Silicon Brain — Technical Brief
Status: Design / Vision overlay. Not a product. Not a measured result.
Date: 2026-08-23
Belt: Vision named separately from Science. This brief states both.
The canonical engineering copy lives at docs/SILICON-BRAIN-BRIEF.md in this repository (path is valid once this pull request is merged). The text below is kept in lockstep so the public docs viewer can open it.
Interpretation Boundary
This document does not demonstrate:
- That CrystalCore is a brain, a mind, or a consciousness
- That H1 or H2 have been proven
- That a recorded 1M dual-gate fail on this encoder is a CrystalCore product result
- Language-model quality, hardware bandwidth, or a costed BOM
- Non-local, dimensional, or brain-to-brain information transfer
- Fitness for Optimus, xAI, or any external pitch
- Affiliation with any company named in the mythos
Interpersonal neural coupling (Kelsen et al., Neuroscience & Biobehavioral Reviews 132:1249–1262) is ordinary sensory-mediated neuroscience. It is not evidence for this brief.
Plate XXXV is Vision-layer atlas art. It is not a schematic of running hardware.
Two layers, kept apart
| Layer | What “Silicon Brain” means here | Status |
|---|---|---|
| Vision | Atlas name for a future crystalline receiver/core: memory as geometry, attention as routing, sovereignty as the refusal to impersonate a mind. See Plate XXXV. | Dreamed |
| Science | An engineering research programme around long-context attention primitives (H2) and earlier unproven H1 work, sitting beside real, narrower systems that already run (consent transport, companion gate). | Mixed: H1 frozen unproven; H2 methodology locked, configs frozen, dual-gate fail recorded on this encoder; production core/ is the consent/record stack, not a brain |
The name is a metaphor with a measurement programme attached. The metaphor does not get to spend the measurements.
What actually exists (Science)
Verified against CrystalArchitect/TerAustralis-Incognita-Code on 2026-08-23.
| Surface | What it is | What it is not |
|---|---|---|
core/crystalcore/, core/crystal-core/ |
Consent gate, Starline P2P, receipts, record kernel | A neural architecture |
docs/H2/, results/final-test-v0.1/ |
Isolated experimental surface on main: sparse/hierarchical attention candidates vs dense reference on a frozen synthetic associative-memory encoder. Dual-gate at 1M recorded as fail. |
A language model; CrystalCore production attention |
| H1 | Frozen. Unproven. Untouched by H2. | — |
| Site / Pages | Built from main. |
Evidence of an H2 product |
H2 locked hypothesis (1M, evaluated 2026-08-23 — fail on this encoder):
LCB(Q_primitive / Q_full) ≥ 0.90 and BW_primitive / BW_full ≤ 0.40, simultaneously.
Quality LCB passed; analytical BW/ref did not (A 0.5638, B 0.5614). Simultaneous gate not met. See results/final-test-v0.1/. Synthetic associative-memory encoder, fp32, analytical bandwidth. Not a language-model result.
Receiver, attention, multi-agent — engineering only
Earlier atlas language used Receiver / Core / Coexisting Layers. Mapped to engineering, and only this far:
- Receiver — a clean input channel. In H2 this is the query that addresses bound keys. It is not a sensory organ and not a person.
- Core — the frozen encoder + attention primitive under test. Weights, tokenizer-equivalent vocab, task, and precision stay locked; only the attention mechanism may change.
- Multi-agent — multiple processes exchanging consent-gated memory (Starline) is real, narrow code. Transient correlation between human brains during shared tasks is a different, classical phenomenon. Neither is evidence of the other.
No claim that silicon “syncs” with cortex.
Why the atlas plate exists
Plate XXXV — The Silicon Brain is a gold-on-indigo atlas plate. Left hemisphere: hierarchical block-sparse attention (H2 Candidate A). Right: content-addressed retrieval (H2 Candidate B). Ghosted behind both: dense full attention, the reference, not a product.
It exists so the Vision layer has a canonical image that labels the split instead of hiding it. The plate is allowed to be beautiful. It is not allowed to be a result.
What would have to happen before the name earns more weight
In order, and not sooner:
- H2 validation recorded (protocol step, not a proof).
- One config frozen per candidate — recorded 2026-08-23; see
docs/H2/FROZEN-CONFIGS-V0.1.md. - Immutable final test, 32k → 1M, dual gate — recorded 2026-08-23. Dual-gate fail (quality LCB pass, bandwidth fail). See
results/final-test-v0.1/. - Independent reproduction of the encoder result.
- A separate, later decision — not implied by (3) — about whether any primitive belongs near production CrystalCore.
- Hardware path, if any, with a costed BOM and a safe operating envelope. None is open.
Until a separate later decision — not implied by the 1M encoder gate — “Silicon Brain” remains a name plus a locked methodology and a recorded fail on this encoder.
Cost / compute
H2 currently records wall-clock on the experimental runner and analytical fp32 element counts. There is no dollar BOM and no hardware PMU. A future prototype, if ordered, should carry an itemised cost boundary the way a buildable lab plate does. That boundary is not on this brief.
Track status (unchanged by this document)
| Track | Status |
|---|---|
| H1 | Frozen · Unproven |
| H2 | Methodology locked · Steps 1–6 recorded · dual-gate fail at 1M on this encoder · no product claims |
| CrystalCore vs dense @ 1M | Unproven |
| External / xAI pitch | Hold |
Licence
CC BY-NC-ND 4.0. Vision art and this brief are not a specification of shipped software.