The record
Benchmarks & claims — tiered, so they survive review.
Every figure carries its tier. Measured executed & sha-pinned · Exact proved & machine-gated · Vision declared trajectory.
An 8B model stays correct at a million tokens. The substrate alone does better still.
Official BABILong prompts, 50 samples per cell, Llama-3.1-8B at temperature 0. Unaided the model falls 0.68 → 0.22 → 0.10 → 0.02 → 0.00 across 4k to 1M. Given toral memory it holds 1.00 flat across that whole 250× range. And with no language model at all, the Klein-168 field scores 1.00 on qa3 and qa4 — the two hardest cells — where that same 8B, handed the same memory, manages only 0.90 and 0.82.
Scope: the flat curve is the qa1 task. qa2–qa5 are measured at 4k only so far (0.32→1.00, 0.54→0.90, 0.38→0.82, 0.86→0.90); extending them across the range is open work, not a finished result.
| Result | Value | Where | Tier |
|---|---|---|---|
| Language stratified one-hot by the V₄ gauge | 100% | semantic bus | Measured 7/7 |
| Tera-Wilson V₄ gauge holonomy throughput (commodity CPU) | 2.94×10¹² ops/s | AVX-512, sha-pinned | Measured |
| Cross-vendor byte-identity (Intel · AMD · ARM), 250k shards × 3 hosts | identical digests | replay | Measured |
| The Agnostiglot — one shared context in three languages | Rust≡Py≡JS | tri-runtime | Measured |
| Small model + substrate memory (BABILong recall to 1M) | 0.05 → 1.00 | 0.5B CPU reader | Measured |
| Substrate answering alone, no language model (BABILong qa3/qa4) | 1.00 / 1.00 | n=50/cell; beats 8B+substrate 0.90 / 0.82 | Measured |
| Shared timeline, constant state | 32 B at 100M | state size measured; throughput varies ~4x by code path, unpinned | Measured |
| Million-qubit GHZ topological emulation | 5.4 s / 114 MB | deterministic geometry | Measured |
| Holy enumeration of the Leech shell (196,560 min. vectors) | 45·16ᵏ law | two derivations | Exact ×2 |
| Bennett-reversible calculus — logical erasure of the alphabet | 0 bits | reversible ledger | Exact 21/21 |
| The single irreversible act (commit), priced exactly | 84 bits | the ledger | Exact |
| The exact tower from one twin-prime seed | 8 · 24 · 120 | E8/Leech/Craig | Exact 19/19 |
| Structural dedupe on real token streams | 47–58% | routing | Measured |
| Role-keyed cache vs LRU (small capacity) | 1.6–2.2× | toral cache | Measured |
| Path certification / tamper detection | 5000/5000 | gauge holonomy | Measured |
| Gated corpus (this paper's governance) | 664 / 664 gates | 71 laboratories | Exact |
No task-accuracy uplift for a learned model is claimed — by design. The wins are on the axes the substrate owns: determinism, reversibility, byte-reproducibility, recall-per-budget, cost, and coverage.
AI's growth is a power and freshwater problem. This is built the opposite way.
GPU-scale training and inference are driving datacenter electricity and cooling-water demand upward fast. This substrate's architecture removes the cause rather than optimizing around it: no GPUs, a ~100 MB working set, $0 marginal cost, deterministic and reversible — power is paid only at the single commit, and everything else is free by construction.
| Efficiency property | Value | Tier |
|---|---|---|
| Compute substrate | CPU-only, no GPU | Architectural |
| Working set | ~100 MB | Measured |
| Marginal cost per op | $0, deterministic | Measured |
| Reversible motion / commit floor | 0 bits / 84 bits | Exact |
| Embed-call dedupe / cache reuse | 47–58% / 1.6–2.2× | Measured |
| Collapsing AI's energy & water footprint | the design goal | Trajectory |
Honest boundary: we claim no measured joule and no specific water figure — the thermodynamic values are exact lower-bound targets, and the environmental impact is the stated design goal, not a metered result. What is measured is the architecture: no GPUs, constant memory, reversible, deterministic.
Security that actually works — because meaning here is constructed geometry, not learned weight.
Access is capability-addressed by self-inverse involutions (units whose square is the identity mod 35) — there is no sequential secret-key pipeline to leak. Custody is derive-don't-store: no derived state is persisted; everything re-emerges byte-identically from append-only, hash-chained seeds, and any tamper avalanches and fails replay. It is a foundation designed for the post-quantum era, not retrofitted to it.
| Security mechanism | Property | Tier |
|---|---|---|
| Capability addressing | involution units, unit²≡1 | Exact |
| Custody / integrity | derive-don't-store, replay-verified | Measured |
| Tamper detection (single-symbol) | avalanche ≥ 0.97–1.00 | Measured |
| Provenance | append-only hash chain | Structural |
| Post-quantum internet · signalling · modems | the roadmap | R&D |