t81-foundation

T81 Foundation — Comprehensive Repository Research & Architecture Reassessment

Document ID: docs/research/T81_FULL_REASSESSMENT_2026-10-25.md Date: October 25, 2026 Status: Complete / Final Research Assessment Author: Independent Principal Research Engineer & Systems Auditor Target Repository: T81 Foundation (t81dev/t81-foundation) Pinned Snapshot Baseline: Commit 962fb5ef24b012f91a1a3d1cb18a12d2145e66f0 (Main Branch) Continuity Baseline: Commit 860558b4ad1cf4160af30860b00b15c46165acd5

A. Executive Assessment

The T81 Foundation repository represents an experimental scientific research framework and software engineering substrate focused on two main research domains:

  1. Track R (Scientific Computing & Alternative Representations): Investigating whether ternary arithmetic, balanced ternary representations, Base-81 digit grouping, adaptive tiering, 3-state logic primitives, or ternary weight formats produce measurable or theoretically defensible computational, storage, or execution advantages over equivalent binary systems (Canonical Hypotheses H01–H06).
  2. Track F (Foundation Systems Engineering & Governance): Building and evaluating deterministic virtual execution, content-addressed artifact storage (CanonFS), pre-side-effect policy authorization (Axion), immutable execution provenance, and detached accountable evidence verification (Canonical Hypothesis H07).
  3. Track M (AI-Assisted Research Methodology): Evaluating prompt-to-code provenance, error rate tracking, and trustworthy research iteration under AI assistance (Canonical Hypothesis H08).

Key Reassessment Findings

B. Architecture Inventory

A comprehensive audit of the primary subsystems in the repository yields the following implementation, testing, and operational status:

Subsystem Implemented Role Key Source Paths Test & Verification Coverage Implementation Status
TISC ISA (v1.9.0) 81-opcode instruction set architecture for T81 execution include/t81/isa/opcodes.hpp, src/isa/ tests/cpp/test_isa.cpp, CTest #391–396 Frozen (v1.9.0)
T81VM Interpreter Virtual machine executing TISC bytecode with DCP determinism src/vm/vm.cpp, include/t81/vm/vm.hpp tests/cpp/vm_determinism_property_test.cpp Beta Runtime
Axion Policy Kernel Pre-dispatch policy evaluation engine (eval_axion_call) src/axion/, include/t81/kernel/axion/ tests/cpp/axion_agent_invoke_policy_test.cpp, CTest #154 Alpha Governance
CanonFS Content-addressed storage and foreign filesystem interchange (RFC-00D1) src/core/fs/, include/t81/fs/ tests/cpp/canonfs_test.cpp, CTest #84–85 Beta Identity FS
T81Lang Tooling Frontend lexer, parser, AST, IR generator, and bytecode compiler src/lang/, include/t81/lang/ tests/cpp/test_lang.cpp, t81 code check/build Beta Language
Model Intake Bridge GGUF/Safetensors model header inspection & tensor inventory src/tools/model/, tools/cli/ai/ t81_weights_load_test (CTest #166) Metadata Bridge
Determinism Facilities Soft-float math (RFC-0030 CORDIC), bit-exact containers, CanonHash-81 src/core/isa/math/, vm/determinism/ CTest #355–362 Functional / DCP
Ternary / Base-81 Codecs SWAR SIMD emulation, Base-81 tryte packing, 2-bit codecs src/core/codec/, src/isa/math/quantization/ tests/cpp/test_swar.cpp, test_base81_simd.cpp Experimental / Frozen
Accountable Evidence (RFC-00D4) Detached evidence bundle generator & verifier examples/governance/accountable-computation/ t81_accountable_experiment (CTest #4) Research Reference

External Dependencies

C. Historical Lineage & Hanoi Research Inheritance

The historical progression of the T81 computational concepts spans four distinct phases:

[ Knuth Seminumerical Algorithms ]
 (Balanced ternary, positional number systems, 3-valued logic)
               |
               v
[ Hanoi CWEB Exploration ]
 (CWEB implementation, 41 historical source blobs, experimental arithmetic)
               |
               v
[ T81Lang / TISC ISA / T81VM Engine ]
 (TISC v1.9.0 opcode frozen specification, C++ interpreter, SWAR ternary vectorization)
               |
               v
[ Axion Policy OS & CanonFS Core ]
 (Fail-closed pre-dispatch governance, content-addressed storage, decision bundles)
               |
               v
[ T81 Foundation Stack ]
 (Deterministic execution, policy-gated AI runtimes, accountable computation framework)

Archaeological Findings

  1. Hanoi CWEB Inheritance: Code in experimental/hanoi/ preserves the CWEB-based exploration of ternary state machines. Historical audits identified 41 shared source blobs transferred from Hanoi.
  2. Arithmetic Defect Correction: Early Hanoi implementations suffered from carry-propagation boundary errors and ambiguous zero representations in signed-digit representations. These were mathematically analyzed and corrected during TISC ISA stabilization (docs/research/TERNARY_ARITHMETIC_AND_RADIX_COMPARISON.md).
  3. Transformation of Scope: The research shifted from attempting to build a general-purpose desktop operating system or native ternary hardware CPU toward evaluating deterministic, policy-governed runtime boundaries for AI artifacts and accountable evidence.

D. Hypothesis Assessment (Master Canonical H01–H08 Register)

The project evaluates eight canonical hypotheses under its Master Continuity Charter (docs/research/continuity/MASTER_CONTINUITY_CHARTER_v1.0.md and docs/research/continuity/HYPOTHESIS_REGISTER_v1.json). Canonical identifiers and statements are preserved as authoritative:

H01: Native Ternary Storage & Logic Cost Reduction (Track R)

H02: Balanced Ternary Signed Arithmetic & Rounding (Track R)

H03: Base-81 Digit Grouping & Representation Interface (Track R)

H04: Adaptive Representation Tiering (Track R)

H05: Three-State Algebra Primitives & Architectures (Track R)

H06: Ternary Weights & Activations for AI Inference (Track R)

H07: Representation-Independent Governed Execution (Track F)

H08: AI-Assisted Iterative Research & Trustworthy Discovery (Track M)

E. Historical Experiment Evidence Mapping

To maintain clarity without renumbering or altering the canonical H01–H08 register, empirical benchmark experiments in the repository map to the hypothesis framework as follows:

Benchmark / Experiment Target Surface / Command Evaluated Hypotheses Primary Findings & Scoped Verdicts
Radix Comparison experiments/radix_comparison/ H01, H03 Software tryte addition is $10.6\times$ slower than native 64-bit binary ALU addition. Falsifies software CPU ternary arithmetic speedup.
Packed Ternary MatVec experiments/packed_ternary_matvec/ H02, H03, H06 2-bit packed weight format saves $75\%$ memory footprint, but SIMD bit-shuffle decode tax limits execution speed to $75.3\%$ of INT8 AVX2 in cache.
Accountable Computation (Baseline-X) examples/governance/accountable-computation/ & experiments/research_001/ H04, H05, H06, H07 Wasmtime + OPA host FFI capability baseline (Arm iii) achieves identical $0.0\%$ breach rates and $100\%$ detached evidence auditability with $10\times$ lower host C++ LOC. Proves accountable computation is a protocol, not a runtime requirement.
Cross-Platform Determinism (DCP) tests/cpp/vm_determinism_property_test.cpp H07 Bit-exact cross-platform execution and RFC-0030 soft-float transcendentals pass $100\%$ of property-based tests across x86_64, ARM64, and WebAssembly.
CanonFS Byte Identity & Interchange src/core/fs/ & CTest #84–85 H07, H08 Content-addressed storage (CanonHash-81) and foreign filesystem interchange (RFC-00D1) deliver immutable byte identity across systems.

E. Claims Audit

An audit of claims in former documentation against implemented code and test evidence reveals:

Category Claim Summary Empirical Status Source Evidence Reassessment Finding
ISA / Runtime “TISC v1.9.0 is a frozen, deterministic instruction set.” SUPPORTED include/t81/isa/opcodes.hpp, CTest #391–396 Core 81 opcodes and bit-exact execution are fully implemented and tested.
Policy Governance “Axion policy checks prevent unauthorized side effects before dispatch.” SUPPORTED src/axion/, CTest #154 Pre-dispatch eval_axion_call fails closed and logs policy denials before state mutation.
Artifact Storage “CanonFS provides content-addressed artifact storage and interchange.” SUPPORTED src/core/fs/, CTest #84–85 CanonFS import, export, and foreign interchange (RFC-00D1) function as specified.
Deterministic Soft-Float “RFC-0030 soft-float transcendentals provide cross-platform parity.” SUPPORTED src/core/isa/math/, CTest #355 Soft-float CORDIC math provides bit-exact parity across x86_64 and ARM64.
Ternary CPU Speedup “Software Base-81 / SWAR ternary execution accelerates computation on binary CPUs.” CONTRADICTED experiments/radix_comparison/ Software tryte addition is $10.6\times$ slower than native 64-bit binary ALU addition.
MatVec Inference “2-bit packed ternary MatVec executes faster than INT8 AVX2.” CONTRADICTED experiments/packed_ternary_matvec/ VPSHUFB 2-bit MatVec reaches $75.3\%$ of INT8 AVX2 speed due to SIMD decode instruction tax.
Custom VM Requirement “Accountable computation requires a custom VM and kernel.” UNSUPPORTED examples/governance/accountable-computation/ Wasmtime + OPA + RFC-00D4 evidence schema provides equivalent policy and audit guarantees.
Hardware Platform “T81 runs on physical ternary microprocessors.” UNSUPPORTED Repository inspection No physical ternary hardware exists in the codebase; all execution uses binary hardware emulation.

F. Technical Debt and Risks

  1. Process Security & Shell Traversal:
    • Command execution in CLI tools should continue avoiding std::system and shell expansion to prevent command injection risks.
    • External processes must strictly use t81::support::run_process (include/t81/support/process.hpp) with explicit argument vectors.
  2. Windows PATH Parsing:
    • PATH environment variable entries on Windows may contain surrounding quotes (e.g. "C:\Program Files"), requiring string trimming before passing to std::filesystem::path.
  3. Durable Witness Isolation:
    • Security audits (experiments/durable_witness/) established that an isolated SQLite WAL witness store remains vulnerable to snapshot rollback without an external monotonic checkpoint or log. HMAC receipts must sign all metadata fields (witness_id, payload_hash, total_amount, timestamp).
  4. Build Warnings:
    • Apple Clang compiler warnings regarding dangling references to temporaries in examples/research/validated-descriptors/runner.cpp line 144 should be cleaned up during maintenance.

G. Research Opportunities

  1. Galois Field $\text{GF}(3^9)$ vs $\text{GF}(2^8)$ Storage & Parity: Evaluating Base-81 Galois Field arithmetic ($\text{GF}(3^9)$) for Reed-Solomon storage parity in CanonFS versus hardware-accelerated binary Galois Field instructions (GF2P8AFFINEINVQB).
  2. Multi-Precision Signed-Digit (RSD) Crossover: Microbenchmarking Avizienis carry-free addition across 128-bit to 512-bit integer widths to identify the exact crossover point where carry-free digit addition overcomes digit-unpacking overhead on binary SIMD vector units.
  3. Deterministic AI Execution & Trace Provenance: Validating soft-float transcendentals (RFC-0030) and trace hashing across large neural network inference execution graphs.

H. Next Discriminating Experiments & Research Sequence

Research experiments proceed along established sequence tracks defined in the Master Continuity Charter:

1. Track R Research Sequence: H04-R01 & H04-R02 (Adaptive Representation Discriminator)

2. Track F Candidate: Cross-Platform Deterministic AI Execution & Trace Parity (Candidate 3)

I. README Change Rationale

The root README.md was rewritten to align project documentation with empirical source code evidence:

  1. Clear Project Identity: Replaced vague architectural promises with a precise description of T81 Foundation as an experimental research program investigating deterministic execution, policy governance, provenance, and alternative computational representations.
  2. Honest Hypothesis Ledger: Included explicit results for H01–H08, openly disclosing negative experimental findings (such as software ternary CPU arithmetic speedup being falsified).
  3. Reproducible Instructions: Provided verified, exact CMake, Ninja, CTest, and Python regression commands that execute cleanly in the codebase today.
  4. Removed Promotional Badges & Claims: Removed unsupported performance badges, claims of physical ternary hardware, and assertions that custom VMs are strictly necessary for accountable evidence.

J. Evidence Ledger

This site is open source. Improve this page.