Skip to content

Repository files navigation

Conditioned Kernel

Paper DOI Evidence DOI

The model supplies linguistic possibility; the substrate determines what becomes an answer.

Paper: The Black Box Is Smaller Than the Experience (preprint v0.1) — doi:10.5281/zenodo.21683054 · frozen code+evidence snapshot: doi:10.5281/zenodo.21683073 (tag paper-v0.1, commit 1695295).

Local-first experiment harness for substrate-conditioned generation. The language model is treated as a replaceable text-transduction kernel. Effective system behavior is relocated into the substrate that surrounds inference: persistent state, context compilation, validation, repair, acceptance, and rendering.

Built as an edge product, not a desktop prototype that might fit later. Default runtime profile: orin_nano_8gb (Jetson Orin Nano 8 GB class).

Temple of Two research project. Fully local. No cloud dependency. No sensors. No autonomous tools in v0.

Research name Substrate-Conditioned Generation (working synonym)
Project / package Conditioned Kernel (conditioned-kernel / ck)
Edge target Jetson Orin Nano 8 GB · ARM64 · one model · short ctx
Runtime Ollama at localhost:11434
Default profile orin_nano_8gb (configs/edge/)
v0 target Measurable substrate gain under edge budgets

Thesis

Once a local model crosses a minimum linguistic threshold, substrate design should predict system behavior more strongly than model identity does.

Bare models may differ widely. When run through the same compiled arrival packet, constrained output schema, deterministic validation, and one repair loop, they should converge toward the same functional behavior (state updates, constraint obedience, continuity)—not stylistic sameness.

Success condition (v0)

Conditioned Kernel succeeds if the same small local model, when run through the substrate, becomes more coherent, more state-faithful, more continuous, and more repairable than when run bare—and if those gains survive a model swap within the tested size band.

Quick start

cd conditioned-kernel
python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

# Requires Ollama running locally with a small model, e.g.:
#   ollama pull qwen3.5:0.8b
ollama serve   # if not already running

ck status
ck edge
ck edge --list
ck smoke
ck ask "Summarize the current design intent in one short paragraph."

# Multi-turn session (substrate keeps a byte-capped recent dialogue):
ck chat
# resume later (same state dir): ck chat
# clear dialogue memory only:     ck chat --new-session

# Opt-in desktop headroom only (product decisions still use orin_nano_8gb):
ck --profile desktop_dev status

On Jetson:

bash scripts/jetson_bootstrap.sh

Offline tests (no Ollama required):

pytest -q

See docs/EDGE_SPEC.md.

Architecture

terminal → substrate_state → compile(arrival packet)
        → generate(Ollama) → return_path(parse → validate → assess)
        → accept | repair(one pass) | reject
        → terminal render + persistent receipts
Module Role
state Load/write filesystem substrate
compile Bound arrival packet from state + input
generate Ollama client (chat_json / generate_raw)
return_path Parse, validate, assess, repair, accept
cli Terminal surface

v0 does not stream model tokens to the terminal. The substrate buffers the full candidate before acceptance.

Default models

Primary experimental window: 0.5B–1.5B. Stretch lower with SmolLM2-class probes; stretch upper carefully on Jetson Orin Nano 8GB.

Default Studio kernel (orin_nano_8gb): qwen3.5:0.8b with think=false via the Ollama API. Also useful tags: qwen2.5:0.5b, qwen2.5:1.5b, granite4:350m, tinyllama:1.1b.

Experiment discipline

See docs/EXPERIMENT_PROTOCOL.md.

Conditions include bare, budget-matched bare, static persona, full Conditioned Kernel, and ablations. Metrics split structural recovery (parse/schema/repair) from semantic substrate gain (faithfulness/continuity on valid candidates).

Studio observations (public)

Lived companion / Interior View field notes, plus the published raw evidence they cite (evidence/ — full session traces, ledgers, and state, hash-manifested, published at the operator's explicit direction):

docs/observations/ — index, privacy rules, and session notes (Context Field, not_responsive, night phenomenology).

Laboratory run receipts stay under docs/adaptive/. Philosophy: docs/PURPOSE_AND_RIVER.md.

Lineage

Synthesis of Temple Two public work:

  • Context Field Conditioning — structure of delivery changes outcomes
  • Phenomenological Compass — organizing posture separate from answering model
  • T2Helix — pre/post chokepoints, local storage, redaction
  • Sovereign Stack — continuity and governance outside any single model instance

v0 does not import the full stack. Bridge surfaces are P3 after the core experiment stabilizes.

Repo layout

conditioned-kernel/
├── src/conditioned_kernel/   # package
├── state/                    # default substrate files
├── logs/                     # receipts / history (live writes; frozen session copies land in evidence/)
├── experiments/              # matrix runners + probes
├── tests/                    # offline unit tests
└── docs/
    ├── adaptive/             # Laboratory run contracts & receipts
    ├── observations/         # public Studio field notes (this is the showcase)
    └── *.md                  # thesis, architecture, purpose, protocol

License

Apache-2.0. Copyright 2026 Anthony J. Vasquez Sr. / Temple of Two.

About

Substrate-conditioned generation: treat the LLM as a replaceable transduction kernel and relocate system behavior into the substrate (state, compiled arrival packet, validation, repair, acceptance). Edge-first, fully local, Jetson Orin Nano 8GB default.

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages