the R&D archive, by the numbers

byte-identical

gcc ≡ clang

the compiler compiles itself, byte-for-byte

qc2 self-hosts, diverse-double-compiled (gcc ≡ clang) — same bytes both ways.

CaveatA real subset of the language; of 570 chapters independently audited, ~59% are confirmed genuinely real (60 stubs quarantined) — not the full frontier yet.

byte-identical

9 frameworks · 46 tests

one spec, the same working app everywhere

FBL compiles a single wiring spec to output that matches across every target framework.

CaveatA deterministic wiring spec, not arbitrary-program generation.

151.7

tok/s

a real Qwen runs on our own stack

Full interactive generation in QUANTA-emitted PTX — token-for-token == llama.cpp, zero cuBLAS.

Caveat0.5B model; 7B decode is at parity, and prefill still trails llama.cpp.

2×

at 128k context

stays fast when the context is huge

Flat at 128k where a transformer climbs — measured against llama.cpp.

CaveatLong-context regime specifically; not a claim of being fastest everywhere.

92.8%

of cuBLAS

a competitive GEMM, bit-exact, on our PTX

WMMA GEMM from QUANTA-emitted PTX, bit-exact against the reference.

Caveat92.8% is PARITY, explicitly NOT fastest — cuBLAS is still ahead.

87% → 0.0%

it refuses instead of bluffing

Confabulation driven to zero — watch it cite or refuse under a 16-turn adversary.

CaveatThe cold-judge stranger score is still 3/10; the no-bluff floor holds, the polish doesn't yet.

20 GB → 0.70 GB

60M → 22.2M citeable facts

a citeable knowledge store you could hold on a phone

28.6 bytes/fact, every answer traced to a source (Canberra, Austen, da Vinci, Tokyo).

CaveatReconstruction of stored facts, not prediction — a retrieval store, not a bigger LLM.

21 ms × 377 MiB

one 3090

a capable mind on hardware you already own

M7 brain, 3 of 4 meters green, pure stack — one consumer GPU, no data centre.

CaveatLoss 0.84 is reconstruction; “calculated > trained at LM scale” is still open.

go / no-go = GO

a memory that actually sticks

does-it-stick, run & validated on real Qwen-0.5B embeddings (cos 0.87 ≈ random 0.90).

CaveatNeeds dreaming or it's recency-only; recall is approximate, not verbatim.

2.2 KB

on a Z80

a real neural net on a 1980s home computer

A trained network runs, net 2.2 KB, on a ZX-Spectrum-class Z80 — the efficiency thesis, extreme.

CaveatA demonstration of the low-end floor, not a production model.

4.2× · 3×

faster · smaller

the compiler's solver beats the naive path

Register allocation 25.8 → 9.6 cycles/call; a cost model picks the strategy — 0.8 TFLOP/s, no BLAS.

CaveatOn the measured kernels; not yet a winning kernel authored end-to-end in QUANTA.

OPEN

“calculated > trained” at LM scale — not yet proven

The moat claim needs downstream evals; today's 0.84 is reconstruction, not prediction.

CaveatWe keep this visible on purpose — it's the honest edge that turns 🔬 into 🟢.