench_archive_n3.py

# Ship-wake-extend loop (chained-15 → +8 EMNIST K..Z)
python3 experiments/bench_chained_extend.py

CSVs and *_run*.log files land in outputs/. Run logs from every reported panel are committed; CSVs are gitignored.

Layout

trioron-project/
├── README.md                    # this file
├── MANUAL.md · QUICKSTART.md    # donor API manual; 5-min reproduction
├── docs/TRIORON_MANUAL.md       # canonical short reference (subordinate to the spec)
├── docs/handoff/HANDOFF.md      # cross-session state of record (rewritten every session)
├── paper/v3/spec.md             # Trioron 2.0 architecture spec — source of truth
├── trioron/                     # the package (pip install trioron)
│   ├── api.py                   # PUBLIC SURFACE — import from here
│   ├── core/                    # cell, epigenome, graph, envelope, arena, construct
│   ├── phenotype/               # how genes express into ops (linear, dendrite, …)
│   ├── bases/                   # construction recipes (seeded, minimal, developmental, …)
│   ├── learning/                # credit, frustration, dream, manifold, router
│   ├── lifecycle/               # growth, evolution, ship, graft, compact
│   ├── pcll/                    # Phasecyte (phase-coherent lock-in) + nest + wake/dream
│   ├── evolution/ · viz/        # multi-substrate controller; recorder / viewer
│   ├── compat/                  # v1 ↔ v2 bridge
│   └── legacy/                  # v1 modules (donor API implementation, benches, competitors)
├── archive/experiments/         # research drivers (world/, progenitor/, …) — not packaged
├── experiments/                 # paper bench scripts (CSV + log outputs)
├── outputs/                     # bench CSVs (gitignored) + run logs (committed)
├── paper/                       # paper.tex / paper.pdf / refs.bib
├── tour/                        # static Canvas demo + phasecyte showcase; GitHub Pages
├── hf_space_build/              # Hugging Face Space deployment build
└── tests/                       # unit tests

Status

• Step 1: TrioronLayer + tests
• Step 2: TrioronNetwork + 2-task continual-learning verification
• Step 3: Scripted incubation environment
• Step 4: Three-condition growth trigger (plateau / rank / grad-stability)
• Step 5: Cellular division routine
• Step 6: Pruning loop
• Step 7: Hard ceilings (cap_bytes pre-flight)
• Step 8: Benchmark vs same-param fixed MLP (falsification gate cleared)
• Phase 4.5: Dreaming phase (replay / compress / purge / archive)
• Manifold replay (storage-free pseudo-rehearsal)
• Dream archive (Phase 1 row-lock + Phase 2 int8 quant)
• BF16 mixed-precision deployment substrate
• Ship-wake-extend loop (chained-15 → chained-23)
• Five-family competitor sweep (PackNet / HAT / Online EWC / LwF / hippo) at n = 10
• Multi-branch absorption + L0 handshake translator (R · S factorization)
• Tour: 13-scene Canvas demo at https://marcrockhat.github.io/trioron-project/tour/
• Full integrated paper draft (paper/paper.tex, 29 pages)
• ArXiv submission (pending endorsement)
• PyPI release (pip install trioron); 0.3.x adds Phasecyte nest + wake/dream (trioron.pcll)
• Embodied organism as a package API ("declare your drives", no hand-written masters)
• Deployment script + ready-to-use checkpoint for Orange Pi 5B

Disclosure

This work was carried out in collaboration with two personified AI assistants in defined supporting roles: Gemma (Gemini Pro 3.1, academic-advisory) and Chloe (Claude Opus 4.7 1M-context, engineering). Human-led problem framing and final decision-making; AI-supported implementation, analysis, and writing. The human author holds sole responsibility for all claims, methodological choices, and interpretations. Per recent editorial guidance (Nature 2023, WAME 2023), AI systems are not listed as authors of record.

License

MIT. Copyright © 2026 Marcelinus R Hatorangan.