From 2314f5a6aec7672bcde6f1eeeb305a83c7d7647d Mon Sep 17 00:00:00 2001 From: Dmitrii Vasilev Date: Mon, 18 May 2026 18:21:00 +0000 Subject: [PATCH] docs: TRI-NET unified positioning in PhD architecture (TRI-17) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Add TRI-NET: Unified Positioning (TRI-17) section to ARCHITECTURE.md and Hardware Realization: TRI-NET Three-Chip Stack section to PHD-RESEARCH-PROGRAM-AND-DISSERTATION.md. Changes are purely additive — no existing content removed. Rebased on current main to resolve conflicts. Closes #707 --- docs/ARCHITECTURE.md | 39 +++++++++++++++++++ docs/PHD-RESEARCH-PROGRAM-AND-DISSERTATION.md | 36 +++++++++++++++++ 2 files changed, 75 insertions(+) diff --git a/docs/ARCHITECTURE.md b/docs/ARCHITECTURE.md index 05c10d751..6b13dc124 100644 --- a/docs/ARCHITECTURE.md +++ b/docs/ARCHITECTURE.md @@ -7,6 +7,45 @@ This document ties **mathematics**, **cognitive / agent architecture**, and **la --- +## TRI-NET: Unified Positioning (TRI-17) + +> **TRI-NET — verifiable open silicon stack for trustworthy AI: identity (Φ), reasoning (E), inference (Γ). One math anchor, three chips, zero closed IP.** + +TRI-NET is the hardware embodiment of the Trinity S³AI architecture: three chips manufactured on the TTSKY26b shuttle (TinyTapeout SKY130A), each implementing a distinct specialization of the φ²+φ⁻²=3 identity in fabricated silicon. + +**DOI:** [10.5281/zenodo.19227877](https://doi.org/10.5281/zenodo.19227877) +**Shuttle:** TTSKY26b (TinyTapeout SKY130A) — submitted 2026-05-19 +**Performance target:** ~1 GOPS @ ~50 MHz @ ~1 W ternary (projected) + +### TRI-NET chip specializations + +| Chip | Role | Tagline | Project ID | Tiles | RTL Modules | +|------|------|---------|-----------|-------|-------------| +| **Φ Phi** | Identity layer — root-of-trust, attestation, die-unique | *"The boot you can prove."* | #4914 | 1×1 | 51 | +| **E Euler** | Verification layer — symbolic AI, CLARA 10-gap safety lattice, host I/O | *"Reasoning you can audit."* | #4915 | 8×2 | 90 | +| **Γ Gamma** | Inference layer — neuromorphic cortex, 20-PE GF16 mesh, full GF multiplier portfolio | *"Cortex you can measure."* | #4913 | 8×4 | 105 | +| **TRI-NET** | Full stack | *"Three chips. One anchor. Provable compute."* | — | — | **246 total** | + +### Cross-die canonical anchor 0x47C0 + +All three chips independently assert `{uio_out, uo_out} = 0x47C0` on reset. The derivation chain is: + +``` +φ²+φ⁻²=3 → Lucas number L₂=3 → GF16 dot4(1,2,3,4)=0x47C0 +``` + +This constitutes **Theorem 36.1 (TG-TRIAD-X)**: a mathematically grounded silicon witness that the three chips are instances of the same formal object. No closed-IP chip can independently reproduce this chain from first principles; TRI-NET's Apache-2.0 RTL makes it fully verifiable from gate level. + +### Architectural philosophy + +1. **Φ = Identity layer** — root-of-trust via Lucas-chain POST (φ²+φ⁻²=3), die-unique HWRNG, CLARA bounded-rationality restraint (Gap-4). Proves: *"I exist and have not been substituted."* +2. **E = Verification layer** — full DARPA CLARA 10-gap safety lattice (Kleene K3 ALU, Datalog, SAT/ASP solvers, proof-trace audit, BLAKE3 receipts), ternary VSA matmul. Proves: *"This computation is correct and explainable."* +3. **Γ = Inference layer** — 8 cortical LIF columns, FHRR holographic binding, full GF4–GF256 multiplier portfolio, thermal-aware AVS. Proves: *"AI inference is happening with verifiable energy."* + +See **`docs/TRI_NET_WHITEPAPER.md`** and **`docs/TRI_NET_API.md`** for full module inventories. The DOI [10.5281/zenodo.19227877](https://doi.org/10.5281/zenodo.19227877) archives the complete RTL submission. + +--- + ## 1. Trinity identity — one constraint, three readings The identity **φ² + 1/φ² = 3** (golden ratio φ) is treated as a **single organizing equation** with three simultaneous readings: diff --git a/docs/PHD-RESEARCH-PROGRAM-AND-DISSERTATION.md b/docs/PHD-RESEARCH-PROGRAM-AND-DISSERTATION.md index 3199c8c18..0601fca41 100644 --- a/docs/PHD-RESEARCH-PROGRAM-AND-DISSERTATION.md +++ b/docs/PHD-RESEARCH-PROGRAM-AND-DISSERTATION.md @@ -35,6 +35,42 @@ See **`docs/ARCHITECTURE.md`** for the strand decomposition and repository map. --- +## Hardware Realization: TRI-NET Three-Chip Stack + +TRI-NET is the **silicon embodiment** of the PhD program's central verifiable-AI thesis. Three chips submitted on the TTSKY26b shuttle (TinyTapeout SKY130A, 2026-05-19) instantiate the three-strand architecture (φ²+φ⁻²=3) as fabricated RTL. + +**DOI:** [10.5281/zenodo.19227877](https://doi.org/10.5281/zenodo.19227877) +**Total RTL modules:** 246 (Phi 51 + Euler 90 + Gamma 105) + +### Chip-to-chapter traceability + +| Chip | Strand | PhD chapter alignment | +|------|--------|-----------------------| +| **Φ Phi** (#4914, 51 modules, 1×1 tile) | Strand I — Identity | Chapter 4 (identity / attestation): silicon root-of-trust, die-unique HWRNG, Lucas POST as constructive proof of φ²+φ⁻²=3. Chapter 8 (governance / integrity): attestable boot as engineering-ethics primitive. | +| **E Euler** (#4915, 90 modules, 8×2 tile) | Strand II — Verification | Chapter 6 (AR / CLARA pipeline): the only known hardware implementation of all 10 DARPA CLARA AI Safety Gaps (Kleene K3 ALU, mini-Datalog, SAT/ASP solvers, proof-trace audit, BLAKE3 receipts). Chapter 7 (hardware & numerics in silicon): wishbone host bus, ternary VSA matmul 8×8 and 16×16. | +| **Γ Gamma** (#4913, 105 modules, 8×4 tile) | Strand III — Inference | Chapter 7 (hardware & numerics in silicon): 8 cortical LIF columns, 20-PE GF16 mesh, FHRR holographic binding (Chapter 32 VSA), full GF4–GF256 multiplier portfolio, dual-library zoning. Chapter 3 (ternary logic): on-chip ternary ALU and BitNet b1.58 MLP as silicon artifact of K3 theory. | + +### Theorem 36.1 as silicon witness + +The **cross-die canonical anchor** `{uio_out, uo_out} = 0x47C0` — present on reset in all three chips — constitutes the **silicon witness of Theorem 36.1 (TG-TRIAD-X)**. The derivation chain: + +``` +φ²+φ⁻²=3 → Lucas L₂=3 → GF16 dot4(1,2,3,4)=0x47C0 +``` + +This transforms WP2 (GoldenFloat & sacred-physics numerics) and WP4 (CLARA-style AR in ternary logic) from paper theorems into **independently reproducible fabricated evidence**. Any examiner can clone the Apache-2.0 RTL and verify the chain from gate-level simulation to physical silicon. + +### Academic significance + +- **WP1** (formal semantics): TRI-NET’s 246-module RTL corpus is the largest single-author open-silicon artifact implementing K3 ternary semantics; it provides a concrete formal substrate for the type-system and invariant chapters. +- **WP4** (CLARA): Euler chip’s 10-gap hardware lattice is novel evidence that bounded-trace AI safety constraints are **implementable at sub-1 W** in open silicon, not just as software libraries. +- **WP5** (FPGA / MAC / ISA bridge): Gamma chip’s 8-column LIF cortex and full GF multiplier set supply the measurement and resource-bound data required by Chapter 7; target ~1 GOPS @ ~50 MHz @ ~1 W ternary (projected). +- **WP6** (governance): all three chips carry the same 0x47C0 anchor, making the three-chip stack a physical instantiation of the PHI LOOP seal discipline — every tapeout commit is traceable to a ring-gated PR. + +See **`docs/ARCHITECTURE.md`** section “TRI-NET: Unified Positioning (TRI-17)” and **`docs/TRI_NET_WHITEPAPER.md`** for full technical detail. + +--- + ## 2. Central hypothesis (defensible PhD spine) **Hypothesis (working):** A **spec-first** pipeline combining **ternary (K3) logical structure**, **GoldenFloat-class numerics**, and **machine-checked conformance vectors** yields **more auditable and safer** neurosymbolic AI stacks than ad-hoc binary toolchains where semantics live in scattered scripts and notebooks.